Citation
Quantitative Metrics to Evaluate Image Quality for Computed Radiographic Images

Material Information

Title:
Quantitative Metrics to Evaluate Image Quality for Computed Radiographic Images
Creator:
PITCHER, CHRISTOPHER D. ( Author, Primary )
Copyright Date:
2008

Subjects

Subjects / Keywords:
Arithmetic mean ( jstor )
Distance functions ( jstor )
Dosage ( jstor )
Image contrast ( jstor )
Image processing ( jstor )
Images ( jstor )
Imaging ( jstor )
P values ( jstor )
Pixels ( jstor )
Radiology ( jstor )

Record Information

Source Institution:
University of Florida
Holding Location:
University of Florida
Rights Management:
Copyright Christopher D. Pitcher. Permission granted to University of Florida to digitize and display this item for non-profit research and educational purposes. Any reuse of this item in excess of fair use or other copyright exemptions requires permission of the copyright holder.
Embargo Date:
12/18/2004
Resource Identifier:
57731737 ( OCLC )

Downloads

This item has the following downloads:

pitcher_c ( .pdf )

pitcher_c_Page_018.txt

pitcher_c_Page_129.txt

pitcher_c_Page_138.txt

pitcher_c_Page_088.txt

pitcher_c_Page_185.txt

pitcher_c_Page_025.txt

pitcher_c_Page_020.txt

pitcher_c_Page_053.txt

pitcher_c_Page_092.txt

pitcher_c_Page_068.txt

pitcher_c_Page_118.txt

pitcher_c_Page_131.txt

pitcher_c_Page_063.txt

pitcher_c_Page_178.txt

pitcher_c_Page_136.txt

pitcher_c_Page_180.txt

pitcher_c_Page_150.txt

pitcher_c_Page_061.txt

pitcher_c_Page_026.txt

pitcher_c_Page_015.txt

pitcher_c_Page_154.txt

pitcher_c_Page_075.txt

pitcher_c_Page_016.txt

pitcher_c_Page_058.txt

pitcher_c_Page_109.txt

pitcher_c_Page_167.txt

pitcher_c_Page_112.txt

pitcher_c_Page_076.txt

pitcher_c_Page_161.txt

pitcher_c_Page_066.txt

pitcher_c_Page_103.txt

pitcher_c_Page_019.txt

pitcher_c_Page_023.txt

pitcher_c_Page_114.txt

pitcher_c_Page_148.txt

pitcher_c_Page_077.txt

pitcher_c_Page_027.txt

pitcher_c_Page_083.txt

pitcher_c_Page_147.txt

pitcher_c_Page_040.txt

pitcher_c_Page_102.txt

pitcher_c_Page_081.txt

pitcher_c_Page_001.txt

pitcher_c_Page_065.txt

pitcher_c_Page_174.txt

pitcher_c_Page_122.txt

pitcher_c_Page_135.txt

pitcher_c_Page_181.txt

pitcher_c_Page_046.txt

pitcher_c_Page_059.txt

pitcher_c_Page_116.txt

pitcher_c_Page_166.txt

pitcher_c_Page_196.txt

pitcher_c_Page_009.txt

pitcher_c_Page_190.txt

pitcher_c_Page_056.txt

pitcher_c_Page_186.txt

pitcher_c_Page_097.txt

pitcher_c_Page_191.txt

pitcher_c_Page_073.txt

pitcher_c_Page_087.txt

pitcher_c_Page_172.txt

pitcher_c_Page_160.txt

pitcher_c_Page_062.txt

pitcher_c_Page_171.txt

pitcher_c_Page_168.txt

pitcher_c_Page_117.txt

pitcher_c_Page_064.txt

pitcher_c_Page_034.txt

pitcher_c_Page_127.txt

pitcher_c_Page_105.txt

pitcher_c_Page_100.txt

pitcher_c_Page_145.txt

pitcher_c_Page_044.txt

pitcher_c_Page_007.txt

pitcher_c_Page_090.txt

pitcher_c_Page_189.txt

pitcher_c_Page_005.txt

pitcher_c_Page_049.txt

pitcher_c_Page_133.txt

pitcher_c_Page_125.txt

pitcher_c_Page_113.txt

pitcher_c_Page_164.txt

pitcher_c_Page_024.txt

pitcher_c_Page_144.txt

pitcher_c_Page_095.txt

pitcher_c_Page_045.txt

pitcher_c_Page_010.txt

pitcher_c_Page_123.txt

pitcher_c_Page_070.txt

pitcher_c_Page_029.txt

pitcher_c_Page_142.txt

pitcher_c_Page_132.txt

pitcher_c_Page_082.txt

pitcher_c_Page_096.txt

pitcher_c_Page_074.txt

pitcher_c_Page_041.txt

pitcher_c_Page_071.txt

pitcher_c_Page_085.txt

pitcher_c_Page_038.txt

pitcher_c_Page_032.txt

pitcher_c_Page_115.txt

pitcher_c_Page_022.txt

pitcher_c_Page_021.txt

pitcher_c_Page_108.txt

pitcher_c_Page_187.txt

pitcher_c_Page_128.txt

pitcher_c_Page_120.txt

pitcher_c_Page_152.txt

pitcher_c_Page_039.txt

pitcher_c_Page_156.txt

pitcher_c_Page_139.txt

pitcher_c_Page_198.txt

pitcher_c_Page_175.txt

pitcher_c_Page_140.txt

pitcher_c_Page_028.txt

pitcher_c_Page_057.txt

pitcher_c_Page_165.txt

pitcher_c_Page_030.txt

pitcher_c_Page_099.txt

pitcher_c_Page_091.txt

pitcher_c_Page_050.txt

pitcher_c_Page_067.txt

pitcher_c_Page_149.txt

pitcher_c_Page_107.txt

pitcher_c_Page_080.txt

pitcher_c_Page_094.txt

pitcher_c_Page_151.txt

pitcher_c_Page_086.txt

pitcher_c_Page_137.txt

pitcher_c_Page_011.txt

pitcher_c_Page_119.txt

pitcher_c_Page_192.txt

pitcher_c_Page_143.txt

pitcher_c_Page_052.txt

pitcher_c_Page_155.txt

pitcher_c_Page_159.txt

pitcher_c_Page_013.txt

pitcher_c_Page_173.txt

pitcher_c_Page_121.txt

pitcher_c_Page_089.txt

pitcher_c_Page_195.txt

pitcher_c_pdf.txt

pitcher_c_Page_158.txt

pitcher_c_Page_177.txt

pitcher_c_Page_008.txt

pitcher_c_Page_183.txt

pitcher_c_Page_093.txt

pitcher_c_Page_157.txt

pitcher_c_Page_055.txt

pitcher_c_Page_184.txt

pitcher_c_Page_054.txt

pitcher_c_Page_014.txt

pitcher_c_Page_035.txt

pitcher_c_Page_111.txt

pitcher_c_Page_006.txt

pitcher_c_Page_110.txt

pitcher_c_Page_163.txt

pitcher_c_Page_169.txt

pitcher_c_Page_182.txt

pitcher_c_Page_033.txt

pitcher_c_Page_047.txt

pitcher_c_Page_153.txt

pitcher_c_Page_003.txt

pitcher_c_Page_031.txt

pitcher_c_Page_098.txt

pitcher_c_Page_162.txt

pitcher_c_Page_012.txt

pitcher_c_Page_042.txt

pitcher_c_Page_141.txt

pitcher_c_Page_004.txt

pitcher_c_Page_170.txt

pitcher_c_Page_197.txt

pitcher_c_Page_199.txt

pitcher_c_Page_051.txt

pitcher_c_Page_072.txt

pitcher_c_Page_130.txt

pitcher_c_Page_188.txt

pitcher_c_Page_176.txt

pitcher_c_Page_194.txt

pitcher_c_Page_193.txt

pitcher_c_Page_084.txt

pitcher_c_Page_134.txt

pitcher_c_Page_037.txt

pitcher_c_Page_017.txt

pitcher_c_Page_104.txt

pitcher_c_Page_124.txt

pitcher_c_Page_069.txt

pitcher_c_Page_036.txt

pitcher_c_Page_060.txt

pitcher_c_Page_043.txt

pitcher_c_Page_106.txt

pitcher_c_Page_146.txt

pitcher_c_Page_002.txt

pitcher_c_Page_078.txt

pitcher_c_Page_179.txt

pitcher_c_Page_126.txt

pitcher_c_Page_101.txt

pitcher_c_Page_079.txt

pitcher_c_Page_048.txt


Full Text












QUANTITATIVE METRICS TO EVALUATE IMAGE QUALITY FOR COMPUTED
RADIOGRAPHIC IMAGES















By

CHRISTOPHER D. PITCHER


A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL
OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT
OF THE REQUIREMENTS FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY

UNIVERSITY OF FLORIDA


2004
































To my wife, Brandy Pitcher, for all of her love and support and to my parents, Al and Sue
Pitcher, who have made this journey possible.















ACKNOWLEDGMENTS

I would like to give my most heartfelt thanks to everyone who has assisted me in

the completion of this work. In particular, I would like to give a special thanks to Dr.

David Hintenlang, chairman of my supervisory committee, and Dr. Manuel Arreola, co-

chairman of my supervisory committee. I thank them both for their continued support and

encouragement. I would never have been able to complete this work without their

expertise and guidance. I would also like to give special thanks to the members of my

supervisory committee, Dr.Wesley Bolch, Dr. Kathleen Hintenlang, Dr. Zhihui Fang, Dr.

William Properzio and Dr. Jonothan Williams, for their support and understanding. I

would also like to extend my deepest appreciation to Dr. Lynn Rill for serving as an

alternate committee member at my oral defense on such short notice.

I would like to thank the students of the Nuclear and Radiological Engineering

Department with whom I have had the pleasure of working for the past three years. I

would especially like to thank James Brindle for the use of his kitchen table. Without that

piece of furniture the qualifying exam would have been the end of my scholastic career.

I would also like to thank the United States Army Medical Department for giving

me this opportunity as well as continued funding of my research.

Finally I would like to thank my family for all of their love and support. Most

importantly I would like to thank my parents, Al and Sue Pitcher, for instilling in me the

values and morals to succeed in life; and my wife Brandy, whom I love and adore with all

of my being, for her unwavering love, devotion, support and encouragement.

















TABLE OF CONTENTS

page

A C K N O W L E D G M E N T S ................................................................................................. iii

LIST OF TABLES ...... .... .... .................... ......... .......... ............ vii

LIST OF FIGURES ......... ......................... ...... ........ ............ xi

ABSTRACT ........ .............. ............. .. ...... ..................... xiv

CHAPTER

1 IN TRODU CTION ................................................. ...... .................

2 BACKGROUND INFORMATION ................................... ....................................8

C om puted R adiography ......................................... ................................................. 8
Picture Archiving and Communications System .......................... ................. 10
M athematical Transforms ........................................ ....... ............... 1
Fourier Transform ...................................................... .. ... ........... ....... 11
R adon Transform ............. ........................................ ............ 12
N oise Pow er Spectrum ................................................. ............ 13
Modulation Transfer Function ......................... ..... ..............16
Detective Quantum Efficiency ...........................................................................17
E exposure ......... .............. ................................... ...........................18
C ontrast-D etail A naly sis................................................ ....................................... 18
V isu al P perception ............ .......................................................... .. .... ...... ... 18
Contrast-to-Noise Ratio ......... ............... ...... ............... 19
C ontrast-D etail Phantom s......................................................... ............... 20
Observer Perception Study ........................................................... ..................... 21
C N R C calculations ........... ....... .......... ................................. 2 1
M easurem ent of D ose .................................... ....................... ............... 22
A nthropom orphic Phantom ........................................ ........................... 23
Dosimeter ....................................................... 23
Determination of Effective Dose........................ ......................... 26

3 LITERATURE REVIEW .......................................................................... 33

N oise P ow er Spectrum ..................................................................... ....................33









M odulation Transfer Function ......... .................................................. ...................35
Detective Quantum Efficiency ............................................................................37
C ontrast-D etail A naly sis.......................................... ............................. ........... 8
M O SFE T D osim etry .......... ..... ............................................................ .. .... ....... 4 1

4 EXPERIMENTAL EQUIPMENT......................................... ......................... 42

Im aging Sy stem ................................................................42
Software .......... ......... .. ........................ ...............42
N P Sc, M T F c and D Q E c ..................................................................... ..................43
Clinical Contrast D etail Score ............................................................................. 44
D o sim etry .............................................................................4 5

5 EXPERIMENTAL METHODOLOGY........................ ... ......................... 48

Im ag e A cq u isitio n .......................................................................................... .. 4 8
E xperim ental Setup .............................. ........................ .. ........ .... ............4 8
A acquisition P aram eters.......................................................................... .... ... 49
Im age R etrieval ............................ ....... ............. .. ........... 50
C clinical N oise Pow er Spectrum ............................................................. ............... .... 50
Variation with Peak-Tube Potential and Current-Time Product.........................50
V ariation w ith Processing Option ............................................ ............... 51
Clinical M odulation Transfer Function ........................................... .....................51
Edge Angle Verification....................... ...... ...... .. ...................51
Variation with Peak-Tube Potential, Current-Time Product and Processing
O p tio n ................. ................................................................5 2
Clinical Detective Quantum Efficiency........................ ....................................52
Clinical Contrast D etail Score ............................................................................. 55
O b serve er Study ..............................................................55
CNR Calculations ........... .................................... .... ................56
V ariation w ith Processing Option ............................................ ............... 57
Anthropomorphic Phantom Viewer Study ...................................... ............... 57
D o sim etry .......................................................................... 5 8

6 RESULTS AND DISCU SSION ........................................... .......................... 65

N PSc ........................... ...... .... ........................ ........ 65
Variation with Current-Time Product ...................................... ............... 65
Variation with Peak-Tube Potential ....................................... ............... 66
V ariation w ith Processing Option ............................................ ............... 67
M T F c ......................... ..........................................................6 7
Scan V ersus Subscan D irection....................................... ......... ............... 68
IM TFc R eproducibility ............................................. .............................. 68
Edge A ngle R eproducibility ...................................................... ..... .......... 69
N um ber of E SF D ata Points ........................................ .......................... 69
Variation with Current-Time Product ...................................... ............... 70
Variation with Peak-Tube Potential ....................................... ............... 72


v









V ariation w ith Processing Option ............................................ ............... 73
Dynam ic Im age M manipulation ........................................................ ............... 74
D Q E c ..............................................................................7 7
Variation with Current-Time Product ...................................... ............... 78
Variation with Peak-Tube Potential ....................................... ............... 79
V ariation w ith Processing Option ............................................ ............... 79
C D S c .............................................................. ................ 8 0
O b server Study .....................................................................80
CN R T D term nation .................. ........ ..................................... ............... 82
Automated Phantom Scoring The Average Observer ......................................84
Automated scoring reproducibility.................... ... ...................... 84
Variation with current-tim e product................................. ............... 85
Variation with peak-tube potential ..................................... ......... ......... 86
Variation with processing option ...................................... ............... 86
Automated Phantom Scoring The Ideal Observer................ .....................86
Anthropomorhic Phantom Evaluation ............................................. ............... 88
E effective D ose C alculation ......... ..................................................... .............. 89

7 CONCLUSIONS AND FUTURE WORK .......................................... .........114

APPENDIX

A INPSc MATLAB M-FILE AND DESCRIPTION...........................................125

B IMTFc MATLAB M-FILE AND DESCRIPTION....... ...................................134

C MATLAB M-FILE FOR CONTRAST-DETAIL PHANTOM SCORING AND
DESCRIPTION ...................................... .......................142

D M C N P IN P U T F IL E S .................................................................... ..................... 155

E RESULTS OF THE POST-HOC STATISTICAL TESTS .....................................166

LIST OF REFEREN CES ........................................................... .. ............... 179

BIOGRAPHICAL SKETCH ............................................................. ............... 184
















LIST OF TABLES


Table pge

2-1 ICRP Publication 60 tissue weighting factors .................. ...............32

5-1 Technique factors for the current-time product variation study...............................62

5-2 Technique factors for the peak-tube potential variation study................................63

5-3 Processing options com pared ................................. ............... ............... 63

5-4 Published mass energy-absorption coefficients for air. ........................................63

5-5 Technique factors for the TO.10 phantom. ................................... ............... 63

5-6 Technique factors for the UF Radiology Phantom................................. ..........64

5-7 Five-point scale for subjective image quality evaluation.................. ........... 64

6-1 Average INPScs and their associated standard deviations...................................106

6-2 Average INPScs and their associated standard deviations...................................106

6-3 Average INPScs and their associated standard deviations...................................106

6-4 Edge comparison for the IMTFc (mm -) in the scan and subscan directions......... 107

6-5 Calculation based variability in the IMTFc........... ..................................107

6-6 Image based variability in the IM TFc......................................... ............... 107

6-7 Edge angle in degrees for repeated setups of the edge device and repeated
im ages of the sam e setup............................................... ............................. 107

6-8 IMTFcs for various ESF lengths in the scan direction ............. ................108

6-9 IMTFcs for various ESF lengths in the subscan direction. ...................................108

6-10 Average IMTFcs and their associated standard deviations.............................. 108

6-11 The INPSc and IMTFc value for different window and level settings. ................108









6-12 The value of Q for each peak-tube potential. ................................................. 109

6-13 In bucky exposure measurements at 60 kVp for the DQEc variation with
current-tim e product study. ............................................ ............................ 109

6-14 In bucky exposure measurements for the DQEc variation with peak-tube
potential study. ......................................................................109

6-15 Number of visible objects in the TO.10 phantom. ............................................109

6-16 Number of visible objects in the UF Radiology phantom............................... 110

6-17 ANOVA p-values for total score comparisons by observer group. .......................110

6-18 ANOVA p-values for total score comparisons by viewing condition .................10

6-19 Average threshold object for the TO. 10 phantom ................................................. 110

6-20 Average threshold object for the UF Radiology phantom. ................................ 111

6-21 CNRT for each object size and each phantom ..................................................111

6-22 Observer study and CDSc comparison.............. ....... ........................... 111

6-23 Variation of the CDSc with current-time product.............................. ..............12

6-24 Radiology residents evaluation of anthropomorphic phantom images ................12

6-25 The exposed bone sites and their associated BMFs and BSFs.............................112

6-26 Mass energy-apsorption coefficients for the four tissue types at 35 keV. .............112

6-27 Calculation of the Effective dose at 60 kVp.......................................................... 113

6-28 Calculation of the absorbed dose to the bone marrow. ................ .....................113

6-29 Calculation of the absorbed dose to the bone surface. ................ .....................113

6-30 Effective doses for clinical pediatric current-time product levels..........................113

7-1 Metrics evaluated at 60 kVp, 1.0 mAs with the chest PA processing option. .......123

7-2 Object diameters in image pixels for each row of both
contrast-detail phantom s. ............................................. .............................. 123

7-3 Physical dimensions in mm of the objects in the proposed contrast-detail
phantom using drilled holes for subject contrast..............................124









7-4 Physical dimensions in mm of the objects in the proposed contrast-detail
phantom using lead disks for subject contrast....................................................... 124

C-l The column and row (x,y) indicies for the objects in the TO. 10 phantom. ...........153

C-2 Diameter of objects in each row ................. ..................................... 154

E-1 Post-hoc INPSc mean comparisons for variations with current-time product.......166

E-2 Post-hoc INPSc mean comparisons for variations with peak-tube potential. ........167

E-3 Post-hoc INPSc mean comparisons for variations with processing option ..........167

E-4 Post-hoc scan IMTFc mean comparisons for variations with
current-tim e product. ...................................................................... ...................168

E-5 Post-hoc subscan IMTFc mean comparisons for variations with current-time
product .......................................................................... ........ ....... 169

E-6 Post-hoc scan IMTFc mean comparisons for variations with peak-tube
potential ..................................... .......................... .... ..... ........ 170

E-7 Post-hoc subscan IMTFc mean comparisons for variations with
peak-tube potential. .......................... ........................... .. ...... ............. .. 170

E-8 Post-hoc IMTFc mean comparisons for variations with processing option in the
sc an d ire ctio n ................................................................... ................ 17 1

E-9 Post-hoc IMTFc mean comparisons for variations with processing option in the
subscan direction ........................................................................................ .......... 171

E-10 Post-hoc scan IDQEc mean comparisons for variations with
current-tim e product. ...................................................................... ...................172

E-11 Post-hoc subscan IDQEc mean comparisons for variations with current-time
product................ .......... ................... ....................... 173

E-12 Post-hoc scan IDQEc mean comparisons for variations with peak-tube
potential .............. ....... ............................................................ ....... 174

E-13 Post-hoc subscan IDQEc mean comparisons for variations with
peak-tube potential. .......................... ........................... .. ...... ............. .. 174

E-14 Post-hoc IDQEc mean comparisons for variations with processing option in the
scan direction ......... ... ........................... ................................................ 175

E-15 Post-hoc IDQEc mean comparisons for variations with processing option in the
subscan direction. ................. ........... .. ......... ....... ......... 175









E-16 Post-hoc CDSc mean comparisons for variations with current-time product
for the TO 10 phantom ...... ........................... ......................................... 176

E-17 Post-hoc CDSc mean comparisons for variations with current-time product
for the U F R adiology phantom ..................................... ............................ ........ 177

E-18 Post-hoc CDSc mean comparisons for variations with processing option for the
T O 10 p h an to m ............................................... ............... ................ 17 8

E-19 Post-hoc CDSc mean comparisons for variations with processing option for
the UF Radiology phantom ........................................... ............................ 178
















LIST OF FIGURES


Figure pge

2-1 The process of photostimulable luminescence in a PSP. ........................................27

2-2 Optical spectra used in CR (adapted from Bushberg)............................................27

2-3 Representation of two images (a and b) and their associated Fourier transforms
(c a n d d ) ......................................................................... 2 8

2-4 Graphical representation of the Radon transform. ............................................28

2-5 Representation of an ideal edge profile and an ideal slit.......................................29

2-6 Radiograph of the UF radiology phantom.................. .... ...................... 29

2-7 Radiograph of the TO.10 phantom .......... .... .............................. .... ........... 30

2-8 Radiograph of Bower's anthropomorphic phantom.............................. .............30

2-9 Cross-section of a p-type MOSFET (adapted from Zeghbroeck). .........................30

2-10 MOSFET dosimetry system manufactured by Thomson and Neilson...................31

2-11 Figure showing actual size of the active region of the MOSFET dosimeter ..........31

4-1 One-year old Bower stylized anthropomorphic phantom. .......................................45

4-2 View of the top of the phantom's trunk showing the MOSFET access ports and
the spine (top-center)......... ............................................................ ...... .... ..... 46

4-3 View of the bottom of the phantom's trunk. ......................................................46

4-4 View of the bottom of the phantom's head ............................. .... ...........47

5-1 Im age acquisition setup. ............................................ ..........................................60

5-2 Image acquisition setup for the flat-field images................. ............... ..............60

5-3 Device to ensure proper angulation of edge device. ..............................................61

5-4 Final setup of the edge device for the determination of the MTFc. .........................61









5-5 The Agfa imaging plate inside the open imaging cassette........................................62

5-6 Idealized graphical representation of the grid. ... ................ ................62

6-1 Variation of the INPSc with current-time product at 60 kVp............. ..................90

6-2 Variation of the INPSc with peak-tube potential for a constant exposure level. .....90

6-3 Variation of the INPSc with processing option.......................................................91

6-4 Relative pixel values across the edge from an image acquired at 60 kVp and 1.0
m A s. ............................................................. ................ 9 1

6-5. MTFc for 25 and 128 data points in the scan direction from an edge image
acquired at 60 kVp and 1.0 m A s. ....................................................................... 92

6-6 Variation of the IMTFc in the scan direction with current-time product ...............92

6-7 Variation of the IMTFc in the subscan direction with current-time product. .........93

6-8 Five MTFc curves in the scan direction from five separate edge images acquired
at 0 .4 m A s. ........................................................ ................ 9 3

6-9 Five MTFc curves in the scan direction from five separate edge images acquired
at 3 .2 m A s. ........................................................ ................ 94

6-10 Smoothing effect of averaging the MTFc before integration...............................94

6-11 Average MTFc in the scan direction calculated from five images acquired at
both 0.4 m A s and 3.2 m A s ............................................... ............................ 95

6-12 Variation of the IMTFc in the scan direction with peak-tube potential for a
constant exposure level. .............................. ................ ................ ............. 95

6-13 Variation of the IMTFc in the subscan direction with peak-tube potential for a
constant exposure level. .............................. ................ ................ ............. 96

6-14 Change in the ESF with peak-tube potential. ........................... ................................ 96

6-15 Variation of the IMTFc in the scan direction with processing option. ..................97

6-16 Variation of the IMTFc in the subscan direction with processing option ..............97

6-17 ESFs for the three processing options in the scan direction...............................98

6-18 MTFC for the full-range and the hand AP processing options ..................................98

6-19 The MTFc for different window and level settings in the scan direction ...............99









6-20 Variation of the IDQEc with current-time product in the scan direction ...............99

6-21 Variation of the IDQEc with current-time product in the subscan direction. ........100

6-22 Variation of the IDQEc with peak-tube potential in the scan direction...............100

6-23 Variation of the IDQEc with peak-tube potential in the subscan direction. ..........101

6-24 Variation of the IDQEc in the scan direction with processing option....................101

6-25 Variation of the IDQEc in the subscan direction with processing option..............102

6-26 Variation of the DQEc in the subscan direction with processing option .............102

6-27 CNRT for each object size as a function of acquisition current-time product for
the T O 10 phantom ........... ... ........................................................ .. .... ... ... .. 103

6-28 CNRT for each object size as a function of acquisition current-time product for
the UF Radiology phantom ........................................... ............................ 103

6-29 Variation of the CDSc with processing option for the TO. 10 phantom................104

6-30 Variation of the CDSc with processing option for the UF Radiology phantom. ...104

6-31 Repeated CNR calculations of two adjacent background regions ......................105

6-32 The CDSc for the ideal observer and the TO. 10 phantom.............. ................105

6-33 The CDSc for the ideal observer and the UF Radiology phantom.........................106

B-l Input image for the IM TFC m-file..... ............................................ 139

B-2 Representation of the extracted usable edge data and the coordinate transform.... 140

B-3 MTFc of an ideal three-degree binary edge with a bin width of 0.5 ....................140

B-4 Binned data for the three-degree binary edge. ....................................... .......... 141

C-1 Geometrical relationship used in the localization of an object. ...........................152

D-l A geometrical depiction of the cells defined in the table top input file ................164

D-2. A geometrical depiction of the cells defined in the image receptor input file .......165















Abstract of Dissertation Presented to the Graduate School
of the University of Florida in Partial Fulfillment of the
Requirements for the Degree of Doctor of Philosophy

QUANTITATIVE METRICS TO EVALUATE IMAGE QUALITY FOR COMPUTED
RADIOGRAPHIC IMAGES

By

Christopher D. Pitcher

December 2004

Chair: David E. Hintenlang
Cochair: Manuel M. Arreola
Major Department: Nuclear and Radiological Engineering

Traditional methods of evaluating a computed radiography (CR) imaging system's

performance. The noise power spectrum (NPS), the modulation transfer function

(MTF), the detective quantum efficiency (DQE) and contrast-detail analysis) were

adapted in order to evaluate the feasibility of identifying a quantitative metric to evaluate

image quality for digital radiographic images. The addition of simulated patient scattering

media when acquiring the images to calculate these parameters altered their fundamental

meaning. To avoid confusion with other research they were renamed the clinical noise

power spectrum (NPSc), the clinical modulation transfer function (MTFc), the clinical

detective quantum efficiency (DQEc) and the clinical contrast detail score (CDSc). These

metrics were then compared to the subjective evaluation of radiographic images of an

anthropomorphic phantom representing a 1-year-old pediatric patient.









Computer algorithms were developed to implement the traditional mathematical

procedures for calculating the system performance parameters. In order to easily compare

these three metrics, the integral up to the system Nyquist frequency was used as the final

image quality metric. These metrics are identified as the INPSc, the IMTFc and the

IDQEc respectively. A computer algorithm was also developed, based on the results of

the observer study, to determine the threshold contrast to noise ratio (CNRT) for objects

of different sizes. This algorithm was then used to determine the CDSc by scoring images

without the use of observers.

The four image quality metrics identified in this study were evaluated to determine

if they could distinguish between small changes in image acquisition parameters (e.g.,

current-time product and peak-tube potential). All of the metrics were able to distinguish

these small changes in at least one of the image acquisition parameters, but the ability to

digitally manipulate the raw image data made the identification of a broad indicator of

image quality not possible. The contrast-detail observer study revealed important

information about how the noise content in an image affects the low-contrast detectability

of different sized objects. Since the CNRT for each object size in the contrast-detail

phantoms was almost independent of the exposure level, the minimum CNRT that would

be necessary for an object of that size to be "visible" in a clinical image was identified.

Finally, in order to determine more refined CNRT values (due to possible observer biases

from the physical construction of the contrast-detail phantoms available for this study)

the design of new contrast detail phantoms is proposed.














CHAPTER 1
INTRODUCTION

Image quality is one of the most important aspects of diagnostic radiology. The

concept of image quality has been undergoing a transformation with the widespread use

of digital-projection radiography. Imaging modalities, such as computed radiography

(CR), are beginning to replace standard screen-film imaging systems. While many

aspects of the imaging system remain unchanged, the processing of the image receptor

and the viewing environment for the resulting radiographic images are significantly

different. The radiologist is no longer limited to the single set of data represented by the

light transmitted through a piece of film. They now have the ability to digitally

manipulate the image data set and have access to the full dynamic range of information

contained in that data set.

Traditionally, radiographic image quality is evaluated in two separate ways,

quantitatively or qualitatively. Quantitative methods, such as the calculation of the noise

power spectrum (NPS) or the modulation transfer function (MTF), are tests of the

imaging system performance. They convey the ability of the system to transfer the

anatomical or physiological information of the patient to a radiographic image,

henceforth referred to simply as an image. These two quantities can be combined to

produce the detective quantum efficiency (DQE), which is the currently accepted

standard for imaging system performance. The images used to calculate these metrics are

acquired under conditions that allow the evaluation of the imaging chain without the

interference of additional scattering media (e.g., a patient, exam table, or grid) between









the x-ray source and the image receptor. Semi-qualitative methods, such as contrast-detail

analysis, are tests of an imaging system's performance as perceived by human observers.

These observers subjectively evaluate, or score, images of contrast-detail phantoms. This

identifies the perceived contrast-detectability in images produced by the imaging system.

The results of this type of study can be used to establish threshold contrast-to-noise ratios

(CNRT) based on the size of the objects in the phantom to determine the total number of

visible objects. This type of study will provide a quantitative parameter based on a semi-

qualitative evaluation that does not have to be repeated. Therefore, it would be ideal to

develop a method of applying these evaluations of imaging system performance to a

perception of clinical image quality. System performance evaluations could be performed

with images acquired under more realistic clinical conditions (e.g., with the use of

additional scattering media between the x-ray source and the image receptor). These

quantitative parameters, in conjunction with qualitative evaluations of simulated clinical

images of a pediatric patient, could be used to infer the quality of an image that would be

produced by the imaging system with those same acquisition parameters (e.g., peak-tube

potential and current-time product).

The radiation dose a patient receives in order to produce an image is also a concern

and is especially important for pediatric patients. Pediatric patients may be at the greatest

risk of potential adverse radiation effects for the following reasons: their growing tissues

are more susceptible to radiation effects than mature adult tissues; their skeleton, an

organ of high radiation sensitivity, encompasses a greater fractional distribution of active

bone marrow; their greater post-exposure lifetime increases the possibility for any

radiation-induced effects to manifest; children may be uncooperative and are frequently









subject to a greater number of exposures than adult patients; and pediatric patients

generally have a larger fraction of their anatomy located within the x-ray field compared

to adults having similar exams and projections.'

Despite the precise testing of imaging system performance, the true test of the

quality of an image is whether or not the radiologist can extract the needed diagnostic

information from an image in order to make the correct diagnosis. Therefore, it is

necessary to identify the patient radiation dose that is necessary to produce an image of

minimum acceptable quality that still has the necessary diagnostic information. The most

effective way to do this would be to image patients with a known condition at

continuingly decreasing exposure levels until the diagnosis is no longer possible. This is

not a viable option for several reasons. Acquiring multiple images of the same patient

would dramatically increase the radiation dose, which is precisely what a study of this

type is trying to limit. In addition, any adverse effects that might occur from the radiation

exposure would be magnified for pediatric patients. An alternative to this method is to

use an anthropomorphic phantom in place of real patients. The disadvantages in this

approach are that the phantom only approximates the anatomy of a human and in most

cases there will be no disease present. The radiologist will then have to make a subjective

evaluation of the "quality" of the image. Despite these limitations, the image acquisition

parameters necessary to acquire an image of minimum acceptable quality could be

determined.

Once the acquisition parameters required to obtain an image of minimum

acceptable quality are determined from an anthropomorphic phantom study, quantitative

metrics can be calculated from images obtained with the same acquisition parameters. In









order to ensure that the images used to calculate these metrics are more representative of

a clinical image, additional scattering media will be placed between the x-ray source and

the image receptor to simulate the presence of a pediatric patient. The quantitative value

of these metrics can then be directly linked to the image acquisition parameters required

to produce an image of minimum acceptable quality determined from the

anthropomorphic phantom study. In the future, this will allow the performance of a

simple quantitative test instead of a lengthy observer study to determine the acquisition

parameters necessary to create an image of minimum acceptable quality for a specific

type of exam.

The main objective of this research is to evaluate the feasibility of developing a

quantitative metric that can be used to set image acquisition parameters for CR exams

without the use of complicated observer studies. A metric of this type will allow images

to be produced with the minimal dose to the patient that ensures the image is of sufficient

quality for a radiologist's evaluation. This work will investigate the use of three

traditional quantitative methods and one semi-qualitative method of evaluating

radiological-imaging system performance as a direct measure of clinical image quality:

the calculation of the system NPS, MTF and the DQE, as well as the evaluation and

scoring of contrast-detail phantoms. Other researchers have applied these measures of

system performance as a direct measure of image quality.29 These methods will be

applied specifically to pediatric CR imaging. Since the images used in this research to

calculate these performance parameters are not acquired under the same conditions as in

imaging system evaluation (with the inclusion of additional scattering media), they are

not strictly representative of their definition. The traditional mathematical procedures are









followed in the calculation of the following image quality metrics evaluated in this

research: the clinical noise power spectrum (NPSc), the clinical modulation transfer

function (MTFc) and the clinical detective quantum efficiency (DQEc). The introduction

of additional scattering media does not change the technical aspects of contrast-detail

analysis and the image quality metric derived from this analysis is the clinical contrast-

detail score (CDSc).

The objectives of this research are listed below. They will show if these new

metrics can be used as measures of clinical image quality, using images acquired under

non-traditional conditions, with an inherently digital imaging system.

A. To determine if the traditional methods of calculating the NPS, MTF and DQE can
be applied to images acquired in a non-traditional manner at diagnostic exposure
levels.

B. To quantify the variation of the NPSc and MTFc with current-time product and
peak-tube potential. Since the MTFc is not expected to vary strongly with current-
time product for a set peak-tube potential, it must be determined if the same is true
for a set current-time product and a varying peak-tube potential. If this is the case, it
may not be necessary to calculate the DQEc as a metric.

C. To quantify the effect of any computer processing of the raw data before the image
is viewed on the evaluations of the NPSc, MTFc and the CDSc.

D. To quantify the effect of grayscale inversion on the evaluation of the CDSc. Two
different contrast detail phantoms will be used in this work and each produces
contrast in a different way. One has attenuating objects (creating light objects on a
dark background) and the other radio-opaque objects or holes (creating dark objects
on a light background).

E. To determine if the calculation of the NPSc and the MTFc with the original image
data is directly applicable, as a measure of image quality, to an image displayed on
the radiologist's monitor that has been manipulated (e.g., the dynamic use of
window and level controls). Since the radiologist now has the ability to manipulate
an image during the viewing process, the data set that the radiologist sees has
changed (the data set available for diagnosis is now the brightness of each pixel on
the monitor). Since the image can be continuously changed, the information
available to calculate quantitative image quality metrics is not the same information
that a viewer is seeing.










F. To quantify the effect of this dynamic image manipulation on the determination of
the CDSc through observer studies.

G. To correlate the newly developed image quality metrics with a qualitative
evaluation of anthropomorphic phantom images and identify the radiation dose
associated with the exam parameters used to obtain those images.

The remainder of this dissertation provides a detailed description of the equipment

utilized, approach and methods used, tools developed, results and conclusions reached in

the search for a quantitative metric applicable to clinical image quality. Chapter 2

contains background information and defines critical terminology so that a reader new to

this area can fully understand the remaining chapters. Chapter 3 contains a review of the

current literature that describes the calculation of the previously mentioned system

performance parameters, their applications as measures of image quality and the various

methods used in conducting a contrast-detail study. Chapter 4 provides a detailed

description of the equipment, both hardware and computer software, used in this

research. Chapter 5 contains the experimental methodologies utilized to answer the

questions at the heart of the objectives of this research. Chapter 6 provides the results and

discussion of the experiments that were carried. Chapter 7 summarizes the findings of

this research and provides a description of the applicability of the image quality metrics,

as well as their implications to the research objectives, and then discusses future proposed

work related to this research.

In addition to the main text, four appendices have been included that give a detailed

description of the computer codes and program input files that were utilized. The text of

those codes and input files are provided, as well as a plain language description of the

construction and functionality of those files. It is this description that should assist the






7


reader in implementing the methodology developed through this work. Also included in

the appendices is a description of the benchmarking of the MATLAB codes that ensures

their proper functioning.














CHAPTER 2
BACKGROUND INFORMATION

This chapter first gives an overview of CR adapted from Bushberg10 and from

Siegel and Kolonder". There are many sources detailing the calculation of system

performance parameters and applying them as measures of image quality. These sources

have been compiled and a coherent description of the methods used to calculate an

imaging system's NPS, MTF and DQE is presented. There are many different

methodologies covered under the broad topic of contrast-detail analysis. A summary of

the different aspects of this type of analysis is described. Finally, the calculation of

effective dose is presented.

Computed Radiography

The use of a photostimulable phosphor (PSP) as an image receptor was introduced

in the 1970s but did not come into wide spread use until the turn of the century.10 The

term CR did not come into use until 1981 when Fuji introduced the PSP imaging plate

and named the technique FCR.3 The first commercial CR, the Model 101, was introduced

by Fuji into clinical practice in Japan in 1983.11

A typical CR imaging plate, or PSP, is made from a combination of BaFBr and

BaFI, doped with a small amount of europium (Eu). A PSP acts in a similar manner to the

phosphors used in screen-film radiography (e.g., Gd202S). The difference in the

functioning is the process of photostimulable luminescence. The intensifying screens

used in screen-film radiography emit visible light promptly upon exposure to x rays. A

PSP, while also promptly emitting a small amount of visible light upon exposure to x









rays, traps the majority of the absorbed x-ray energy. The PSP can then be stimulated

with a laser and the trapped x-ray energy is released. The details of this process are

graphically represented in Figure 2-1.

When x-ray energy is absorbed in a CR imaging plate, the electrons of the Eu

atoms (Eu+2) are excited to the conduction band and Eu+3 is produced. These excited

electrons migrate and are trapped by fluorine atoms (F ), which then become non-ionized

F. These are referred to as "F-centers." These trapped electrons form the latent image in

the imaging plate. When the exposed imaging plate is scanned by a red laser, the

electrons trapped in the F-centers are excited back to the conduction band where they

become mobile again. They then can de-excite back into the electron hole of the Eu+3 by

releasing blue-green visible light (see Figure 2-2). This light is collected by a fiberoptic

light guide, which directs the light to a photomultiplier tube where an electronic signal is

generated. This electronic signal is digitized and stored, creating an inherently digital

image. When the imaging plate is first read, not all of the stored energy is released. In

order to reuse the plate, it must be erased by exposure to a high intensity light source.1'11

There are many advantages to acquiring an image this way. First of all, there is no

longer a need for a film processor and there are no longer chemicals to worry about. This

reduces the workload on a number of fronts from mechanical maintenance to waste

disposal (e.g., silver recovery). There is no longer the need for a dark room which will

free up valuable space in what seems to be ever shrinking budgets and facilities. All of

this equates to an overall cost savings for the radiology department. A CR imaging plate

also has a much wider latitude, or dynamic range, than a screen-film system. This means

that a CR plate can produce a usable image over a much wider range of exposures than a









screen-film system. This can dramatically reduce the amount of retakes. This increased

latitude can also lead to problems because the abilities to overexpose and underexpose a

CR plate and still get a good image are both possible. This could lead to an unnecessary

increase in patient dose, a phenomenon called dose creep. While a problem in the early

days of CR, this problem has been addressed by the CR manufacturers. All major

manufacturers of CR systems provide an exposure indicator that allows the technologist

to monitor the exposure level to the plate. Nevertheless, by far the most important

advantage is the ability to digitally manipulate the image, which will aid the radiologist in

the interpretation of the image.

Picture Archiving and Communications System

A picture archiving and communications system (PACS) is an integrated system of

interfaces, computer networks, computer hardware and computer software for the storage

and transfer of images. Imaging systems need to be physically linked to the PACS. In the

case of CR, this is accomplished through the image plate reader. Once the plate is

scanned and the image digitized, it is sent over the local area network to a computer

quality assurance (QA) station. At the QA station the technologist can view the image,

with specialized viewing software, and ensure that it is acceptable for a radiologist's

evaluation. The image is sent to the QA station in a standard format known as Digital

Imaging and Communications in Medicine (DICOM). This format was sponsored by the

American College of Radiology and the National Electrical Manufacturers' Association.

The purpose of this standard image format was to overcome the difficulties of integrating

imaging system components from different manufacturers.

Images are sent from the QA stations over a wide area network for storage and

archiving. There are usually two levels of storage, a short term archive that allows quick









access to the images and a long term archive for the permanent storage of the images.

Database programs are used to manage the large number of images that need to be stored.

The images can then be accessed through the network by the physicians and viewed on a

local PC with specialized viewing software.10

Mathematical Transforms

In order to calculate the NPSc and the MTFc, some specialized mathematical

operations are required. The first is the Fourier transform. This transform converts the

information contained in an image into its frequency components. The second is the

Radon transform. This transform, used primarily in the reconstruction of computed

tomography images, can be used to detect lines in an image. Below is a mathematical, as

well as a conceptual description, of both of these mathematical operations.

Fourier Transform

The two dimensional Fourier transform of a functionf(x,y) is mathematically

represented by

I(u, v)= exp[-2i(ux + vy)]f(x)dxdy (2-1).

The Fourier transform can be conceptually described as a mathematical procedure

that identifies the magnitude and phase of sinusoidal variations in the intensity of the

pixel values as a function of spatial frequency. Spatial frequency is analogous to the

general term of frequency as applied to a time varying signal and has units of inverse

distance. The Fourier transform is best described by example. Figure 2-3 depicts an

image of a sinusoidal pattern (a) and the modulus of the Fourier transform of that image

(b). Note the three bright dots, or peaks, in the center of Figure 2-3-b. The center peak is

the zero frequency value, or DC component, and represents the average brightness across









the whole image. The other two peaks represent the sinusoid in the image. The magnitude

of those peaks is the difference in brightness between the dark and light areas. The

placement of them is the spatial frequency associated with the sinusoid (the inverse of the

distance between the peaks). The orientation of the sinusoid in the image correlates with

the orientation of the peaks in the Fourier transform relative to the zero frequency value.

This is shown in Figure 2-3-c and 2-3-d. In this case a tilted sinusoidal pattern creates a

tilted pair of peaks in the Fourier transform. There are two peaks representing a single

sinusoid because the Fourier transform produces a mirror image of itself creating

redundant information referred to as the negative frequency values. It is important to note

that the Fourier transform just does not identify a single sinusoidal variation, but

simultaneously breaks down the spatial function into a sum of sinusoids that exactly

represents the information contained in the image. Only sinusoids up to a certain

maximum frequency (the Nyquist frequency) can be represented by the Fourier transform

of a digital image. The Nyquist frequency is related to the size of the pixels in the image

as


Nyquist frequency = (2-2).
2 pixel size

Since the Fourier transform contains all of the information in the original image, the

inverse Fourier transform can be applied and the original image can be recovered.12

Radon Transform

The Radon transform is an integral transform that converts a function f(x,y) into a

set of projections p(r,O). Mathematically it can be represented as

p(r, 0)= f(x, y) (r- xcos y sin O)dxdy (2-3)









where 6 is the Dirac delta function. This process is graphically represented in Figure 2-4.

Each projection p, rotated at an angle 0, is a collection of all of the line integrals

perpendicular to the projection at all distances r from the origin. The complete Radon

transform is a collection of the projections for all angles.

Noise Power Spectrum

The NPS shows the ability of an imaging system to process noise as a function of

spatial frequency. The term power spectrum is somewhat of a misnomer in this context. It

is a carryover from electrical engineering. If x(t) is the voltage across (or current through)

a one-ohm resistor, the expectation value of the squared modulus of x(t) is the average

power dissipated. Since the NPS is simply the squared modulus of the Fourier transform

of an image function, the term power spectrum is used.13'14

The NPS can be thought of as the variance associated with a particular frequency

component of an image.4 For a flat-field image, the NPS is the variance associated with a

particular frequency component of the noise in that image. In most cases, noise may be

represented, or approximated, as a stationary Gaussian random process with zero-mean.

Therefore, all of the relevant statistical properties will be contained in the power

spectrum. Unfortunately, exact determination of a power spectrum would require a

perfectly measured, indefinitely long piece of a random function, and would require

indefinitely detailed computations.'3 We are therefore reduced to approximating the NPS.

The power spectrum of a stationary random process can be directly estimated

through the periodogram, which is given by

N 2
NPS(f) = 1 I(x)e J-,2xdx (2-4)
N 0









for a continuous function I(x) of length N.1315 With CR, we are dealing with a discrete

function I(x) that is represented by the pixel values. The NPS estimation then becomes

N-1 -J 2nk 2
NPS = Ax I(kAx)e N (2-5)
NAx) NAx K-o

where the term Ax, the pixel size, inside the squared modulus symbol arises from the

definition of the discrete Fourier transform (DFT).16 The 1/N term is due to Parseval's

theorem that states for discrete functions, the relationship between power as computed in

the spatial domain and as computed in the frequency domain is given by


-h2(k) = H(n)2 (2-6)
k=0 N=0

where H(n) is the Fourier transform of h(k).16 Equation 2-5 can then be simplified to read


NPS(f) = DFT[I(x)]2 (2-7).
N

This can be easily expanded to two dimensions as follows,


NPS(u, v) = NxN DFT[I(x, y)] (2-8)


where( ) represents an ensemble average over many NPS calculations from small

regions of interest (ROIs) in the same image.

In order to use Equation 2-8 to calculate the NPS, the data need to be prepared in

order to obtain the most accurate estimate of the NPS. Due to the nature of this

application of the NPS, only one image is available to calculate each NPS. Therefore, a

tradeoff must be made between the number of NPS calculations that are averaged and the

size of the ROI for each calculation. Utilizing the center 1024 x 1024 region of each









image to avoid nonuniformities near the edges makes sixty-four 128 x 128 ROIs

available for the calculation of the NPS.4

In order to reduce the effect on the low frequency components of the NPS due to

structure in the flat field image, the variation in exposure across the image due to the heel

effect must be removed. In order to accomplish this, a surface is fit to the data and must

be subtracted from the flat-field image.4'5'17

The final data preparation step is to minimize the effect of having only a finite data

set to perform the NPS calculation. A finite data set, mathematically, is a rectangular

truncation function multiplied by what would be the infinite data set.15 This causes a

function of the form sin(f)/fto be superimposed on the NPS in the frequency domain. In

order to reduce this effect, the data must be truncated with a weighting function that

slowly goes to zero at the Nyquist frequency. A function of the following form

1 1 2zc
h(x) = cos 0 2 2 X

called the Hanning function,16 multiplied by the data set before the Fourier transform is

calculated, will accomplish this goal.5'5 It is important that the Hanning function be

normalized so the mean-square value of H(x) is equal to unity in order to preserve the

magnitude of the NPS.2

In order to calculate the DQE, the one-dimensional NPS must be calculated from

the two-dimensional NPS. The one-dimensional NPS can be calculated from a thick slice

comprising four lines of data immediately adjacent to the frequency axes for the scan (u)

and sub-scan (v) directions, respectively. Each data point is assigned a frequency value

of u2 +v2 and then binned. Each bin is averaged to produce the one-dimensional NPS.4









Modulation Transfer Function

The MTF represents the maximum ability of an imaging system to transfer subject

contrast, to the final image, as a function of spatial frequency. There are multiple

methods for determining the MTF of an imaging system.4'18-21 The two most common are

the use of a sharp attenuating edge or an attenuating material with a narrow slit cut

through the material (see Figure 2-5). Since the images of these devices are already

digitized, the pixel values can be easily sampled in the direction of the arrows in Figure

2-5. If these pixel values are plotted, the sharp edge produces the system edge spread

function (ESF) and the narrow slit produces the system line spread function (LSF). The

LSF can also be derived from the ESF. The LSF of a system is the first derivative of the

ESF. The MTF is then the modulus of the DFT of the LSF normalized to the value at zero

frequency.

Since the image is comprised of discrete data, a correction needs to be made for the

effect of finite differentiation with the edge method. The correction is made in the spatial

frequency domain by multiplying the MTF by


Correction Function = (2-10)
since -L
2f)

wherefis frequency andf, is the Nyquist frequency.19

In order to sample the MTF at frequencies higher than the Nyquist frequency, the

slit or edge must be angled with respect to the pixel matrix. The ideal angle is between

one and six degrees. Before the pixel values can be sampled, the edge angle must be

determined. This can be done using the Radon transform. The distance of each pixel from

the edge, along a line perpendicular to the edge, must also be determined. This is done be









transforming the coordinates of the pixels into a coordinate system that is rotated by the

edge angle. The two coordinate systems are related by the following relationships:

Xrotated = X COs 0 + y sin 0 (2-11)

Rotated = y cos x sin 0 (2-12).

The ESF data can then be binned into intervals smaller than the pixel size.20'22 This allows

the calculation of the pre-sampled MTF.

The MTF calculation also suffers from the same problem of rectangular data

truncation as the NPS calculation. Therefore, the Hanning function is multiplied by the

data set before the Fourier transform is calculated. In this case it is important that the

Hanning function be normalized so the mean value of H(x) is equal to unity in order to

preserve the magnitude of the MTF.

Detective Quantum Efficiency

The DQE is a quantity used to describe the overall signal-to-noise ratio (SNR)

performance of a system. Currently, the DQE has become the standard by which digital

x-ray imaging systems are measured in the research environment. It is essentially the

ratio of the output SNR squared to the input SNR squared. The DQE can be calculated

from the MTF and the NPS as follows:


DQE(f) k[T(f)]2 (2-13)
Q.NPS(f)

where k corrects for the gain of the imaging system and Q is the number of photons

incident on the image receptor (in photons/mm2) used to generate the flat-field image.2'5'23









Exposure

The quantities k and Q combined have units of photons/mm2-mR. The

determination of this quantity is related to the calculation of exposure (X) for a specified

x-ray spectrum. Exposure is defined as the absolute value of the total charge of one sign

produced in air when all the electrons liberated by photon interactions in a given mass of

air are completely stopped in air. Exposure can be calculated from the following equation


X j= T. ( e (2-14)
PI Ear WJc

where is the energy fluence of the x-ray spectrum, (ten/p)E,air is the energy dependant

mass energy-absorption coefficient for air and (e/W) is a constant of 33.97 J/C that is

related to the amount of energy required to create an ion pair in air.24

Contrast-Detail Analysis

Visual Perception

The NPS, MTF and DQE are indicators of the performance of an imaging system

up to the point of storage of the final image. In medical imaging, a radiologist's

perception of the displayed visual information is used to make a diagnosis. The

perception of visual information consists of three sequential steps: detection, recognition

and interpretation. Contrast-detail analysis focuses on the task of detection. Various

models have been developed to describe how observers detect visual signals in images.

One approach is the use of a signal-to-noise model. The signal-to-noise model describes

the ability of an observer to detect simple visual signals embedded in a noisy background.

The model characterizes an "ideal observer" who detects signals with a certain likelihood

if their amplitude differs from the background by a set threshold.25 Hendee et al. states,

"Wagner and Brown26 have suggested that the performance of an imaging system can be









characterized quantitatively by describing the ability of an ideal observer to detect simple

visual signals provided by the system."(25 p 295)

There are two aspects to the detection of visual information, visual acuity and

contrast discrimination. Hendee et al.defines visual acuity as the ability of an observer to

extract detailed information from a small region of an image. This is the ability to detect

high spatial frequency signals. Visual acuity is often measured using the Snellen eye test

chart. This chart has black letters of varying sizes on a white background. If the grayscale

on the chart were to be reversed, the ability of observers to recognize the letters from a

distance is impaired. Deficiencies in visual acuity can be corrected with eyeglasses.

Hendee et al. then define contrast discrimination as the ability of the visual system to

distinguish differences in brightness in the image. This is the ability to detect middle and

low spatial frequency signals. Deficiencies in contrast discrimination suggest a neuro-

opthalmological cause. Hendee et al. conclude that, "Contrast discrimination is probably

a more critical feature than visual acuity in determining how well the average person

'sees'."(25 p 297)

Contrast-to-Noise Ratio

The signal in traditional screen film radiography can be represented by a quantity

called radiographic contrast. Radiographic contrast is the difference between the average

optical densities of the region where the object of interest providing the signal is located

in the background.

Radiographic Contrast = ODgna ODBackground (2-15)

With the use of an inherently digital imaging modality such as CR, the imaging system

software often performs a series of processing steps before the image is viewed. One









common form of processing is the subtraction of a constant from all of the pixel values in

the image. This can lead to problems if traditional notions of contrast, such as subject

contrast that is described in the next section (see Equation 2-19), are used. If the constant

that is subtracted is equal to the average pixel value represented by N1, the subject

contrast becomes infinite. Therefore, a more meaningful and frequently used measure of

contrast is the contrast-to-noise ratio (CNR)

CNR = Signal Background (2-16)


where a is the standard deviation of the pixel values in the background.10

Contrast-Detail Phantoms

Many different forms of contrast-detail phantoms have been used to evaluate the

performance of imaging systems.'27-29 While the designs may differ, all have one thing in

common: the presence of circular objects of varying diameter and varying levels of

subject contrast. Subject contrast is the inherent difference in the x-ray attenuating

properties between two regions in an object. If the number of x-rays reaching the imaging

plate directly below the two regions of interest are N1 and N2

N, = Noe i1 (2-17)

N =Noe /2X2 (2-18)

then subject contrast is defined as


Subject Contrast = N1 N2 (2-19).
N,

While all contrast-detail phantoms have this design property in common there are

two distinct ways in which subject contrast can be designed in the phantom. The first is to

start with a phantom of uniform thickness and drill circular holes of varying depth. An









example of this type of phantom is the UF Radiology phantom (see Figure 2-6).27 A

radiograph of this phantom will produce an image with dark signals (circular objects) on

a light background. The second method is to insert circular objects with varying x-ray

attenuating properties. An example of this type of phantom is the Leeds test object

TO. 10' (see Figure 2-7). A radiograph of this phantom will produce an image with light

signals on a dark background.

Observer Perception Study

The first task in contrast-detail analysis is to use human observers to evaluate

images of contrast-detail phantoms. The images are acquired under the conditions that are

to be compared or evaluated. The observer is then asked to evaluate, or score, the images.

Depending on the design of the phantom, the observer is either asked to identify the

presence or absence of the object (or signal) in a specific region of the image or identify

the number of objects that are visible. The definition of what constitutes a present or

visible object is dependent upon the design of the study. The observer is instructed on

these criteria before the images are evaluated. Even with the most explicit instructions,

there is still an amount of subjectivity in each observer's interpretation of those

instructions. Therein lies the inherent qualitative nature in this type of study.

CNR Calculations

In an attempt to speed up the evaluation of quality assurance phantoms, several

researchers have developed computer algorithms to identify, localize and mathematically

evaluate objects and ROIs in those phantoms.3-32 These completely automated algorithms

depend either on some initial preparation of the digital image or an apriori knowledge of


'Leeds Test Objects Ltd, Wetherby Road, Boroughbridge, North Yorkshire, YO51 9UY, UK









the size of some or all of the objects in the image. The methods utilized involved the

convolution of a mask of the object of interest with part or all of the original image. For

the circular objects, the maximum value of the resulting convolution would be the point

of maximum correlation between the mask and the object (i.e., the center of the object).

These methods are extremely sensitive to the size of the object of interest in the image. If

the mask is not exactly the same size as the object, a precise center can not be located. In

order to make an algorithm more broadly applicable to geometrical magnification or

phantom rotation a manual localization procedure can be utilized. If two points in each

image are manually identified, all objects and ROIs can subsequently be successfully

located. Once the positions of the objects in the image have been successfully located,

mathematical operations such as the calculation of CNR can be performed.

Measurement of Dose

There are three aspects to the determination of the radiation dose a patient will

receive from a specific type of radiographic examination. First, an anthropomorphic

phantom that simulates the anatomy of the human body is constructed. This is done so

that the attenuated and scattered x-ray spectrum produced through interactions with the

patient's internal organs is reproduced. Secondly, the amount of absorbed x-ray energy to

those organs must be measured. This requires the use of very small radiation detectors (or

dosimeters) so that the x-ray spectrum in the anthropomorphic phantom is not

appreciably disturbed. Lastly, the amount of measured energy deposition in the dosimeter

must be converted to the amount of energy that would have been deposited in human

tissues, which can be used to calculate a dosimetric quantity.









Anthropomorphic Phantom

Mathematical models have been developed to simulate the anatomy of human

beings. The most widely used of these models was developed by Cristy.33 These models

are based on the size of the "average" person of varying age and gender. Bower34

produced a physical phantom based on the Cristy and Eckerman one-year old pediatric

model. It consists of three physical regions, head, trunk and legs. The head region is

modeled after the Bouchet and Bolch model.35 The phantom is composed of three tissue

substitute types: soft tissue, bone and lung. The composition of these tissue substitutes is

explained in detail by Bower.34 For organs used in the calculation of effective dose that

are internal, small holes are drilled in the phantom so a small dosimeter can be placed at

the centroid of the organs. A radiograph of the anthropomorphic phantom's trunk can be

seen in Figure 2-8.

Dosimeter

The radiation dosimeter needed to measure the amount of energy deposited in the

organs of a pediatric patient must be very small. Not only does a dosimeter have to be

small to measure the absorbed dose to pediatric organs, the measurement of total

accumulated dose requires the use of miniature dosimeters that will not perturb the

radiation field. The implicit size restriction narrows the choice of conventional detectors

to thermoluminescent dosimeters (TLDs) or semiconductor diode detectors.36 Accurate

surface dose measurements can be made with parallel plate ionization chambers, but

because of their size are not suitable for internal organ dose measurements.37 The TLD

measures cumulative dose, but the reading procedure is destructive in the sense that the

stored signal is lost after reading. Although small diodes provide dose rate information,

which can be integrated to yield the total dose for any time duration, the signal due to









total dose is not retained in the device and any signal lost can not be recovered.36 An

alternative option for absorbed dose measurements is the use of metal oxide-silicon

semiconductor field effect transistors (MOSFETs) as radiation dosimeters. The MOSFET

detector has a very small size, with commercially available systems having active

volumes typically 400 um x 500 um x 100 um.38'39 The MOSFET detector also offers the

unique advantage of permanently storing the total cumulative dose signal, which can be

read at any time in a nondestructive manner. This ensures that the dose information will

not be lost during the readout procedure. In addition, the acquired dose signal is dose rate

independent.40 These factors combine to make the MOSFET an ideal detector for

absorbed dose measurements in an anthropomorphic phantom.

The basic MOSFET structure is shown in Figure 2-9. This type of MOSFET is a p-

channel MOSFET that is built on a negatively doped (n-type) silicon substrate. Two of

the terminals of the MOSFET called the source and the drain are situated on top of a

positively doped (p-type) silicon region. The third terminal shown is the gate. Underneath

the gate is an insulating silicon dioxide layer and under this oxide layer is the n-silicon

substrate. The region of the substrate immediately below the oxide layer is known as the

inversion layer or channel region. When a sufficiently negative voltage is applied to the

gate, with reference to the substrate, a significant number of minority carriers (holes in

this case) will be attracted to the oxide-silicon surface from both the bulk of the silicon

and the source and drain regions. Once a sufficient number of holes have accumulated

there, a conduction channel is formed, allowing an appreciable amount of current to flow

between the source and the drain. The gate voltage needed to allow a predetermined

current flow is defined as the threshold voltage (Vth).4041









During irradiation, a positive voltage is applied to the gate. Electron-hole pairs are

generated within the silicon dioxide by the incident ionizing radiation. The electrons

quickly move toward the positively biased gate while the holes migrate toward the Si-

Si02 interface. When the holes get close to the interface, some of them are captured in

long-term trapping sites. These trapped positive charges cause a negative shift in Vth. The

magnitude of this shift is proportional to the radiation dose deposited in the oxide layer.40

The use of a single MOSFET as a dosimeter does have some limitations. A 1C

change in ambient temperature can shift Vth by as much as 4-5 mV. Also, the response of

a single MOSFET detector (Vth) as a function of accumulated dose will exhibit a

nonlinear region at high dose levels.40 Commercially available MOSFET dosimeters

consist of two identical MOSFETs fabricated on the same silicon chip. The two

MOSFETs are operated at two different positive gate biases during irradiation and the

difference between the threshold voltage shifts of the two MOSFETS is representative of

the absorbed dose. Since the response to temperature of each individual MOSFET is

identical, this type of construction renders minimal temperature effects. This type of

construction also reduces the nonlinear response at high dose levels.40

Many in\ estigators : 43 have reported an angular dependence in the sensitivity

of the MOSFET dosimeter. Pomije et al.43 investigated this behavior for diagnostic

exposure levels from 60 kVp to 120 kVp and found that a significant reduction in the

response of the MOSFET occurs when the epoxy bubble faces away from the x-ray

source. The black epoxy bubble covering the active MOSFET can be seen in Figure 2-10

and 2-11. These figures picture a commercially available MOSFET dosimetry system

manufactured by Thomson and Nielsen Electronics Ltd. This effect can be minimized if









care is taken to orient the MOSFET with the epoxy bubble facing the x-ray source during

dosimetric measurements.

Determination of Effective Dose

The currently accepted dosimetric quantity linking absorbed dose and the

probability of stochastic radiation induced effects is effective dose (E).44 Effective dose is

calculated from the following equation

E = w .wR D.,R (2-20)
T R

where R signifies the type of radiation, T signifies the type of tissue exposed, wRand wT

are the radiation and tissue weighting factors, respectively, and DT,R is the absorbed dose

to a particular tissue T from radiation type R. In order to calculate the effective dose to a

patient, the absorbed dose to the organs listed in Table 2-1 must be measured.

The first step in measuring absorbed dose using the MOSFET as a dosimeter is to

calibrate the system. This procedure is described by Bower and Hintenlang.42 This will

result in calibration factors that will be able to convert the MOSFET reading to exposure

in Roentgens (R). After the exposure of the phantom has been measured with the

dosimetry system, those measurements must be converted to absorbed dose. An exposure

of one R produces an absorbed dose of 0.876 rads in air. Therefore, the absorbed dose to

tissue can be calculated as follows,34


Dtssue =0.876.* (k-eP)o. de (en 1ssue .CF. V (2-21)
k enP low, en /oxlde

where /en /p is the average mass energy absorption coefficient for the specified material

in cm2/g, CF is the calibration factor in R/mV, and V is the MOSFET reading in mV.










Conduction Band



Red
Laser F F


X-ray
X-ray ABlue to Green
\A,/'* fe Photon

Eu+2 Eu+3


Valance Band

Figure 2-1. The process of photostimulable luminescence in a PSP.


Red





SBle- Gre en\



300 400 500 600 700 800

Wavelength (nm)


Figure 2-2. Optical spectra used in CR (adapted from Bushberg).1














a. b.




c.4 ,4 d.



Figure 2-3. Representation of two images (a and b) and their associated Fourier
transforms (c and d). Used with the permission of Dr. Stephen Lehar.12

_y




0 x


Figure 2-4. Graphical representation of the Radon transform.










Ideal edge profile


ESF


LSF


Figure 2-5. Representation of an ideal edge profile and an ideal slit. If the pixel values are
sampled in the direction of the arrows, the ESF and LSF can be determined.


Figure 2-6. Radiograph of the UF radiology phantom.


Ideal slit

























Figure 2-7. Radiograph of the TO. 10 phantom.


Figure 2-8. Radiograph of Bower's anthropomorphic phantom.


Figure 2-9. Cross-section of a p-type MOSFET (adapted from Zeghbroeck).41






















Figure 2-10. MOSFET dosimetry system manufactured by Thomson and Neilson.ii The
active MOSFETs are encased in the black epoxy bubbles at the tips of the
long brown strips at the bottom of the figure. The bias supply is the small box
labeled A in the right center of the figure while the readout device is at the top
of the figure.


Figure 2-11. Figure showing actual size of the active region of the MOSFET dosimeter.
The black rectangular region (3 mm in width) contains the MOSFETs.


" Thomson and Nielsen Electronics Ltd, 25B Northside Road, Ottowa, ON, Canada K2H 8S1









Table 2-1. ICRP Publication 60 tissue weighting factors.44
Tissue or Organ Tissue weighting factor
Gonads 0.2
Bone marrow (red) 0.12
Colon 0.12
Lung 0.12
Stomach 0.12
Bladder 0.05
Breast 0.05
Liver 0.05
Esophagus 0.05
Thyroid 0.05
Skin 0.01
Bone Surface 0.01
Remainder 0.05














CHAPTER 3
LITERATURE REVIEW

Noise Power Spectrum

The NPS has been calculated by many researchers for a wide variety of imaging

systems and image receptors. The choice of normalization of the NPS varies slightly from

researcher to researcher but the underlying methodology is consistent (the squared

modulus of the Fourier transform of a flat-field image). In all cases, the researchers are

attempting to determine the noise properties of the imaging system, or systems, in

question.

Flynn and Samei2 determined the two-dimensional NPS of a CR system utilizing

two different imaging plate resolutions. The measurements were made by exposing the

imaging plate to a spatially uniform x-ray beam while simultaneously measuring the

exposure level to the imaging plate. One-hundred and forty-four subregions in a 12 x 12

array were used in the NPS calculation. For the low resolution mode with pixels of 200

rim, the subregions were 128 x 128. For the high resolution mode with pixels of 100 rim,

the subregions were 256 x 256. Even though the x-ray field is assumed to be uniform, in

reality slight variations exist across the image receptor due to the heel effect. The

exposure measurement and the mean signal level in different regions of the image were

used to correct for large-scale nonuniformities (e.g. the heel effect). The data in each

subregion was truncated with the Hanning function before the two-dimensional Fourier

analysis generated the NPS. The subregions were averaged to produce the final NPS. The









one-dimensional NPS was estimated by averaging the central row or column and +5 rows

or columns from the two-dimensional NPS.

Fetterly and Hangiandreou5 calculated the NPS of a Lumisys ACR-2000 single

plate desktop CR reader with a method similar to Flynn and Samei.2 The primary

differences in their methodology were the size of the subregions used to calculate the

NPS and the subtraction of a planar fit from each subregion before calculation of the

NPS. Fetterly and Hangiandreou utilized 225 100 x 100 ROIs from each uniform field

image. This selection was stated to be a matter of convenience. The resultant NPS

measurements were judged to contain suffient detail and the frequency increments were

equal to those of their MTF measurements.

Dobbins et al.4 measured the NPS of four generations of CR imaging plates by

acquiring a flat-field image with 0.5-mm Cu added filtration. The two-dimensional NPS

was computed directly. The center 1024 x 1024 portion of a single image for each plate

was used to calculate the NPS in order to avoid nonuniformities near the edges. This

portion of the image was subdivided into 64 128 x 128 ROIs. They observed that this size

for the ROIs was the smallest size that could be used without appreciably changing the

shape of the average NPS curve near zero frequency. In order to remove the background

trends created by the heel effect, a planar ramp was fitted to the data of each ROI and

subtracted from the ROI before the Fourier analysis. All of the NPS calculations for each

ROI were averaged to generate the final NPS. The method used by Flynn and Samei2 to

generate the one-dimensional NPS from the two-dimensional NPS was found to be

insufficient by Dobbins et al.4 Instead, the pixels of a thick slice on either side of the

axes, excluding the axes, were binned to generate the one-dimensional NPS. The









frequency value assigned to each pixel was computed as the radial distance of that pixel

from the zero-frequency pixel.

Modulation Transfer Function

Multiple methods for determining the MTF are presented in the literature. The

methods vary from the use of a radio-opaque square, a thin slit cut into a radio-opaque

material, a small pin-hole cut into a radio-opaque material and a thin wire. In all cases an

image is acquired of one of the test devices listed above and a LSF is generated. This LSF

is then used to calculate the MTF through the use of the Fourier transform. Each method

is usually given a name that represents the test device used. For example, the edge

method would utilize the radio-opaque square and the edges of that square are used in the

calculation of the MTF.

Fetterly and Hangiandreou5 used an angled edge technique to determine the MTF of

a Lumisys ACR-2000 single plate desktop CR reader. The edge device was custom made

and consisted of a 250-[m thick square of lead laminated between two sheets of acrylic,

each 1 mm in thickness. A single image was used to evaluate the MTF in both the scan

and subscan directions. Fetterly and Hangiandreou processed the storage phosphors in a

dimly lit room with only the monitor of the Lumisys control workstation providing light.

This leads to the conclusion that the edge images were acquired with the bare imaging

plate and it was not in a normal cassette. The angle the edge device made with respect to

the pixel matrix was calculated to be between two and five degrees for all images. A 4-

cm x 8-cm region of the image encompassing the edge was extracted and the pixel values

were averaged in bins 0.2 times the pixel size. This produced an over-sampled ESF. The

resulting LSF was zero-padded so that the frequency values of the MTF matched those of

the NPS calculated in the same study. The MTF was then corrected for finite element









differentiation as described by Cunningham and Fenster.19 The formula used to calculate

this correction is given in Equation 2-10. Cunningham and Fenster state that the MTF

resulting from finite differentiation of the ESF contains an error that remains unimportant

only if the sampling rate is approximately four times the Nyquist frequency or greater.

Samei and Flynn2 constructed a sharp, attenuating edge device and designed a

procedure for the measurement of the MTF of a digital storage phosphor radiography

system. The edge device was constructed from a 250-im thick lead foil laminated

between two 1-mm-thick sheets of acrylic. Images of the edge device were acquired with

the presence of an additional three-mm of aluminum filtration between the x-ray source

and the image receptor. The edge device was angled between one and six degrees with

respect to the pixel matrix. An 8-cm x 8-cm region containing only the edge was

extracted from the image. In order to determine the precise edge angle, the data was

processed with the following steps; the gray-scale image is converted to a binary image

through a thresholding procedure, a binary line along the edge is produced through the

application of a gradient operation and the edge angle is then determined with the use of

the Hough transform. The data were then placed in bins with sub-pixel bin widths of 0.1

times the pixel width. The resulting LSF was then truncated with a Hanning filter to

eliminate the high-frequency content of the measurement not associated with the edge.

The pre-sampled MTF was calculated through Fourier analysis of the LSF. This method

was compared to a slit method similar to Fujita et al.21 It was determined that the edge

method provided similar results to the slit method.

The use of a narrow slit to determine the MTF of a CR imaging system is described

by Fujita et a.21 This method was also employed by Dobbins et al.4 Images of a 0.01-mm









slit that was angled less than 2 degrees with respect to the pixel matrix were acquired.

The slit was oriented almost perpendicular to the scanning direction. Four rows of data,

each five pixels in width, were used in the determination of the 'finely' sampled LSF.

The data was binned using the simple geometrical relationships according to their relative

positions. The Fourier transform of this LSF results in the presampled MTF.

Detective Quantum Efficiency

The majority of the researchers discussed above that measured an MTF also

measured an NPS in order to calculate a DQE. The basic definition of the DQE is

consistent throughout the literature.2,4,5,10,23 In order to calculate the DQE the ideal SNR

must be determined. This is the SNR of the x-ray beam incident upon the image receptor.

The ideal SNR is defined as the incident number of x-ray quanta.6'10'23 To calculate this,

the x-ray spectrum of the imaging system must be determined. Bradford et al. 1

eliminated the need for this calculation by using the same image acquisition conditions as

Dobbins et al.4 and utilizing the already published values. Launders et al.23 used computer

simulations based on published methods and tables. Flynn and Samei2 used a

semiempirical x-ray spectra model based on work by Storm.45 Dobbins et al.4 used the

ideal SNR value provided by the manufacturer. Fetterly and Hangiandreou5 estimated the

x-ray spectra used experimentally with the tungsten anode spectral model using

interpolating polynomials (TASMIP) program developed by Boone and Seibert.46

Fetterly and Hangiandreou5 investigated the effects of different x-ray spectra on the

DQE of a CR system. In order to obtain different quality x-ray spectra the beam was

filtered with various thicknesses of a patient equivalent phantom (PEP), aluminum and

copper. The PEP was constructed from 1mm of type 1100 aluminum secured between

two 2.48-cm thick pieces of acrylic. Depending on the peak tube potential up to six PEP









phantoms were used. In addition, up to 80 mm of aluminum and 8.1 mm of copper were

utilized. X-ray beams between 70 and 120 kVp were investigated. The term that corrects

for the system gain in the calculation of the DQE was eliminated by converting the image

pixel values to exposure values before the calculation of the NPS. This methodology was

also followed by Flynn and Samei.2

Contrast-Detail Analysis

Lu et al.29 utilized a custom made contrast detail phantom to compare computed

radiography (Kodak CR 400) and film-screen combination (Speed 400) systems in

regards to patient dose, technique settings and contrast-detail detectability. The phantom

was constructed from a 26.5-cm x 26.5-cm Lucite sheet that was 2.5-cm thick. Two

hundred and twenty-five holes (arranged in a 15 x 15 square pattern) were drilled in the

Lucite ranging from 0.3 to 8 mm in both diameter and depth. The phantom was imaged in

an under-table bucky with the use of a 10:1 grid. The phantom was placed on top of

additional Lucite sheets to simulate tissues that generate scattered radiation. The Lucite

and the phantom were placed on the examination table. Varying amounts of scattering

media were used and the source to image distance (SID) was adjusted so the geometric

magnification of the phantom in the image remained constant for different thicknesses of

scattering media. The CR images were printed onto film for a hard copy reading

comparison with the images acquired directly on film. In order to compare the use of CR

with a higher peak-tube potential x-ray beam, the soft copy CR images were read and the

data manipulation tools of the diagnostic workstation were employed. Four physicists

evaluated the images by scoring the threshold target depth for each object size. An object

detection ratio was calculated by dividing the number of detected objects by the total

number of objects in the phantom. Lu et al.29 concluded that using a higher peak-tube









potential setting and additional aluminum filtration would reduce the patient entrance

skin dose without compromising the contrast-detail detectability, which was compensated

by the contrast manipulation on soft-copy workstations.

Padgett and Kotre6 utilized the Leeds T020 test object to evaluate the effect of

"dead pixels" on image quality in direct digital radiography based on selenium detectors.

The Leeds T020 test object creates contrast by using x-ray attenuating objects so the

objects in the image are light compared to the background. Images of the contrast-detail

phantom were obtained from a Hologic Direct Ray EPEX system with the anti-scatter

grid in place. No additional scattering media was placed between the x-ray source and the

image receptor. Four experienced observers scored the images and the threshold contrast

(CT) was determined for each object size. The threshold contrast was the contrast of the

object at the threshold of visual detectability. The degradation in observer performance

was found to be similar to the reduction in the relative DQE (DQE normalized to zero

frequency) with the loss of active pixels. The loss of active pixels was simulated in the

images prior to scoring.

Aufrichtig4 and Aufrichtig and Xue2 investigated the low contrast threshold

detectability for an amorphous silicon x-ray detector designed for digital radiography and

a standard thoracic screen-film combination. The contrast-detail phantom utilized in this

study is the commercially available CDRAD contrast-detail phantom manufactured by

Nuclear Associates, Carle Place, NY. The phantom is constructed from a 26.5-cm x 26.5-

cm Plexiglass plate that is 1-cm thick. The Plexiglass plate is divided into 255 separate

square regions (15 rows x 15 columns) of equal size. There are holes drilled in the

plexiglass creating objects that vary logarithmically in diameter and depth from 0.3 to 8.0









mm. The first three rows of square regions, containing objects 5.0 mm and larger, have

one object in the center of each square. The remaining rows have two identical objects in

each square, one in the center and the other in one of the four corners. The design of these

rows allows for the use of a four-alternative forced choice (4-AFC) experimental

methodology. The observer is forced to choose in which corer the object is located. This

type of study design can reduce the subjectivity of the observers' detection thresholds.

Twelve images of the CDRAD phantom were sequentially acquired for each image

receptor. The images were acquired to simulate a clinical chest exam; the image receptor

was in an upright bucky, an anti-scatter grid was utilized and a 12.7-cm acrylic absorber

was placed between the CDRAD phantom and the x-ray source. The images acquired

with the flat-panel detector were printed to film with a laser imager for hard-copy

evaluation. In order to utilize the 4-AFC methodology, the first three rows of the phantom

were not evaluated by the observers. Six observers with normal or corrected vision were

used. A signal detection model that is beyond the scope of this discussion was used to

calculate the threshold contrast level where a 75 percent probability of correct object

identification was reached.

Rong et al. 48 compared the low-contrast performance of an amorphous silicon-

cesium iodide based flat-panel digital chest radiography system to those of a screen-film

and a CR system by measuring their contrast-detail curves. The CDRAD contrast-detail

phantom was also used in this study. Images were acquired with the CDRAD phantom in

the center of the x-ray field against the chest detector-grid assembly and a 0.5-mm thick

copper plate was placed at the tube output to simulate patient attenuation. All images

were processed and printed according to their respective clinical protocols. The images









were evaluated by three physicists, a graduate student, and an undergraduate student.

Rong et al. state that, "previous studies have shown that for images of a contrast-detail

phantom there were no significant differences in observer responses between radiologists

and nonradiologists and that there was no noticeable improvement in the readers'

performance with increased experience."'48' p2330-2331) The minimum detectable object

contrast was then determined for each object size. A procedure to correct for the false

identification of an object in one of the four covers on a subregion, supplied by the

manufacturer, was applied to the raw scoring data before development of the contrast-

detail curves. Rong et al. concluded that the flat-panel system demonstrated a

significantly better low-contrast performance than the screen-film or CR systems as

compared by the contrast-detail curves.

MOSFET Dosimetry

Bower and Hintenlang42 evaluated the characteristics of a patient dose verification

system that uses high sensitivity MOSFET dosimeters that were developed for low dose

measurements. The dosimeters were evaluated at clinical diagnostic energy levels and

were found to perform well at these energies. The system was found to have a sensitivity

that was precise enough for many physics applications. It was also determined that the

size of the dosimeters would not interfere with diagnostic image quality. Bower and

Hintenlang concluded that, "The dosimeters have a good angular response, a very linear

response with dose and are extremely small. These characteristics make them good

candidates for use in a variety of dosimetry applications."'42' p 204















CHAPTER 4
EXPERIMENTAL EQUIPMENT

Imaging System

All images utilized in this research were acquired in the Radiology Department's

Examination Room 1 in Shands Teaching Hospital at the University of Florida. The x-ray

source is a Dunlee$ Duratron x-ray tube with a Picker VPE 3-phase generator. The x-ray

tube insert has a target angle of 13.5 degrees with a 0.6-mm and a 1.2-mm focus. The

inherent filtration of the tube-collimator assembly is 2.35-mm aluminum equivalent. The

system has a half-value layer of 2.73 mm of aluminum at 60 kVp. The table assembly

was manufactured by Picker and contains a reciprocating anti-scatter grid. The grid is

36.2 cm x 48.0 cm with a grid ratio of 12:1 and 103 lines per inch. The grid has a focal

distance of 36 to 40 inches.

The CR system is the Agfa** ADC system with a Compact plate reader and the

MD30 code 15 imaging plates. The 24-cm x 30-cm high-resolution imaging plates were

used. The plate-reader combination generates 8.8 pixels per mm in the final image.

Software

There were various commercially available computer packages used in the

completion of this work. Matlab, a high-level technical computing language produced




SDunlee, 555 North Commerce Street, Aurora, IL 60504

SPhilips Medical Systems, 22100 Bothell Everett Highway, P.O. Box 3003, Bothell, WA 98041-3003

*Agfa Corporation, 100 Challenger Road, Ridgefield Park, NJ 07660









by The Mathworks, Inc.t, was used in the calculation of the image quality metrics

investigated in this research. TASMIP, developed by Boone and Seibert46, was used to

generate the x-ray spectra needed in the evaluation of the image quality metrics. The

General Monte Carlo N-Particle Transport Code Version 4C (MCNP4C), produced by

the Diagnostics Applications Group at Los Alamos National Laboratory,n was used to

verify input parameters to TASMIP and simulate additional x-ray spectra for different

experimental setups. All statistical analysis was performed using SPSS for Windows,

Release 11.5.0, which is produced by SPSS Inc.

The images that were retrieved from the PACS at Shands hospital are in DICOM

format. Matlab was not able to read the images directly from the PACS. The program

ImageJ, developed by Wayne Rasband at the National Institutes of Health was used to

make the images readable by Matlab. This was accomplished by saving the DICOM

images as text images.

NPSc_ MTFc and DQEc

Acrylic was used to simulate patient scatter in the acquisition of the images used to

calculate the NPSc while maintaining a flat field. Two blocks of acrylic, each 40 cm x 32

cm x 2.5 cm were utilized. In addition to the acrylic, a 254-tm thick tungsten square (10

x 10 cm) was used as the edge device in the measurement of the MTFc. This device was

manufactured by Electronic Space Products International.*" The edges of the tungsten

square were manufactured to be straight with a precision of 50 im. In order to reproduce

t The Mathworks, Inc., 3 Apple Hill Drive, Natick, MA 01760

: Los Alamos National Laboratory, P.O. Box 1663, Los Alamos, NM 87545

SPSS Inc., 233 South Wacker Drive, 11th Floor, Chicago, IL 60606
*** Electronic Space Products International, 1050 Benson Way, Ashland, OR 97520









the angle of the edge, a device was custom made in-house to position the tungsten square.

Exposure measurements were made with the Radcalttt MDH (model 1015) ion chamber.

The standard probe was used for table top measurements and the pancake probe was used

for measurements in the bucky. The generator ripple was measured with a Kiethley=

35080A kVp voltage divider, with the 50-kV 150-kV pack, and a Tektronix

THS720A 100 MHz oscilloscope.

Clinical Contrast Detail Score

In addition to the acrylic used to simulate patient scatter, two contrast-detail

phantoms were used in this study. The first is the UF Radiology phantom shown in

Figure 2-6. This phantom was constructed from a 20 cm x 15 cm x 2.5 cm piece of

Lucite.27 The phantom contains nine object sizes ranging from 12.7 mm to 0.8 mm in

diameter. For each object size, seven contrast steps are created from drilled holes ranging

from 10.0 mm to 0.25 mm in depth. The objects are arranged in a square pattern. As

pictured, decreasing object size is arranged in columns and decreasing contrast steps are

arranged in rows. The second is the Leeds test object TO. 10 shown in Figure 2-7. This

phantom contains twelve object sizes with each size having nine contrast steps. The six

largest object sizes are arranged in a circular pattern around the periphery of the phantom

while the six smallest object sizes are arranged in rows in the phantom's center. The

objects range in diameter from 11.1 mm to 0.25 mm. The contrast steps for each object

size and the relative subject contrast between the different object sizes are not the same as



t Radcal Corporation, 426 West Duarte Road, Monrovia, CA 91016

SCardinal Health Radiation Management Services, 6045 Cochran Road Cleveland, OH 44139-3303

Tektronix, Inc, 14200 SW Karl Braun Drive, P. O. Box 500, Beaverton, OR 97077









in the UF Radiology phantom. The nominal contrast values, as listed in the instruction

manual, range from 0.012 to 0.93 but are not specified for a specific energy.49

Dosimetry

In order to measure the effective dose for different radiographic examinations, an

anthropomorphic phantom of a one-year old, produced by Bower34, and the MOSFET

AutoSenseTM Dose Verification System, manufactured by Thomson and Nielsen

Electronics Ltd, were utilized. The anthropomorphic phantom, previously mentioned in

Chapter 2, is shown in Figure 4-1. The arms are contained in the elliptical trunk while the

head and legs are not permanently attached. The trunk of the phantom is 21.5 cm in

length and 13.2 cm in width. Figure 4-2 and Figure 4-3 show the top and bottom of the

phantom's trunk respectively. Figure 4-4 shows the bottom of the phantom's head. The

small holes (3 mm in diameter) provide access to the body organs of interest for insertion

of the MOSFET dosimeters. The large hole in the head is for the spine, which protrudes

from the top of the trunk. The high sensitivity MOSFET dosimeters, TN-1002RDI, and

the dual bias supply, TN-RD-22, were utilized in this study. The dual bias supply allows

for either high sensitivity or standard bias in one piece of equipment.


Figure 4-1. One-year old Bower stylized anthropomorphic phantom.




























Figure 4-2. View of the top of the phantom's trunk showing the MOSFET access ports
and the spine (top-center).


Figure 4-3. View of the bottom of the phantom's trunk.




























Figure 4-4. View of the bottom of the phantom's head. The large hole on the left is for
insertion of the spine.














CHAPTER 5
EXPERIMENTAL METHODOLOGY

The majority of the research objectives stated in Chapter 1 deal with comparing the

NPSc, MTFc, DQEc and CDSc from images acquired and/or presented under varying

conditions. In order to make this comparison quantitative, a single metric was defined for

each image quality parameter. Since the NPSc, MTFc and DQEc vary over a range of

spatial frequencies, the integral under these curves, up to the Nyquist frequency, was used

for comparisons. These integral values were labeled INPSc, IMTFc and IDQEc

respectively, where the I stands for integral in each case. This methodology has been

utilized by previous researchers.50 The total number of visible objects in an image of the

contrast detail phantom (either perceived by a human observer or inferred from CNRT

values calculated from the observer studies) was used for the CDSc.

Image Acquisition

Experimental Setup

All images were acquired with a standard experimental setup that mimics clinical

practice. The clinical examination setup for a chest/abdomen/pelvis (CAP) exam of a

pediatric patient of similar size to the anthropomorphic phantom was chosen. This setup

can be seen in Figure 5-1. The edge test device and the TO. 10 phantom were imaged on

the exam table between two 2.5-cm blocks of acrylic. Flat-field images were acquired

with only the two blocks of acrylic in the x-ray field. Since the UF Radiology phantom

was constructed from a 2.5-cm block of Lucite it was imaged on the exam table, the

drilled holes facing up, with only one 2.5-cm block of acrylic on top of the phantom. The









test devices were placed between the two blocks of acrylic to simulate the evaluation of

the image quality metrics at a location inside the patient. The two contrast-detail

phantoms were oriented so that the portion of each phantom with the smallest objects and

the lowest subject contrast were in the portion of the image with the highest exposure.

The imaging cassette was placed in the bucky and the SID was set at 40 inches. The

automatic collimation was used in all cases except when imaging the anthropomorphic

phantom. In those cases the x-ray field was tightly collimated on the phantom's trunk.

Finally, the small focal spot (0.6 mm) was used for all images.

Acquisition Parameters

The medical physics staff at Shands Hospital has conducted extensive studies to

optimize the technique settings for various exams with the Agfa CR imaging system. The

x-ray generator technique settings (peak-tube potential and current-time product) that

would normally be used for a CAP of the anthropomorphic phantom are 60 kVp and 1.0

mAs. The image quality metrics were evaluated with images acquired around this

exposure level.

In addition to technique settings, there are many ways the raw image data can be

collected and processed before it is viewed on the QA station. These settings are selected

by the user before the image plate is scanned. The two most important parameters for this

research were the exposure class and the processing options associated with a specific

type of exam. The exposure class sets the gain of the photomultiplier tube in the plate

reader. This setting is chosen based on the exposure to the imaging plate. The CR

imaging system at Shands is calibrated so that an image acquired in the bucky with the

automatic exposure control (AEC) is scanned with exposure class 400. The exposure

class is analogous to the traditional speed of a screen-film imaging system. If the









exposure to the plate is doubled, the exposure class needs to be reduced by half. This will

produce the desired "optical density," or brightness, in the final image. The average

brightness level over the entire image is shown at the QA station and is called the

exposure level. The desired exposure level is 2.2.

There are three levels of processing options to describe a specific type of

radiographic procedure. Unless otherwise stated, the following processing options were

used for all images in this work. The primary and secondary processing option used was

system diagnosis. These options allowed for the minimum amount of processing

available with this imaging system. The tertiary processing options used were either flat-

field or full range. The flat-field option was used for the determination of the NPSc and

the full range option was used for all other images. In order to minimize any additional

variations in the acquisition parameters being evaluated, a single imaging plate was used

for the acquisition of all images that were directly compared (unless otherwise noted).

Image Retrieval

After the images were acquired, the consistency of the radiation output from the x-

ray tube, and therefore the level of exposure to the imaging plate, was verified for each

image acquired by recording the exposure level indicator at the QA station. The images

were then sent to the PACS. Most of the images, except those utilized in observer studies,

were retrieved from the PACS via CD. These images were then transferred to a personal

computer (PC) for processing.

Clinical Noise Power Spectrum

Variation with Peak-Tube Potential and Current-Time Product

In order to determine the NPSc, flat-field images with 5 cm of added acrylic (see

Figure 5-2) were acquired and mathematically evaluated the same way as a traditional









NPS calculation. This calculation was detailed in Chapter 2 and a Matlab algorithm, or

m-file, was developed to determine the INPSc. This m-file and an explanation of its

functionality can be seen in Appendix A. To compare the variation of the INPSc with

current-time product and peak-tube potential, images were acquired for each of the

technique settings listed in Tables 5-1 and Table 5-2. The last column in each table gives

the exposure class used to scan the imaging plate. This was based on an AEC exposure

with the CPS setup. For the peak-tube potential variation study, the current-time product

was adjusted to maintain as constant an exposure level as possible to the imaging plate.

The exposure level was measured in the bucky with the MDH ion chamber. The

measured exposure level for each set of technique factors is listed in Table 5-2.

Variation with Processing Option

In order to quantify the effect of other processing options, flat-field images were

acquired at 60 kVp and 1.0 mAs as previously described but two different exam types

were selected to process the raw image data. The processing options compared are listed

in Table 5-3.

Clinical Modulation Transfer Function

Edge Angle Verification

The edge alignment device was constructed to produce an edge angle of three

degrees with respect to the pixel matrix (see Figure 5-3). The following setup procedure

was tested to ensure reproducibility of the edge angle. The first block of acrylic was

placed on the exam table so the long edge was aligned with the line down the center of

the table. The alignment device, fitting snugly over the edge of the acrylic, was put in

place. The hinged arm (the right arm in Figure 5-3) was set to the correct angle by a small

peg. The edge device was placed on the acrylic and aligned with the right arm. The right









edge of the tungsten square was secured to the acrylic with tape. The alignment device

was removed and the second block of acrylic was placed on top of the first. This

procedure ensured the left and upper edges of the tungsten square (as shown in Figure 5-

3) were near the center of the image so one image could be used to calculate the MTFc in

both the scan and subscan directions (see Figure 5-4).

The imaging plate does not fit tightly in the imaging cassette (see Figure 5-5), so

even if the alignment device was exactly reproducible, slight variations in the plate

position could change the edge angle with respect to the pixel matrix. Two series of

images were acquired to investigate the reproducibility of the edge device setup. In the

first series, the setup was repeated for each image. In the second series, the edge device

was set up once and multiple images were taken. These images were mathematically

evaluated the same way as the traditional MTF calculation detailed in Chapter 2. A

Matlab m-file was developed to determine the IMTFc. This m-file and an explanation of

its functionality can be seen in Appendix B.

Variation with Peak-Tube Potential, Current-Time Product and Processing Option

To compare the variation of the IMTFc, images of the edge device were acquired

with the same technique settings and processing options as listed above for the INPSc.

The images were then evaluated with the IMTFc Matlab m-file.

Clinical Detective Quantum Efficiency

Once the NPSc and MTFc had been determined for a given set of acquisition

parameters, the DQEc could be determined. Before this calculation could take place, the

parameters k and Q in Equation 2-13 were determined. This was done through the

simulation of the x-ray spectrum at the imaging plate. The x-ray spectrum output from

the x-ray tube housing was simulated using TASMIP and MCNP4C. This procedure will









be explained later in this chapter. The details of this procedure are not necessary for this

discussion and it is assumed the output spectrum is known. A simulation was performed

in MCNP4C based on the physical setup of the NPSc measurement. The MCNP4C input

file for the simulation of the x-ray spectrum in the bucky at the imaging plate is shown in

Appendix D. This simulation was performed twice. The first was with a thin sheet of lead

simulating the thickness of the grid in the x-ray field. The second was without the lead in

the x-ray field. A weighted average was performed on the two resulting spectra to get the

simulated spectrum at the imaging plate. The weights were based on the proportion of the

x-ray field obstructed by the thickness (t) of the lead strips in the grid to the gaps (g)

between the strips (see Figure 5-6). The grid ratio is defined as h divided by g.

After the x-ray spectrum incident on the imaging plate was determined the value of

-2
Q (photons-mm -2mR1) was calculated. The x-ray spectrum from MCNP4C was

normalized so that when multiplied by a constant would give a true spectrum of

photons/mm2. Before the exposure in mR could be calculated from equation 2-14, the

energy dependant mass-energy absorption coefficients (ien/p) for air were calculated

from published values (see Table 5-4).24 The simulated spectrum was calculated at one

keV intervals, but Iein/p values are only published at a few energies between 10 and 80

keV. The intermediate values were calculated by fitting several curves to the known data.

The energy range between 0 keV and 80 keV was subdivided into five regions; 0-

10, 10-20, 20-40 40-60 and 50-80. The Ien/p values for the 10-20 keV range were fit with

a power series of the form

en (E)= 2x10-6E "1738 (5-1).

The n/ values for the 20-40 keV range were fit with a power series of the form
The ein/p values for the 20-40 keV range were fit with a power series of the form









"en (E)= 5x10-6E 29433 (5-2).
P

The Ien/P values for the 40-60 keV range were fit with a polynomial of the form

Pen (E) = 80.5E2 9.865E + 0.3326 (5-3).
P

The Ien/P values for the 50-80 keV range were fit with a polynomial of the form

Pen (E) = 23.333E2 3.576E + 0.1611 (5-4).
P

This last energy range overlapped with the previous one so that three Ien/P values

were available for the curve fitting but Equation 5-4 was only used to calculate the len/P

values above 60 keV. The R2 values describing how well the above equations fit the data

are 1.0, 0.9992, 1.0 and 1.0 respectively, with 1.0 being the best possible fit. The len/P

values for the 0-10 keV range were determined from linear back-interpolation using the

known eLn/P value for 10 keV and the calculated value for 9 keV. Since two equations

were fit to each known data point except at 10 keV and 60 keV, the published values

were used at those energies.

The multiplicative constant mentioned previously was changed until the sum of the

exposure (in mR) for each energy bin in the spectrum was one. The exposure for each

energy was calculated using Equation 2-14. The photon spectrum was then summed over

all energies to give a total number of photons per mm2 to produce one mR at the imaging

plate. This value was then used in the calculation of the DQEc based on the measured

exposure level in the bucky for a given set of technique factors.

The parameter k, representing the gain of the imaging system for each set of

acquisition parameters, was determined from the average pixel values from the center









portion of the flat-field images used in the determination of the INPSc. A ROI about the

size of the MDH pancake probe used to measure the in bucky exposure was extracted and

averaged from each flat-field image. These values were then averaged for each set of

acquisition parameters. These average pixel values were then normalized to one set of

acquisition parameters. This gave a relative system gain which included not only the

change in technique factors, but also the exposure class.

Clinical Contrast Detail Score

Observer Study

Images were acquired of the two contrast detail phantoms with 5 cm of total acrylic

scattering media. The images were acquired at 60 kVp with the acquisition parameters

listed in Table 5-5 and Table 5-6. The exposure classes were based on an AEC exposure

of each phantom setup and adjusted accordingly for the change in current-time product.

As previously mentioned, the system-diagnosis/full-range processing options were used.

Two sets of these images were sent to the PACS. The first set was the normal images and

the second was a set of those same images with the grayscales inverted. All of the images

were then available for scoring at a diagnostic workstation in the pediatric reading room

at Shands.

Thirteen human observers were utilized to score the images of the contrast-detail

phantoms. There were four third- or fourth-year radiology residents, four experienced

medical physicists and five medical physics graduate students. All images were evaluated

on the same diagnostic viewing workstation. All of the images of one phantom were

evaluated in a single viewing session. Therefore, each observer had two viewing sessions

during this study. The images were presented to each observer in the same randomly

selected order based on the current-time product acquisition value. For each current-time









product value four images were read in a specific order. The normal image was scored

first without the use of any of the image processing tools available at the workstation.

Next, the grayscale inverted image was scored without the use of any processing tools.

The normal image was then scored again with the use of the window and level controls.

Finally, the grayscale inverted image was scored with the use of the window and level

controls. Each observer was allowed to get familiar with the functioning of the window

and level controls at the workstation with a test image of the phantom before the scoring

began. Each observer was also given the following instructions before scoring began.

1. Identify the number of objects you can see of each object size. An object is
considered visible if it is an entire circle with a clear circular outline.

2. Score each row with your first impression. Do not spend more than a few seconds
on each row.

3. Operate the window and level controls as you see fit to best score the image.

4. Do not score past an object that is "not" visible.

Once all of the images were scored, one of the images the viewer had already scored was

presented again as a check of observer reproducibility.

CNR Calculations

The observer study produced an average threshold object number for each object

size for the four image scoring conditions. In order to calculate the CNR associated with

this threshold of visibility, a Matlab m-file was developed for each phantom to calculate

the CNR of each object. This m-file and an explanation of its functionality can be seen in

Appendix C. Since both m-files function essentially the same and differ only by input

parameters, only the TO. 10 m-file is presented.









Variation with Processing Option

The m-file in Appendix C can also automatically score images of the contrast-detail

phantoms used in this study. This algorithm's automatic scoring ability is based on the

combination of the CNR calculations and the observer data. The reasoning behind this

procedure will be presented in the next chapter. This ability allowed the scoring of

images acquired with the acquisition parameters listed in Table 5-5 and Table 5-6 for

each of the processing options listed in row two and three of Table 5-3 without the use of

human observers.

Anthropomorphic Phantom Viewer Study

In order to evaluate the image acquisition parameters needed to produce an image

of minimum acceptable quality, a series of images were acquired of the anthropomorphic

phantom. The phantom was placed on the exam table and the imaging cassette was

placed in the bucky. No additional scattering material was placed in the x-ray field. The

images were acquired with the acquisition parameters listed in Table 5-6 (based on an

AEC exposure of the phantom) for both of the processing options listed in rows one and

two of Table 5-3. The images were then sent to the PACS for retrieval at the same

diagnostic viewing station used in the contrast-detail observer study. Since Shands has

already undergone a procedure to optimize their technique factors, the technique factors

recommended by the technologist to image the anthropomorphic phantom were initially

assumed to produce an image of minimal acceptable quality (60 kVp and 1.0 mAs). In

order to verify that assumption the radiology residents were asked to rate all other images

to this standard image. In a process similar to Rill,51 the images were presented one at a

time against the standard image and the radiologist was asked to score the new image on

a five-point scale (see Table 5-7). An image with a lower exposure level than the standard









image with an average rating of three or higher will replace the standard image as the

image of minimum acceptable quality.

Dosimetry

Four MOSFET dosimeters were used in the measurement of absorbed dose to the

individual organs of the anthropomorphic phantom. This was done because only four of

the dosimeters had a total accumulated exposure resulting in less than 10,000 mV. Under

this level of exposure, the dosimeters do not need to be recalibrated. Three measurements

were made at each of the relevant organ sites needed to calculate the effective dose. Each

measurement was performed with the same physical setup used in the anthropomorphic

phantom viewer study with technique settings of 60 kVp and 300 mAs.

In order to perform the calculation of effective dose using Equation 2-20, the

average mass energy absorption coefficients were needed. This involved determining the

average energy of the x-ray beam. An accurate method for determining diagnostic x-ray

spectra is through the use of TASMIP. The TASMIP algorithm requires three input

parameters; peak tube potential, generator ripple and aluminum-equivalent-inherent

filtration. The peak tube potential, the effective kV and the generator ripple were

physically measured with an oscilloscope. The inherent filtration provided by the

manufacturer was verified through the measurement of the x-ray unit's aluminum half-

value layer (HVL) and the use of MCNP4C. The MCNP4C input file that performs this

simulation is shown in Appendix D.

The half-value layer was measured with the use of the MDH ion chamber, a test

stand designed specifically for the HVL measurement and varying thicknesses of

aluminum. An initial exposure measurement was performed without any aluminum in the

x-ray field. Varying thicknesses of aluminum were added and an exposure measurement









was taken for each aluminum thickness until the exposure level was less than one-half the

value of the initial exposure measurement. This data was then used to determine the

thickness of aluminum that would be needed to reduce the exposure to one half of the

initial exposure measurement. This procedure was repeated for 50, 60, 70 and 80 kVp.

Starting with the manufacturer's stated inherent filtration as an input to TASMIP, a

spectrum was generated and used in a simulation of the x-ray system. The simulation was

performed twice. The first was with the measured amount of aluminum, determined from

the half-value layer measurement, in the x-ray field. The second was without the

aluminum in the x-ray field. The exposure to a simulated detector was calculated for each

simulation. When the first simulation produce a reduction in the exposure to the

simulated detector of one-half the value of the second simulation, the inherent filtration

parameter used as the input to TASMIP was considered correct. This did not occur with

the first x-ray spectrum so the inherent filtration value was changed until the proper

exposure reduction was seen. The mean energy was calculated from this spectrum. The

effective dose was then calculated from the measured exposure level and linearly scaled

with current-time product to get the effective doses at the diagnostic exposure levels used

in the evaluation of the image quality metrics.












- X-ray Tube Housing



Edge for MTF


SAcrylic Blocks
I


B Exam Table


Bucky and Grid


\Imang Plate


Figure 5-1. Image acquisition setup.


Figure 5-2. Image acquisition setup for the flat-field images. The two 2.5-cm blocks of
acrylic are placed on the exam table and the x-ray tube is set at an SID of 40
inches.


40" SID


I I


_j



























Figure 5-3. Device to ensure proper angulation of edge device.


Figure 5-4. Final setup of the edge device for the determination of the MTFc.




























Figure 5-5. The Agfa imaging plate inside the open imaging cassette.

t g





h





Figure 5-6. Idealized graphical representation of the grid.

Table 5-1. Technique factors for the current-time product variation study.
k\'p mAs Exposure C'lss
60 0.4 800
60 0.8 800
60 1.0 800
60 1.2 800
60 2.0 800
60 3.2 400









Table 5-2. Technique factors for the peak-tube potential variation study.


50 12 200
60 5.6 200
70 3.0 200
80 2.1 200


Table 5-3. Processing options compared.
Processing Option Level
Exposure Class
Primary Secondary Tertiary
System Diagnosis System Diagnosis Flat Field 800
Pediatric Chest Chest PA 800
Pediatric Upper Extremity Hand AP 800


Table 5-4. Published mass energy-absorption coefficients for air.


10 4.61
15 1.27
20 0.511
30 0.148
40 0.0668
50 0.0406
60 0.0305
80 0.0243


Table 5-5. Technique factors for the TO. 10 phantom.


60 0.4 800
60 0.8 800
60 1.0 800
60 1.2 800
60 2.0 800
60 3.2 400


mAs Exposure Class


kVp


Energy (keV)


([ten/P)E,air (I/cm)


kVp


mAs


Exposure Class









Table 5-6. Technique factors for the UF Radiology Phantom.


60 0.4 800
60 0.8 800
60 1.0 800
60 1.2 800
60 2.0 400
60 3.2 200


Table 5-7. Five-point scale for subjective image quality evaluation.
Score Image Quality Compared to Standard Image
1 Much Worse
2 Worse
3 Same
4 Better
5 Much Better


mAs Exposure Class


kVp














CHAPTER 6
RESULTS AND DISCUSSION

NPSc

The full two-dimensional NPSc was calculated directly from the flat-field images

in this study. The pixel values in the raw image data were used directly and were not

converted to exposure values before calculation of the NPSc. Unlike the MTFc, which

must be calculated in the laser scan and subscan directions separately, the two-

dimensional NPSc can be calculated from a single image and the one-dimensional NPSc

in the scan and subscan directions is then calculated from the two-dimensional NPSc as

previously described. The INPSc is the integral under the full two-dimensional NPSc.

Even though the full two-dimensional NPSc has the positive, as well as the redundant

negative frequency values present, the INPSc was defined in this way for ease of

evaluation.

Variation with Current-Time Product

The variation of the INPSc with current-time product is shown in Figure 6-1. Five

images at each current-time product were used to calculate an average INPSc. The INPSc

decreased as the current-time product was increased. A one-way analysis of variance

(ANOVA) was performed to compare the means and check for statistically significant

differences. The ANOVA p-value was 0.000 so differences exist among the means.

Additionally, the Levene statistic was used to test for homogeneity of the variances. The

p-value was 0.000 so the assumption of equal variances was not used in the post-hoc

tests. Three multiple comparison post-hoc tests that do not assume equal variances were









performed. The results of the post-hoc tests are shown in Appendix E (Table E-l).

Tamhane's T2 is a conservative pairwise comparisons test based on a t test. Dunnett's T3

is a pairwise comparisons test based on the Studentized maximum modulus. Lastly,

Games-Howell is a more liberal comparison test.

The INPSc decreases with increasing current-time product. This was the expected

result since the NPSc is a measure of the noise level in an image and as the current-time

product is increased more photons go into the formation of the image and the quantum

noise level is reduced. The differences in the INPSc were statistically significant for all

current-time product comparisons. The INPScs determined at 0.4 mAs had the largest

variance. The average values and their associated standard deviations (C) are shown in

Table 6-1.

As the current-time product is increased, the magnitude of change in the INPSc for

equal increases in current-time product does not remain constant. For example, the

difference in the average INPSc for a doubling of the current-time product as determined

from 0.4 mAs and 0.8 mAs is 2.222x 106 while the difference is 9.497x 107 between 1.0

mAs and 2.0 mAs. This shows that the INPSc is more sensitive to changes in exposure

level at lower exposure levels. This was the expected result since the CR imaging plate is

a non-linear photon detection system.

Variation with Peak-Tube Potential

The variation of the INPSc with peak-tube potential is shown in Figure 6-2. Five

images at each peak-tube potential were used to calculate an average INPSc. The current-

time product was adjusted in the acquisition of the flat-field images used to test the

INPSc variation with peak-tube potential in the attempt to maintain a constant exposure

level of 1.0 mR (3.876 C/kg). The mA and time were adjusted independently at each









peak-tube potential to maintain the exposure level as close to 1.0 mR (3.876 C/kg) as

possible. The INPSc decreased slightly as the peak-tube potential was increased from 50

kVp to 60 kVp, then remained statistically constant up to 80 kVp. The ANOVA p-value

was 0.000 so significant differences exist among the means (in this case the INPScs

determined at 50 kVp and 60 kVp statistically differed). The Levene p-value was 0.000

so the assumption of equal variances was not used in the post-hoc tests. The results of the

post-hoc tests are shown in Table E-2.

Variation with Processing Option

The variation of the INPSc with the three different processing options previously

discussed is shown in Figure 6-3. Ten images processed with each processing option

were used to calculate an average INPSc. Ten images were used instead of five because

of the increased variance in the INPSc when the hand AP processing algorithm was used.

The INPSc decreased with the chest PA processing option and increased with the hand

AP processing option. The ANOVA p-value was 0.000 so statistical differences exist

among the means. The Levene statistic p-value was 0.001 so the assumption of equal

variances was not used in the post-hoc tests. The results of the post-hoc tests are shown in

Table E-3. The slight decrease in the INPSc when the chest PA processing option was

used indicated that the algorithm performs a smoothing operation on the raw image data.

The significant increase in the INPSc when the hand AP processing option was used

indicated that an edge enhancement process was performed.

MTFc

Two methods of determining the MTFc were initially performed; the use of a

narrow slit and the use of a sharp attenuating edge. Initial images of a 20-[m slit at the

diagnostic exposure levels investigated in this work did not produce a viable data set to









calculate the LSF. With the low exposure levels and the presence of simulated patient

scattering media the slit was barely visible in the image. There was not enough data

present in the image to generate a usable LSF and the data that was available was

extremely noisy. The images of the 250-[tm thick tungsten edge device were easily used

to generate an ESF. Therefore, the investigations of the MTFc that follow were all

accomplished with the edge method.

Scan Versus Subscan Direction

The Agfa Compact CR reader scans the imaging plate along its' longest dimension.

Since a single image is used to calculate the MTFc in both the scan and subscan

directions, these metrics are determined using two perpendicular edges of the tungsten

square. To ensure that both edges produced the same results, several images were

acquired with both edges in the scan and subscan direction. The images were acquired

with 60 kVp and 1.0 mAs. It should be noted that the Matlab m-file used to calculate the

IMTFc was designed to evaluate the subscan MTFc using a vertical edge that is parallel

to the scan direction (see Figure B-l). In order to evaluate the MTFc in the scan direction

the image must be rotated 90-degrees counter-clockwise so that the top edge of the

tungsten square in Figure B-l is properly oriented for evaluation. The average IMTFc

from five images for both edges in the scan and subscan directions are listed in Table 6-4.

A Student's t-test was performed on the data and it was determined that both edges

functioned equally well in both directions.

IMTFc Reproducibility

There are two main sources of variability in the IMTFc calculation. The first is due

to the manual selection of the usable edge data from the edge image. This caused a

variation in the value of the IMTFc for multiple evaluations of the same image. The









second is the difference in the IMTFc as calculated from multiple images acquired with

the same acquisition parameters. Both of these sources of variability were evaluated. The

IMTFc calculated five times from a single image acquired with 60 kVp and 1.0 mAs is

shown in Table 6-5. This evaluation showed a maximum variability of 0.006 mm-1. The

IMTFc calculated once from five separate images is shown in Table 6-6. This evaluation

showed a maximum variability of 0.017 mm1.

Edge Angle Reproducibility

The actual angle the edge device made with respect to the pixel matrix was

calculated by the Matlab IMTFc algorithm. The target edge angle was three degrees. The

actual angle of the edge in the images for both repeated edge setups and repeated images

with a single setup are listed in Table 6-7. The average for the repeated setups was 2.8

degrees with an overall variability of 0.61 degrees. The average for repeated images of

the same setup was 2.75 degrees with an overall variability of 0.50 degrees. Combining

these results, the 95-percent confidence interval for the edge angle used to determine the

IMTFc is 2.79 +0.49 degrees. The edge angle was the same for the scan and subscan

direction in each case.

Number of ESF Data Points

The number of data points selected about the edge for the ESF has an effect on the

magnitude of the IMTFc. This is primarily due to increased exposure just under the edge

device from scattered x rays. This phenomenon causes a shoulder on the attenuated side

of the ESF that reduces the sharpness of the edge response (see Figure 6-4). If more

pixels are used for the ESF, more of the shoulder is incorporated and the magnitude of

the IMTFc is decreased.









Due to the way the FFT is calculated, the number of data points in the final MTFc

is one half the number of data points in the LSF. Since the LSF is zero padded to 256 data

points before the calculation of the MTFc, there are 128 data points in the MTFc

independent of the length of the ESF. Therefore, the frequency resolution of the MTFc

appears to be independent of the ESF length. This is not the case since the length of the

LSF is artificially increased to 256 data points. This process is analogous to adding

additional data points to the MTFc by interpolating between the values that would be in

the MTFc based on the true length of the ESF. The IMTFc for ESFs of varying lengths

are shown in Table 6-8 and Table 6-9. An image of the edge device acquired with 60 kVp

and 1.0 mAs was used to calculate the IMTFcs in Table 6-8 and Table 6-9.

There is approximately a 20-percent reduction in the IMTFc when 250 data points

are used compared to the use of 25 data points in both the scan and subscan directions.

This is not a large decrease and since the shoulder on the attenuated side of the ESF is a

direct result of the clinical nature of the edge image acquisition, it was decided that some

of the shoulder should be included in the "clinical" edge response. In addition, the

smaller number of data points used does produce a smoother MTFc (see Figure 6-5), but

at the cost of frequency resolution. Since the MTFc is used to compare different images,

the choice of data length is somewhat arbitrary as long as the same methodology is used

for each calculation. Therefore, in order to see the effect of the ESF shoulder 128 data

points were selected for further calculations of the IMTFc.

Variation with Current-Time Product

The variation of the IMTFc with current-time product is shown in Figure 6-6 and

Figure 6-7. Five images at each current-time product were used to calculate an average

IMTFc. The IMTFc appears to decrease with increasing current-time product. The









ANOVA p-value was 0.002 for the scan direction and 0.000 for the subscan direction so

differences appear to exist among the means. The Levene statistic p-value was 0.001 for

the scan direction and 0.029 for the subscan direction so the assumption of equal

variances was not used in the post-hoc tests. The results of the post-hoc tests are shown in

Table E-4 and Table E-5. Even though the ANOVA had a p-value of 0.002 for the scan

direction, when the post-hoc tests were performed and the observed significance level

was adjusted for the fact that multiple comparisons were made, the IMTFc means were

not statistically different. The IMTFcs for the subscan direction showed statistically

significant differences between those determined at low current-time product values and

those determined at high current-time product values. A difference was not seen between

the IMTFcs determined from 0.4 to 1.0 mAs and those determined from 1.0 to 3.2 mAs.

The IMTFcs determined at 0.4 mAs showed the largest variation. The average

values and their associated standard deviations are shown in Table 6-10. This larger

variation, as well as the overall larger average magnitude, can be attributed to the

increased noise level in the edge imaged acquired at 0.4 mAs. The increased noise causes

increases in the high frequency components of the MTFc. As the current-time product

was increased and the noise level in the edge image decreased, the MTFc became more

stable. This behavior can be seen in Figure 6-8 and Figure 6-9. If the MTFc is averaged

over many images these variations start to get averaged out and the MTFc starts to

stabilize (see Figure 6-10). Even though these increases in the high frequency

components are reduced when the MTFcs are averaged before integration, there are still

slight increases in the average MTFc determined at lower current-time product values

(see Figure 6-11). Since the maximum variation of the IMTFc is approximately 0.04









mm-1 and the reproducibility of the IMTFc calculation is on the order of 0.02 mm-1, these

small increases in the high frequency components of the MTFc from increased noise in

the edge image account for the increase in the IMTFc at low current-time product values.

Variation with Peak-Tube Potential

The variation of the IMTFc with peak-tube potential is shown in Figure 6-12 and

Figure 6-13. Five images at each current-time product were used to calculate an average

IMTFc. The IMTFc showed a definite increase with peak-tube potential. The ANOVA p-

value was 0.000 for both the scan and subscan directions so differences exist among the

means. The Levene statistic p-value was 0.230 for the scan direction and 0.564 for the

subscan direction so the assumption of equal variances was assumed in the post-hoc tests.

Three multiple comparison post-hoc tests that assume equal variances were

performed. The results of the post-hoc tests are shown in Table E-6 and Table E-7. Both

Tukey's honestly significant difference test (HSD) and the Bonferroni test are multiple

comparisons tests that adjust the observed significant level for the fact that multiple

comparisons are made. Tukey's HSD is based on the Studentized range statistic while the

Bonferroni test is based on Student's t statistic. The least significant difference (LSD) test

is a pairwise multiple comparison test that is equivalent to multiple individual t tests

between all pairs of groups. The LSD test does not adjust the observed significance level

for multiple comparisons. All of the comparisons showed statistical differences in the

average IMTFcs at all peak-tube potentials in both directions.

The increase in the IMTFc with peak-tube potential was a direct result of the shape

of the ESF. It was previously noted that the ESFs at 60 kVp have a shoulder on the

attenuating side of the edge as a result of scatter. As the peak-tube potential is increased,

the scattering of photons under the edge device is decreased and the shoulder on the ESF









is reduced (see Figure 6-14). This leads to an increase in the magnitude of the MTFc and

as a result an increase in the IMTFc. Using the MTFc as a measure of image quality

could lead to the conclusion that the higher the peak-tube potential used the better the

image quality. This is not true for all applications because more factors than just the

ability of the system to transfer subject contrast need to be considered. If an imaging

system had a perfect MTFc (a value of one for all spatial frequencies) but such a high

peak-tube potential was used that the object being imaged presented very little subject

contrast, the image would still be of poor quality. So even though the IMTFc increases

with increasing peak-tube potential, more factors need to be considered before a

statement can be made regarding the quality an image would have if the same acquisition

parameters were used.

Variation with Processing Option

The variation of the IMTFc with the three different processing options is shown in

Figure 6-15 and Figure 6-16. Five images processed with each processing option were

used to calculate an average IMTFc. The IMTFc increased when both the chest PA and

the hand AP processing algorithms were applied to the raw image data. The ANOVA p-

value was 0.000 for both the scan and subscan directions so differences exist among the

means. The Levene statistic p-value was 0.019 for the scan direction and 0.178 for the

subscan direction so the assumption of equal variances was not used in the post-hoc tests

for the scan direction but was used for the subscan direction. The results of the post-hoc

tests are shown in Table E-8 and Table E-9. The IMTFc means were statistically different

for all processing options. The increase in the IMTFc for the chest PA and the hand AP

processing options was directly attributed to the enhancement of the ESF (see Figure 6-

17). The shoulder on the attenuated side of the edge is almost completely eliminated,









which lead to an increase in the magnitude of the MTFc and therefore an increase in the

IMTFc. This occurs because both processing options perform a smoothing of the bright

areas (those that would represent bone in a clinical image) after the edge enhancement.

This leads to not only a better edge response but a flattening of the data under the edge

device which is a bright area in the image similar to a bone. The hand AP processing

option was so effective with this enhancement that the MTFc has a value greater than one

for the low frequency components in the scan direction (see Figure 6-18). This behavior

was not noticed in the subscan direction because the edge response was not reproduced as

well in that direction, but the IMTFc is still increased with the use of the hand AP

processing option. This enhancement of the MTFc does have its limitations. In the case of

the hand AP processing option the noise level in the image is increased. If the noise level

in the original image data is sufficiently low, this noise increase may be tolerable in a

clinical setting. For the chest PA processing option, the increase in noise over the entire

image caused by the edge enhancement is corrected by a smoothing operation which

leads to an overall noise reduction.

Dynamic Image Manipulation

The calculation of the INPSc and the IMTFc in the previous two sections was

performed on the raw, or digitally processed, image data before any changes were made

by the viewing software or the viewer. The monitors of the diagnostic display stations

used in the Radiology Department at Shands are capable of displaying 4096 gray scale

levels so a 12-bit image would be accurately represented. Therefore, the raw image data

used for calculations in this study should be represented by the pixel brightness values

displayed on the monitor. Unfortunately, the viewing software automatically adjusts the

window and level settings on the image to optimize the display. The level is simply the









midpoint of the pixel values displayed. The window is the range of pixel values that are

displayed around the level. This automatic adjustment changes the image data that the

viewer sees so the displayed image is no longer the same as the raw image data. Once the

image is displayed, the viewer also has the ability to dynamically change the window and

level settings. This causes a continuous change in the displayed image data. These

changes have an effect on the INPSc and the IMTFc, but it is not practical to capture the

displayed image data to perform calculations. In order to demonstrate the range of this

effect, the window and level of a flat-field and an edge image (acquired at 60 kVp and

1.0 mAs) were altered with the ImageJ program. The INPSc and the IMTFc were then

calculated (see Table 6-11 and Figure 6-19) from these manipulated images. ImageJ only

allows the user to save an image that has had the window and level adjusted if it is an 8-

bit grayscale image. Since the images acquired from PACS are 12-bit images, they were

first converted to 8-bit images before the adjustments were made. The window and level

were adjusted one at a time in both the positive and negative directions to the full extent

allowed by the ImageJ program. After each adjustment the image was saved as a text

image so it could be read by Matlab.

As the window was increased to its maximum slope, the flat-field image became an

almost uniform image with all pixel values set to 255 except a few around the periphery

where the imaging plate was outside of the x-ray field during image acquisition. This

reduced the INPSc to nearly zero. The reason the INPSc was not exactly zero is discussed

in Appendix A. The increase in window caused the edge image to resemble the perfect

binary image discussed in Appendix B. All of the pixel values outside of the edge device

were set to 255 while all of the pixel values under the edge device were set to zero. This









caused a dramatic increase in the IMTFc. It should be noted that this is the most extreme

case and in clinical practice as the window is increased, so is the perceived noise level.

The noise level begins to vanish as the window approaches its maximum slope, but so

does any subject contrast. Therefore, once again more factors than just the MTFc need to

be considered when assessing the quality of an image. As the window was decreased to

its minimum slope, the pixels in the flat-field image were set to a narrow range of values

in the middle of the grayscale range. This caused a reduction in the INPSc but was not as

drastic an effect as the window increase. The window decrease had a similar effect on the

edge image except the pixels were set to two narrow ranges of values corresponding to

those pixels outside of the edge device and those under the edge device respectively. This

did not have an effect on the IMTFc because the relative edge response remained

unchanged.

As the level was increased to its maximum value, the flat-field image became a

uniform image as almost all of the pixel values increased beyond the usable 8-bit

grayscale level of 255. This caused the same effect on the INPSc as the increase in

window. This increase in level caused the majority of the pixels outside of the edge

device to increase in value beyond the usable grayscale level and were therefore set to a

value of 255. This flattened the tail of the ESF on the non-attenuated side of the edge

resulting in a small increase in the IMTFc. As the level was decreased to its minimum

value, the pixel values in the flat-field image were all reduced by a constant but none of

the pixel values decreased below zero so the INPSc was not affected. This decrease in

level caused the majority of the pixel values under the edge device to decrease below the

minimum grayscale level and therefore be set to zero. This significantly flattened out the









shoulder on the attenuated side of the ESF and caused an increase in the IMTFc. The

MTFcs for these changes in window and level are shown in Figure 6-19.

It should be noted that the increase in level affected the non-attenuated side of the

ESF because the raw image data from the PACS has an inverted grayscale compared to

the images viewed on the display stations. The raw image data has high pixel values for

high radiation exposure and low pixel values for low radiation exposure. This is exactly

opposite from images acquired on film where a high radiation exposure leads to a darker

area on the film. This increased exposure results in less light being transmitted through

the film causing the perception of a low "number." In order to make the raw image data

in CR to appear like a traditional film, the grayscale is inverted before display. For this

reason, the ESFs shown in this chapter have been grayscale inverted for display purposes

so they appear in the traditionally accepted way.

DQEc

The DQEc was calculated from the NPSc and MTFc data previously shown in this

chapter. In order to calculate the DQEc, the parameters k and Q in Equation 2-13 were

calculated. The value of k for this application of the DQEc was determined to be unity.

Since the raw image data is corrected for background trends by subtracting a fitted

surface from the original data before the calculation of the NPSc, the mean pixel value of

each ROI used in the NPSc calculation is nearly zero. This procedure corrected for the

total gain of the imaging system including any change in exposure class. The effect of the

system gain on the image noise content is not accounted for by the background trend

correction. A correction for this was not desired because the altered noise content is what

is initially displayed on the monitor as 12-bit brightness levels during image viewing. In

order to calculate Q, the input parameters for TASMIP were determined. The half-value









layer (HVL) was determined to be 0.2733 mm of Al at 60 kVp. Simulating the HVL

measurement in MCNP4C the x-ray tube housing inherent filtration was determined to be

2.35-mm Al-equivalent. The generator ripple was determined to be approximately five

percent across the peak-tube potential range of interest. Using the x-ray spectrum

generated by TASMIP, the flat-field setup was simulated in MCNP4C and the parameter

Q was calculated for each peak-tube potential (see Table 6-12). The parameter Q and was

then normalized by the exposure measured in the bucky for each set of acquisition

parameters (see Table 6-13 and Table 6-14).

Variation with Current-Time Product

The variation of the IDQEc with current-time product is shown in Figure 6-20 and

Figure 6-21. The IDQEc showed the same trend with current-time product as the INPSc.

This decrease in the average IDQEc with current-time product is due to the parameter Q

in Equation 2-13. The NPSc decreases with current-time product but the Q value

increases at a greater rate causing a decrease in the IDQEc. Physically this was

determined to show that as the current-time product is increased the efficiency of the

imaging system per unit photon decreases. The ANOVA p-value was 0.000 for both the

scan and subscan directions so differences exist among the means. The Levene statistic p-

value was 0.098 for the scan direction and 0.152 for the subscan direction so the

assumption of equal variances was assumed in the post-hoc tests for both directions. The

results of the post-hoc tests are shown in Table E-10 and Table E-11. All of the average

IDQEcs showed statistically significant differences for two of the post-hoc tests in the

scan direction and all of the post-hoc tests in the subscan direction. The Bonferroni test

showed statistically equivalent IDQEcs determined at 0.8, 1.0 and 1.2 mAs.









Variation with Peak-Tube Potential

The variation of the IDQEc with peak-tube potential is shown in Figure 6-22 and

Figure 6-23. The IDQEc showed a decreasing trend with increasing peak-tube potential.

The ANOVA p-value was 0.000 for both the scan and subscan directions so differences

exist among the means. The Levene statistic p-value was 0.470 for the scan direction and

0.885 for the subscan direction so the assumption of equal variances was assumed in the

post-hoc tests for both directions. The results of the post-hoc tests are shown in Table E-

12 and Table E-13. The decrease in the average IDQEc was statistically significant for at

least two of the post-hoc tests in the scan direction for all increases in peak-tube potential.

The only statistically significant difference in the average subscan IDQEc with a change

in the peak-tube potential was for 80 kVp. The average subscan IDQEc determined from

lower peak-tube potentials were not statistically different as determined by the post-hoc

tests that adjust the error rate for multiple comparisons.

This decrease in the IDQEc with increasing peak-tube potential was once again

driven by the parameter Q but the MTFc also added to the decrease. Since the exposure

level was kept constant for all of the peak-tube potentials, the NPScs were almost the

same. Both the MTFc and the Q value increased with increasing peak-tube potential, this

lead to the decrease in the IDQEc with increasing peak-tube potential.

Variation with Processing Option

The variation of the IDQEc with processing option is shown in Figure 6-24 and

Figure 6-25. The average IDQEc increased with the chest PA processing option but did

not change with the hand AP processing option in both the scan and subscan directions.

The ANOVA p-value was 0.000 for both the scan and subscan directions so differences

exist among the means. The Levene statistic p-value was 0.026 for the scan direction and









0.028 for the subscan direction so the assumption of equal variances was not assumed in

the post-hoc tests for both directions. The results of the post-hoc tests are shown in Table

E-14 and Table E-15. All three post-hoc tests showed a statistically significant increase in

the average IDQEc with the use of the chest PA processing option but no statistical

change occurred with the use of the hand AP processing option.

The use of the integral of a function as a means of comparing two functions does

have its limitations. As the NPSc and MTFc changed with different acquisition

parameters (current-time product and peak-tube potential), the direction of the change

(positive or negative) was consistent for all spatial frequencies. The IDQEc did not show

this same behavior with a change in the processing option. The DQEc in the subscan

direction for all three processing options is shown in Figure 6-26. The IDQEc for the full-

range and the hand AP processing options are statistically the same but the DQEcs have

dramatically different shapes. This shows that the processing options are truly application

specific and the magnitude of the DQEc they produce at any particular spatial frequency

value would depend on what is being imaged. If small objects are to be imaged, the ideal

DQEc would have a large magnitude at high spatial frequencies while in contrast if large

objects are to be imaged, the ideal DQEc would have a large magnitude at low spatial

frequencies. This shows that the IDQEc could be a misleading indication of overall image

quality.

CDSc

Observer Study

Each of the contrast-detail phantoms used in this study was scored under four

different viewing conditions: the standard image without the use of window and level, the

standard image with the use of window and level, and the inverted image under the same









two conditions. The total number of visible objects, or score, for each phantom under all

viewing conditions is shown in Table 6-15 and Table 6-16. These are the combined

results of all observers. A one-way ANOVA was performed on the total score across all

current-time product values and viewing conditions for each phantom to determine if

there was a difference in the performance of the observers based on specialty. Of the

three groups of observers, the radiology residents and the students performed statistically

equivalent for both phantoms. The medical physicists performed statistically equivalent

to the other groups for the UF Radiology phantom but scored statistically higher by 7.5

percent for the TO. 10 phantom (the p-value was 0.001). This is attributed to the

subspecialties within the medical physicists group. Despite the same viewing instructions,

the nuclear medicine physicist and the mammography specialist scored higher than the

other medical physicists. Therefore, the total scores shown in Table 6-15 and Table 6-16

are representative of an average observer across several radiological specialties.

Since the number of visible objects changes with current-time product, the

ANOVA was repeated for the mean total score of each group across all viewing

conditions at each current-time product independently. This was done to ensure that the

increased variance caused by the total score changing with current-time product did not

affect the results. The p-values for this analysis are shown in Table 6-17. The only

difference found between the groups was with the TO. 10 phantom at 2.0 mAs. Post-hoc

tests showed that only the medical physicists and the radiology residents scoring of the

contrast-detail images statistically differed but the conclusion regarding inter-group

performance was verified.









To determine if there were any differences in the scores due to the different

viewing conditions, the same type of ANOVA was performed for the different viewing

conditions as was done for the observer groups. The p-values for the mean total score for

each viewing condition across all groups and current-time product values was 0.549 and

0.098 for the TO. 10 and the UF Radiology phantom, respectively. The p-values for the

ANOVAs performed at each current-time product independently are shown in Table 6-

18. This analysis showed that there was no statistically significant difference in observer

performance for the different viewing conditions.

CNRT Determination

The purpose for determining the CNRT for each object size was to automate the

scoring of the contrast-detail phantoms. Since there was no statistical difference in

observer performance for the four viewing conditions the observer scores for the standard

image with the use of window and level were used for all subsequent analysis. This

viewing condition was chosen because it best represents the true clinical viewing

conditions. Utilizing the Matlab m-file shown in Appendix C the CNR for all of the

objects in each phantom was calculated for each image used in the observer study. It was

then necessary to determine the object that was considered the threshold of visibility for

each object size, or row, across all observers. Since each observer counted the total

number of visible objects in each row, these values were averaged to determine the object

that the average observer would consider the threshold of visibility. A non integer value

was allowed to add precision to the calculation of the CNRT. These average values are

listed in Table 6-19 and Table 6-20. The CNRT was then calculated for this threshold

object by linearly interpolating between the CNRs for the two objects on either side if the

average threshold object.









The CNRTS for each phantom are shown in Figure 6-27 and Figure 6-28. The

CNRT for each row is nearly independent of the exposure level to the image plate

especially for the UF Radiology phantom. This behavior became less pronounced as the

object size decreased. It should be noted that CNRTs were not calculated for all of the

rows in each phantom. The last three rows (smallest three object sizes) of the TO. 10

phantom and the last row of the UF Radiology phantom were excluded from the CNRT

analysis. This was done because the small size of the objects produced only a few pixels

in the image. The small number of pixels led to large relative errors in the reproducibility

of the CNR calculations.

The manual image registration procedure discussed previously and incorporated in

the Matlab m-file shown in Appendix C does have a slight disadvantage to a completely

automated procedure. Depending on the exact location of the user input from the image

during the registration procedure (see Appendix C), the CNR calculations for each object

can vary. Even with this limitation the manual registration procedure produced better

results than a completely automated procedure. If a completely automated procedure

produced inadequate localization of the objects there was no way to correct the problem.

The m-file produced the same CNR values each time but they were incorrect. The m-file

using the manual registration procedure is capable of exactly reproducing the CNR

calculations for multiple analyses of the same image if the registration inputs are

identical. This ability to exactly reproduce the CNR calculations was difficult due to the

slight variations of the user input during the registration procedure so there were

variations in the reproducibility of the CNR calculations. These variations were less than

five percent for the larger objects but increased to as much as 30 percent for the small









objects. To limit the effect of these variations some object sizes were excluded and the

CNR calculations were performed five times on each image and the results were

averaged. The CNRT determined for each object size was then averaged over all current-

time product values to determine the CNRT to be used in the automated scoring algorithm

described below. The CNRT for each object size and each phantom is shown in Table 6-

21.

Automated Phantom Scoring The Average Observer

Once a CNRT was determined for each object size both contrast-detail phantoms

could be scored without the use of observers and a CDSc for the average observer

determined. In order to increase the accuracy of the auto-scoring process the CDSc is an

interpolated value and not a whole number. This procedure is explained in Appendix C.

The total number of visible objects for the object sizes used in this evaluation and the

CDSc for all current-time product values for both phantoms is shown in Table 6-22. The

CDScs in Table 6-22 were calculated from the same images used in the observer study.

The auto-scoring algorithm performs very well except at the lowest current-time product,

the 0.4 mAs level. This is due to the fact that the CNRT is not exactly independent of

exposure level. Since the CNRTs calculated from the images acquired at 0.4 mAs were

slightly below the average CNRT for all object sizes and the CNR decreases with

decreasing subject contrast, the use of a slightly higher CNRT than was actually measured

from the observer data caused fewer objects to be scored as visible. This caused the

CDSc to be smaller than the observers' score at this current-time product level.

Automated scoring reproducibility

Due to the variations that can occur in the calculation of the CNR for each object in

an image, the automated scoring algorithms used to calculate the CDSc (see Appendix C)









were tested for reproducibility. A single image of each phantom was acquired at each of

the current-time product values shown in Table 6-22 and five CDScs were determined

from each image. A one-way ANOVA was performed on the results to see if this

variability affects the ability of the scoring algorithm to detect small changes in exposure

level. The p-value was 0.000 for the analysis of both phantoms so there were statistical

differences in the mean CDSc determined at different current-time product values. Post-

hoc tests showed that all mean CDScs were statistically different for all current-time

product comparisons with a p-value of 0.000 in all cases. This showed that the variation

in the calculation of the CNR does not affect the performance of the scoring algorithm.

Variation with current-time product

Like the NPSc, the CDSc is capable of distinguishing between small changes in

exposure levels. Five images of the TO. 10 phantom were acquired with the current-time

product levels listed in Table 5-5. The variation of the average CDSc with current-time

product determined from these images, as well as the associated standard deviations, is

shown in Table 6-23. The ANOVA p-value was 0.000 so significant differences exist

among the means. The Levene p-value was 0.227 so the assumption of equal variances

was used in the post-hoc tests. The results of the post-hoc tests are shown in Table E-16.

All of the CDScs statistically differed for all current-time product comparisons. Since the

CDSc differentiated a 0.2 mAs change with the TO. 10 phantom, the CDSc variation with

current-time product for the UF Radiology phantom was done only at 0.8, 1.0, 1.2 and

2.0 mAs. These results are also shown in Table 6-23. The ANOVA p-value was 0.000 so

significant differences exist among the means. The Levene p-value was 0.408 so the

assumption of equal variances was used in the post-hoc tests. The results of the post-hoc

tests are shown in Table E-17. All of the CDScs statistically differed for all current-time




Full Text

PAGE 1

QUANTITATIVE METRICS TO EVALUATE IMAGE QUALITY FOR COMPUTED RADIOGRAPHIC IMAGES By CHRISTOPHER D. PITCHER A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLOR IDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 2004

PAGE 2

To my wife, Brandy Pitcher, for all of her l ove and support and to my parents, Al and Sue Pitcher, who have made this journey possible.

PAGE 3

iii ACKNOWLEDGMENTS I would like to give my most heartfelt thanks to everyone who has assisted me in the completion of this work. In particular, I would like to give a special thanks to Dr. David Hintenlang, chairman of my supervis ory committee, and Dr. Manuel Arreola, cochairman of my supervisory committee. I th ank them both for their continued support and encouragement. I would never have been ab le to complete this work without their expertise and guidance. I would also like to give special th anks to the members of my supervisory committee, Dr.Wesley Bolch, Dr. Kathleen Hintenlang, Dr. Zhihui Fang, Dr. William Properzio and Dr. Jonothan Williams for their support and understanding. I would also like to extend my deepest appr eciation to Dr. Lynn Rill for serving as an alternate committee member at my or al defense on such short notice. I would like to thank the st udents of the Nuclear and Radiological Engineering Department with whom I have had the pleasur e of working for the past three years. I would especially like to thank James Brindle fo r the use of his kitchen table. Without that piece of furniture the qualifying exam would have been the end of my scholastic career. I would also like to thank the United Stat es Army Medical Department for giving me this opportunity as well as continued funding of my research. Finally I would like to thank my family for all of their love and support. Most importantly I would like to thank my parents, Al and Sue Pitcher, for instilling in me the values and morals to succeed in life; and my wife Brandy, whom I love and adore with all of my being, for her unwavering love devotion, support and encouragement.

PAGE 4

iv TABLE OF CONTENTS page ACKNOWLEDGMENTS.................................................................................................iii LIST OF TABLES............................................................................................................vii LIST OF FIGURES...........................................................................................................xi ABSTRACT.....................................................................................................................xi v CHAPTER 1 INTRODUCTION........................................................................................................1 2 BACKGROUND INFORMATION.............................................................................8 Computed Radiography................................................................................................8 Picture Archiving and Communications System........................................................10 Mathematical Transforms...........................................................................................11 Fourier Transform...............................................................................................11 Radon Transform.................................................................................................12 Noise Power Spectrum...............................................................................................13 Modulation Transfer Function....................................................................................16 Detective Quantum Efficiency...................................................................................17 Exposure.....................................................................................................................18 Contrast-Detail Analysis.............................................................................................18 Visual Perception.................................................................................................18 Contrast-to-Noise Ratio.......................................................................................19 Contrast-Detail Phantoms....................................................................................20 Observer Perception Study..................................................................................21 CNR Calculations................................................................................................21 Measurement of Dose.................................................................................................22 Anthropomorphic Phantom.................................................................................23 Dosimeter............................................................................................................23 Determination of Effective Dose.........................................................................26 3 LITERATURE REVIEW...........................................................................................33 Noise Power Spectrum...............................................................................................33

PAGE 5

v Modulation Transfer Function....................................................................................35 Detective Quantum Efficiency...................................................................................37 Contrast-Detail Analysis.............................................................................................38 MOSFET Dosimetry...................................................................................................41 4 EXPERIMENTAL EQUIPMENT..............................................................................42 Imaging System..........................................................................................................42 Software......................................................................................................................4 2 NPSC, MTFC and DQEC.............................................................................................43 Clinical Contrast Detail Score....................................................................................44 Dosimetry...................................................................................................................45 5 EXPERIMENTAL METHODOLOGY......................................................................48 Image Acquisition.......................................................................................................48 Experimental Setup.............................................................................................48 Acquisition Parameters........................................................................................49 Image Retrieval...................................................................................................50 Clinical Noise Power Spectrum..................................................................................50 Variation with Peak-Tube Pote ntial and Current-Time Product.........................50 Variation with Processing Option.......................................................................51 Clinical Modulation Transfer Function......................................................................51 Edge Angle Verification......................................................................................51 Variation with Peak-Tube Potential, Current-Time Product and Processing Option..............................................................................................................52 Clinical Detective Quantum Efficiency......................................................................52 Clinical Contrast Detail Score....................................................................................55 Observer Study....................................................................................................55 CNR Calculations................................................................................................56 Variation with Processing Option.......................................................................57 Anthropomorphic Phantom Viewer Study.................................................................57 Dosimetry...................................................................................................................58 6 RESULTS AND DISCUSSION.................................................................................65 NPSC...........................................................................................................................65 Variation with Current-Time Product.................................................................65 Variation with Peak-Tube Potential....................................................................66 Variation with Processing Option.......................................................................67 MTFC..........................................................................................................................67 Scan Versus Subscan Direction...........................................................................68 IMTFC Reproducibility........................................................................................68 Edge Angle Reproducibility................................................................................69 Number of ESF Data Points................................................................................69 Variation with Current-Time Product.................................................................70 Variation with Peak-Tube Potential....................................................................72

PAGE 6

vi Variation with Processing Option.......................................................................73 Dynamic Image Manipulation....................................................................................74 DQEC..........................................................................................................................77 Variation with Current-Time Product.................................................................78 Variation with Peak-Tube Potential....................................................................79 Variation with Processing Option.......................................................................79 CDSC...........................................................................................................................80 Observer Study....................................................................................................80 CNRT Determination...........................................................................................82 Automated Phantom Scoring The Average Observer......................................84 Automated scoring reproducibility...............................................................84 Variation with current-time product.............................................................85 Variation with peak-tube potential...............................................................86 Variation with processing option.................................................................86 Automated Phantom Scoring The Ideal Observer............................................86 Anthropomorhic Phantom Evaluation........................................................................88 Effective Dose Calculation.........................................................................................89 7 CONCLUSIONS AND FUTURE WORK...............................................................114 APPENDIX A INPSC MATLAB M-FILE AND DESCRIPTION...................................................125 B IMTFC MATLAB M-FILE AND DESCRIPTION..................................................134 C MATLAB M-FILE FOR CONTRAST-D ETAIL PHANTOM SCORING AND DESCRIPTION........................................................................................................142 D MCNP INPUT FILES...............................................................................................155 E RESULTS OF THE POSTHOC STATISTICAL TESTS......................................166 LIST OF REFERENCES.................................................................................................179 BIOGRAPHICAL SKETCH...........................................................................................184

PAGE 7

vii LIST OF TABLES Table page 2-1 ICRP Publication 60 tissue weighting factors..........................................................32 5-1 Technique factors for the curr ent-time product variation study...............................62 5-2 Technique factors for the peak -tube potential variation study.................................63 5-3 Processing options compared...................................................................................63 5-4 Published mass energy-absor ption coefficients for air............................................63 5-5 Technique factors for the TO.10 phantom...............................................................63 5-6 Technique factors for the UF Radiology Phantom...................................................64 5-7 Five-point scale for subjective image quality evaluation.........................................64 6-1 Average INPSCs and their associated standard deviations.....................................106 6-2 Average INPSCs and their associated standard deviations.....................................106 6-3 Average INPSCs and their associated standard deviations.....................................106 6-4 Edge comparison for the IMTFC (mm-1) in the scan and s ubscan directions.........107 6-5 Calculation based variability in the IMTFC............................................................107 6-6 Image based variability in the IMTFC....................................................................107 6-7 Edge angle in degrees for repeated se tups of the edge device and repeated images of the same setup........................................................................................107 6-8 IMTFCs for various ESF lengths in the scan direction...........................................108 6-9 IMTFCs for various ESF lengths in the subscan direction.....................................108 6-10 Average IMTFCs and their associated standard deviations....................................108 6-11 The INPSC and IMTFC value for different window and level settings..................108

PAGE 8

viii 6-12 The value of Q for each peak-tube potential..........................................................109 6-13 In bucky exposure measurements at 60 kVp for the DQEC variation with current-time product study.....................................................................................109 6-14 In bucky exposure measurements for the DQEC variation with peak-tube potential study........................................................................................................109 6-15 Number of visible objects in the TO.10 phantom..................................................109 6-16 Number of visible objects in the UF Radiology phantom......................................110 6-17 ANOVA p-values for total score comparisons by observer group........................110 6-18 ANOVA p-values for total score comparisons by viewing condition...................110 6-19 Average threshold object for the TO.10 phantom..................................................110 6-20 Average threshold object for the UF Radiology phantom.....................................111 6-21 CNRT for each object size and each phantom........................................................111 6-22 Observer study and CDSC comparison...................................................................111 6-23 Variation of the CDSC with current-time product..................................................112 6-24 Radiology residents evaluation of anthropomorphic phantom images..................112 6-25 The exposed bone sites and thei r associated BMFs and BSFs...............................112 6-26 Mass energy-apsorption coefficients for the four tissue types at 35 keV..............112 6-27 Calculation of the Effective dose at 60 kVp...........................................................113 6-28 Calculation of the absorbed dose to the bone marrow...........................................113 6-29 Calculation of the absorbed dose to the bone surface............................................113 6-30 Effective doses for clinical pe diatric current-time product levels..........................113 7-1 Metrics evaluated at 60 kVp, 1.0 mAs with the ches t PA processing option........123 7-2 Object diameters in image pixels for each row of both contrast-detail phantoms........................................................................................123 7-3 Physical dimensions in mm of the objects in the proposed contrast-detail phantom using drilled holes for subject contrast....................................................124

PAGE 9

ix 7-4 Physical dimensions in mm of the objects in the proposed contrast-detail phantom using lead disks for subject contrast........................................................124 C-1 The column and row (x,y) indicies for the objects in the TO.10 phantom............153 C-2 Diameter of objects in each row.............................................................................154 E-1 Post-hoc INPSC mean comparisons for variatio ns with current-time product.......166 E-2 Post-hoc INPSC mean comparisons for variatio ns with peak-tube potential.........167 E-3 Post-hoc INPSC mean comparisons for varia tions with processing option............167 E-4 Post-hoc scan IMTFC mean comparisons for variations with current-time product...............................................................................................168 E-5 Post-hoc subscan IMTFC mean comparisons for variations with current-time product....................................................................................................................169 E-6 Post-hoc scan IMTFC mean comparisons for variations with peak-tube potential..................................................................................................................170 E-7 Post-hoc subscan IMTFC mean comparisons for variations with peak-tube potential.................................................................................................170 E-8 Post-hoc IMTFC mean comparisons for variations with processing option in the scan direction..........................................................................................................171 E-9 Post-hoc IMTFC mean comparisons for variations with processing option in the subscan direction....................................................................................................171 E-10 Post-hoc scan IDQEC mean comparisons for variations with current-time product...............................................................................................172 E-11 Post-hoc subscan IDQEC mean comparisons for variations with current-time product....................................................................................................................173 E-12 Post-hoc scan IDQEC mean comparisons for vari ations with peak-tube potential..................................................................................................................174 E-13 Post-hoc subscan IDQEC mean comparisons for variations with peak-tube potential.................................................................................................174 E-14 Post-hoc IDQEC mean comparisons for variations with processing option in the scan direction..........................................................................................................175 E-15 Post-hoc IDQEC mean comparisons for variations with processing option in the subscan direction....................................................................................................175

PAGE 10

x E-16 Post-hoc CDSC mean comparisons for variatio ns with current-time product for the TO.10 phantom...........................................................................................176 E-17 Post-hoc CDSC mean comparisons for variatio ns with current-time product for the UF Radiology phantom...............................................................................177 E-18 Post-hoc CDSC mean comparisons for variations with processing option for the TO.10 phantom.......................................................................................................178 E-19 Post-hoc CDSC mean comparisons for variatio ns with processing option for the UF Radiology phantom....................................................................................178

PAGE 11

xi LIST OF FIGURES Figure page 2-1 The process of photostimul able luminescence in a PSP..........................................27 2-2 Optical spectra used in CR (adapted from Bushberg)..............................................27 2-3 Representation of two images (a and b) and their associated Fourier transforms (c and d)....................................................................................................................28 2-4 Graphical representation of the Radon transform....................................................28 2-5 Representation of an ideal e dge profile and an ideal slit..........................................29 2-6 Radiograph of the UF radiology phantom................................................................29 2-7 Radiograph of the TO.10 phantom...........................................................................30 2-8 Radiograph of Bowers anthropomorphic phantom.................................................30 2-9 Cross-section of a p-type MO SFET (adapted from Zeghbroeck)............................30 2-10 MOSFET dosimetry system manu factured by Thomson and Neilson.....................31 2-11 Figure showing actual size of the active region of the MOSFET dosimeter...........31 4-1 One-year old Bower stylized anthropomorphic phantom........................................45 4-2 View of the top of the phantoms tr unk showing the MOSFET access ports and the spine (top-center)................................................................................................46 4-3 View of the bottom of the phantoms trunk.............................................................46 4-4 View of the bottom of the phantoms head..............................................................47 5-1 Image acquisition setup............................................................................................60 5-2 Image acquisition setup fo r the flat-field images.....................................................60 5-3 Device to ensure proper a ngulation of edge device.................................................61 5-4 Final setup of the edge device for the determination of the MTFC..........................61

PAGE 12

xii 5-5 The Agfa imaging plate inside the open imaging cassette.......................................62 5-6 Idealized graphical repr esentation of the grid..........................................................62 6-1 Variation of the INPSC with current-time product at 60 kVp..................................90 6-2 Variation of the INPSC with peak-tube potential fo r a constant exposure level......90 6-3 Variation of the INPSC with processing option........................................................91 6-4 Relative pixel values across the edge from an image acquired at 60 kVp and 1.0 mAs..........................................................................................................................91 6-5. MTFC for 25 and 128 data points in the s can direction from an edge image acquired at 60 kVp and 1.0 mAs..............................................................................92 6-6 Variation of the IMTFC in the scan direction with current-time product.................92 6-7 Variation of the IMTFC in the subscan direction with current-time product...........93 6-8 Five MTFC curves in the scan direction from five separate edge images acquired at 0.4 mAs................................................................................................................93 6-9 Five MTFC curves in the scan direction from five separate edge images acquired at 3.2 mAs................................................................................................................94 6-10 Smoothing effect of averaging the MTFC before integration...................................94 6-11 Average MTFC in the scan direction calculated from five images acquired at both 0.4 mAs and 3.2 mAs.......................................................................................95 6-12 Variation of the IMTFC in the scan direction with peak-tube potential for a constant exposure level............................................................................................95 6-13 Variation of the IMTFC in the subscan direction w ith peak-tube potential for a constant exposure level............................................................................................96 6-14 Change in the ESF with peak-tube potential............................................................96 6-15 Variation of the IMTFC in the scan direction with processing option.....................97 6-16 Variation of the IMTFC in the subscan direction with processing option................97 6-17 ESFs for the three processing options in the scan direction.....................................98 6-18 MTF C for the full-range and the hand AP processing options..................................98 6-19 The MTFC for different window and level se ttings in the scan direction................99

PAGE 13

xiii 6-20 Variation of the IDQEC with current-time product in the scan direction.................99 6-21 Variation of the IDQEC with current-time product in the subscan direction.........100 6-22 Variation of the IDQEC with peak-tube potential in the scan direction.................100 6-23 Variation of the IDQEC with peak-tube potential in the subscan direction...........101 6-24 Variation of the IDQEC in the scan direction with processing option....................101 6-25 Variation of the IDQEC in the subscan direction with processing option..............102 6-26 Variation of the DQEC in the subscan direction with processing option...............102 6-27 CNRT for each object size as a function of acquisition current-time product for the TO.10 phantom.................................................................................................103 6-28 CNRT for each object size as a function of acquisition current-time product for the UF Radiology phantom....................................................................................103 6-29 Variation of the CDSC with processing option for the T0.10 phantom..................104 6-30 Variation of the CDSC with processing option for the UF Radiology phantom....104 6-31 Repeated CNR calculations of two adjacent background regions.........................105 6-32 The CDSC for the ideal observer and the TO.10 phantom.....................................105 6-33 The CDSC for the ideal observer and the UF Radiology phantom.........................106 B-1 Input image for the IMTF C m-file..........................................................................139 B-2 Representation of the extracted usable edge data and the co ordinate transform....140 B-3 MTF C of an ideal three-degree bina ry edge with a bin width of 0.5.....................140 B-4 Binned data for the thre e-degree binary edge........................................................141 C-1 Geometrical relationship used in the localization of an object..............................152 D-1 A geometrical depiction of the cells defined in the table top input file.................164 D-2. A geometrical depiction of the cells de fined in the image receptor input file.......165

PAGE 14

xiv Abstract of Dissertation Pres ented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy QUANTITATIVE METRICS TO EVALUATE IMAGE QUALITY FOR COMPUTED RADIOGRAPHIC IMAGES By Christopher D. Pitcher December 2004 Chair: David E. Hintenlang Cochair: Manuel M. Arreola Major Department: Nuclear and Radiological Engineering Traditional methods of eval uating a computed radiography (CR) imaging systems performance. The noise power spectrum (NPS), the modulation transfer function (MTF), the detective quantum efficiency (DQE) and contrast-detail analysis) were adapted in order to evaluate the feasibility of identifying a quantitative metric to evaluate image quality for digital radiographic images. The addition of simula ted patient scattering media when acquiring the images to calculate these parameters altered their fundamental meaning. To avoid confusion with other resear ch they were renamed the clinical noise power spectrum (NPSC), the clinical modulati on transfer function (MTFC), the clinical detective quantum efficiency (DQEC) and the clinical contrast detail score (CDSC). These metrics were then compared to the subjectiv e evaluation of radiogr aphic images of an anthropomorphic phantom representing a 1-year-old pediatric patient.

PAGE 15

xv Computer algorithms were developed to implement the traditional mathematical procedures for calculating the system performance parameters In order to easily compare these three metrics, the integral up to the sy stem Nyquist frequency was used as the final image quality metric. These metrics are identified as the INPSC, the IMTFC and the IDQEC respectively. A computer algorithm was al so developed, based on the results of the observer study, to determine the threshold contrast to noise ratio (CNRT) for objects of different sizes. This algorithm wa s then used to determine the CDSC by scoring images without the use of observers. The four image quality metric s identified in this study were evaluated to determine if they could distinguish between small ch anges in image acquisition parameters (e.g., current-time product and peak-tube potential). Al l of the metrics were able to distinguish these small changes in at least one of the im age acquisition parameters but the ability to digitally manipulate the raw image data made the identification of a broad indicator of image quality not possible. The contrast-d etail observer study revealed important information about how the noise content in an image affects the low-contrast detectability of different sized ob jects. Since the CNRT for each object size in the contrast-detail phantoms was almost independent of the exposure level, the minimum CNRT that would be necessary for an object of that size to be "visible" in a clinical image was identified. Finally, in order to determine more refined CNRT values (due to possible observer biases from the physical construction of the contra st-detail phantoms avai lable for this study) the design of new contrast detail phantoms is proposed.

PAGE 16

1 CHAPTER 1 INTRODUCTION Image quality is one of the most import ant aspects of diagnostic radiology. The concept of image quality has been undergoing a transformation with the widespread use of digital-projection radiography. Imaging modalities, such as computed radiography (CR), are beginning to replace standard sc reen-film imaging systems. While many aspects of the imaging system remain uncha nged, the processing of the image receptor and the viewing environment for the resul ting radiographic images are significantly different. The radiologist is no longer limited to the single set of data represented by the light transmitted through a piece of film. Th ey now have the ability to digitally manipulate the image data set and have access to the full dynamic range of information contained in that data set. Traditionally, radiographic image quality is evaluated in tw o separate ways, quantitatively or qualitatively. Quantitative me thods, such as the calculation of the noise power spectrum (NPS) or the modulation tran sfer function (MTF), are tests of the imaging system performance. They convey th e ability of the system to transfer the anatomical or physiological information of the patient to a radiographic image, henceforth referred to simply as an image. These two quantities can be combined to produce the detective quantum efficiency (DQE), which is the currently accepted standard for imaging system performance. The images used to calculate these metrics are acquired under conditions that allow the ev aluation of the imaging chain without the interference of additional scat tering media (e.g., a patient, ex am table, or grid) between

PAGE 17

2 the x-ray source and the image receptor. Semiqualitative methods, such as contrast-detail analysis, are tests of an imaging systems performance as perceived by human observers. These observers subjectively evaluate, or score, images of contrast-detail phantoms. This identifies the perceived contrast-detectability in images produced by the imaging system. The results of this type of study can be used to establish threshold contrast-to-noise ratios (CNRT) based on the size of the objects in the phantom to determine the total number of visible objects. This type of study will provide a quantitative parameter based on a semiqualitative evaluation that does not have to be repeated. Therefore, it would be ideal to develop a method of applying these evaluati ons of imaging system performance to a perception of clinical image quality. System performance evaluations could be performed with images acquired under more realistic clinical conditions (e .g., with the use of additional scattering media between the x -ray source and the image receptor). These quantitative parameters, in c onjunction with qualitative evalua tions of simulated clinical images of a pediatric patient, could be used to infer the quality of an image that would be produced by the imaging system with those same acquisition parameters (e.g., peak-tube potential and current-time product). The radiation dose a patient re ceives in order to produce an image is also a concern and is especially important for pediatric patient s. Pediatric patients may be at the greatest risk of potential adverse radi ation effects for the following reasons: their growing tissues are more susceptible to radiation effects th an mature adult tissue s; their skeleton, an organ of high radiation sensitivity, encompasse s a greater fractional distribution of active bone marrow; their greater post-exposure li fetime increases the possibility for any radiation-induced effects to manifest; childr en may be uncooperative and are frequently

PAGE 18

3 subject to a greater number of exposures th an adult patients; a nd pediatric patients generally have a larger fracti on of their anatomy located with in the x-ray field compared to adults having similar exams and projections.1 Despite the precise testing of imaging syst em performance, the true test of the quality of an image is whether or not the radiologist can extract the needed diagnostic information from an image in order to make the correct diagnosis. Therefore, it is necessary to identify the patient radiation dos e that is necessary to produce an image of minimum acceptable quality that still has the ne cessary diagnostic information. The most effective way to do this would be to image patients with a known condition at continuingly decreasing exposure levels until th e diagnosis is no longe r possible. This is not a viable option for several reasons. Acqui ring multiple images of the same patient would dramatically increase th e radiation dose, which is pr ecisely what a study of this type is trying to limit. In addition, any adve rse effects that might occur from the radiation exposure would be magnified for pediatric pati ents. An alternative to this method is to use an anthropomorphic phantom in place of real patients. The disadvantages in this approach are that the phantom only approxima tes the anatomy of a human and in most cases there will be no disease present. The radiologist will then have to make a subjective evaluation of the quality of the image. Despite these limitations, the image acquisition parameters necessary to acquire an image of minimum acceptable quality could be determined. Once the acquisition parameters requir ed to obtain an image of minimum acceptable quality are determined from an anthropomorphic phantom study, quantitative metrics can be calculated from images obtaine d with the same acquisition parameters. In

PAGE 19

4 order to ensure that the images used to calcu late these metrics are more representative of a clinical image, additional scattering medi a will be placed between the x-ray source and the image receptor to simulate the presence of a pediatric patient. The quantitative value of these metrics can then be directly linked to the image acquisition parameters required to produce an image of minimum acceptable quality determined from the anthropomorphic phantom study. In the future this will allow the performance of a simple quantitative test instead of a lengt hy observer study to determine the acquisition parameters necessary to create an image of minimum acceptable quality for a specific type of exam. The main objective of this research is to evaluate the feasibility of developing a quantitative metric that can be used to set image acquisition parameters for CR exams without the use of complicated observer studies. A metric of this type will allow images to be produced with the minimal dose to the pa tient that ensures the image is of sufficient quality for a radiologists evaluation. This work will investigate the use of three traditional quantitative methods and one semi-qualitative method of evaluating radiological-imaging system performance as a direct measure of clinical image quality: the calculation of the system NPS, MTF and the DQE, as well as the evaluation and scoring of contrast-detail phantoms. Other re searchers have applied these measures of system performance as a direct measure of image quality.2-9 These methods will be applied specifically to pediat ric CR imaging. Since the images used in this research to calculate these performance parameters are not acquired under the same conditions as in imaging system evaluation (with the inclusi on of additional scattering media), they are not strictly representative of their definiti on. The traditional mathematical procedures are

PAGE 20

5 followed in the calculation of the following image quality metrics evaluated in this research: the clinical noise power spectrum (NPSC), the clinical m odulation transfer function (MTFC) and the clinical detective quantum efficiency (DQEC). The introduction of additional scattering media does not change the technical aspects of contrast-detail analysis and the image quality metric derived fr om this analysis is the clinical contrastdetail score (CDSC). The objectives of this research are listed below. They will show if these new metrics can be used as measures of clini cal image quality, using images acquired under non-traditional condi tions, with an inherently digital imaging system. A. To determine if the traditional methods of calculating the NPS, MTF and DQE can be applied to images acquired in a nontraditional manner at diagnostic exposure levels. B. To quantify the variation of the NPSC and MTFC with current-time product and peak-tube potential. Since the MTFC is not expected to vary strongly with currenttime product for a set peak-tube potential, it mu st be determined if the same is true for a set current-time product and a varying pe ak-tube potential. If this is the case, it may not be necessary to calculate the DQEC as a metric. C. To quantify the effect of any computer pr ocessing of the raw data before the image is viewed on the evaluations of the NPSC, MTFC and the CDSC. D. To quantify the effect of grayscal e inversion on the ev aluation of the CDSC. Two different contrast detail phantoms will be used in this work and each produces contrast in a different way. One has atte nuating objects (creating light objects on a dark background) and the other radio-opaque objects or holes (creating dark objects on a light background). E. To determine if the calculation of the NPSC and the MTFC with the original image data is directly applicable, as a measure of image quality, to an image displayed on the radiologists monitor that has be en manipulated (e.g., the dynamic use of window and level controls). Since the radiologist now has the ability to manipulate an image during the viewing process, the data set that the ra diologist sees has changed (the data set available for diagnos is is now the brightness of each pixel on the monitor). Since the image can be continuously changed, the information available to calculate quantitative image qua lity metrics is not the same information that a viewer is seeing.

PAGE 21

6 F. To quantify the effect of this dynamic image manipulation on the determination of the CDSC through observer studies. G. To correlate the newly developed im age quality metrics with a qualitative evaluation of anthropomorphic phantom im ages and identify the radiation dose associated with the exam parameters used to obtain those images. The remainder of this dissertation provides a detailed description of the equipment utilized, approach and methods used, tools de veloped, results and conclusions reached in the search for a quantitative metric appli cable to clinical image quality. Chapter 2 contains background information and defines cri tical terminology so that a reader new to this area can fully understand the remaining chap ters. Chapter 3 contai ns a review of the current literature that describes the calcul ation of the previously mentioned system performance parameters, their applications as measures of image quality and the various methods used in conducting a contrast-det ail study. Chapter 4 provides a detailed description of the equipment, both hardware and computer software, used in this research. Chapter 5 contains the experiment al methodologies utilized to answer the questions at the heart of the objectives of th is research. Chapter 6 provides the results and discussion of the experiments that were ca rried. Chapter 7 summari zes the findings of this research and provides a description of th e applicability of the image quality metrics, as well as their implications to the research objectives, and th en discusses future proposed work related to this research. In addition to the main text, four appendices have been included that give a detailed description of the computer codes and program input files that were utilized. The text of those codes and input files are provided, as well as a plain language description of the construction and functionality of those files. It is this descri ption that should assist the

PAGE 22

7 reader in implementing the methodology develo ped through this work. Also included in the appendices is a descrip tion of the benchmarking of the MATLAB codes that ensures their proper functioning.

PAGE 23

8 CHAPTER 2 BACKGROUND INFORMATION This chapter first gives an overview of CR adapted from Bushberg10 and from Siegel and Kolonder11. There are many sources detai ling the calculation of system performance parameters and applying them as measures of image quality. These sources have been compiled and a cohe rent description of the meth ods used to calculate an imaging systems NPS, MTF and DQE is presented. There are many different methodologies covered under the broad topic of contrast-detail analysis. A summary of the different aspects of this type of analys is is described. Finall y, the calculation of effective dose is presented. Computed Radiography The use of a photostimulable phosphor (PSP) as an image receptor was introduced in the 1970s but did not come into wide sp read use until the turn of the century.10 The term CR did not come into use until 1981 wh en Fuji introduced the PSP imaging plate and named the technique FCR.3 The first commercial CR, the Model 101, was introduced by Fuji into clinical pr actice in Japan in 1983.11 A typical CR imaging plate, or PSP, is made from a combination of BaFBr and BaFI, doped with a small amount of europium (Eu). A PSP acts in a similar manner to the phosphors used in screen-film radiography (e.g., Gd2O2S). The difference in the functioning is the process of photostimulabl e luminescence. The intensifying screens used in screen-film radiography emit visibl e light promptly upon exposure to x rays. A PSP, while also promptly emitting a small amount of visible light upon exposure to x

PAGE 24

9 rays, traps the majority of the absorbed x-ray energy. The PSP can then be stimulated with a laser and the trapped x-ray energy is released. The details of this process are graphically represen ted in Figure 2-1. When x-ray energy is absorbed in a CR im aging plate, the electrons of the Eu atoms (Eu+2) are excited to the conduction band and Eu+3 is produced. These excited electrons migrate and are tr apped by fluorine atoms (F+), which then become non-ionized F. These are referred to as F-centers. Thes e trapped electrons form the latent image in the imaging plate. When the exposed imag ing plate is scanned by a red laser, the electrons trapped in the F-cen ters are excited back to the conduction band where they become mobile again. They then can de-exc ite back into the electron hole of the Eu+3 by releasing blue-green visible li ght (see Figure 2-2). This light is collected by a fiberoptic light guide, which directs the light to a photom ultiplier tube where an electronic signal is generated. This electronic signa l is digitized and stored, cr eating an inherently digital image. When the imaging plate is first read, not all of the stored energy is released. In order to reuse the plate, it must be erased by exposure to a high intensity light source.10,11 There are many advantages to acquiring an image this way. First of all, there is no longer a need for a film proce ssor and there are no longer ch emicals to worry about. This reduces the workload on a number of front s from mechanical maintenance to waste disposal (e.g., silver recovery). There is no longer the need for a dark room which will free up valuable space in what seems to be ever shrinking budgets a nd facilities. All of this equates to an overall cost savings fo r the radiology department. A CR imaging plate also has a much wider latitude, or dynamic range, than a screen-film system. This means that a CR plate can produce a usable image over a much wider range of exposures than a

PAGE 25

10 screen-film system. This can dramatically reduce the amount of retakes. This increased latitude can also lead to problems becau se the abilities to overexpose and underexpose a CR plate and still get a good image are both possi ble. This could lead to an unnecessary increase in patient dose, a phenomenon called dose creep. While a problem in the early days of CR, this problem has been addr essed by the CR manufacturers. All major manufacturers of CR systems provide an exposur e indicator that allo ws the technologist to monitor the exposure level to the plate. Nevertheless, by far the most important advantage is the ability to digi tally manipulate the image, whic h will aid the radiologist in the interpretation of the image. Picture Archiving and Communications System A picture archiving and communications system (PACS) is an integrated system of interfaces, computer networks, computer hardwa re and computer software for the storage and transfer of images. Imaging systems need to be physically linked to the PACS. In the case of CR, this is accomplished through the image plate reader. Once the plate is scanned and the image digitized, it is sent over the local area network to a computer quality assurance (QA) station. At the QA st ation the technologist can view the image, with specialized viewing software, and ensu re that it is acceptable for a radiologists evaluation. The image is sent to the QA sta tion in a standard format known as Digital Imaging and Communications in Medicine (DICOM). This format was sponsored by the American College of Radiology and the Natio nal Electrical Manufact urers Association. The purpose of this standard image format wa s to overcome the difficulties of integrating imaging system components from different manufacturers. Images are sent from the QA stations over a wide area network for storage and archiving. There are usually two levels of stor age, a short term arch ive that allows quick

PAGE 26

11 access to the images and a long term archive for the permanent storage of the images. Database programs are used to manage the large number of images that need to be stored. The images can then be accessed through the network by the physicians and viewed on a local PC with specialized viewing software.10 Mathematical Transforms In order to calculate the NPSC and the MTFC, some specialized mathematical operations are required. The first is the Four ier transform. This transform converts the information contained in an image into its frequency components. The second is the Radon transform. This transform, used prim arily in the reconstruction of computed tomography images, can be used to detect lines in an image. Below is a mathematical, as well as a conceptual description, of bot h of these mathematical operations. Fourier Transform The two dimensional Fourie r transform of a function f ( x,y ) is mathematically represented by dy dx x f vy ux i v u I ) ( )] ( 2 exp[ ) ( (2-1). The Fourier transform can be conceptually described as a mathematical procedure that identifies the magnitude and phase of si nusoidal variations in the intensity of the pixel values as a function of spatial freque ncy. Spatial frequency is analogous to the general term of frequency as applied to a time varying si gnal and has units of inverse distance. The Fourier transform is best described by example. Figure 2-3 depicts an image of a sinusoidal pattern (a) and the modul us of the Fourier tran sform of that image (b). Note the three bright dots, or peaks, in the center of Figure 23-b. The center peak is the zero frequency value, or DC component, and represents the av erage brightness across

PAGE 27

12 the whole image. The other two peaks represen t the sinusoid in the image. The magnitude of those peaks is the difference in bright ness between the dark and light areas. The placement of them is the spatial frequency asso ciated with the sinusoid (the inverse of the distance between the peaks). The orientation of the sinusoid in the image correlates with the orientation of the peaks in the Fourier transform relative to the zero frequency value. This is shown in Figure 2-3-c and 2-3-d. In th is case a tilted sinusoidal pattern creates a tilted pair of peaks in the Fourier transfor m. There are two peaks representing a single sinusoid because the Fourier transform pr oduces a mirror image of itself creating redundant information referred to as the negative frequency values. It is important to note that the Fourier transform just does not identify a singl e sinusoidal variation, but simultaneously breaks down the spatial functi on into a sum of sinusoids that exactly represents the information contained in th e image. Only sinusoids up to a certain maximum frequency (the Nyquist frequency) can be represen ted by the Fourier transform of a digital image. The Nyquist frequency is re lated to the size of th e pixels in the image as size pixel frequency Nyquist 2 1 (2-2). Since the Fourier transform contains all of the information in the original image, the inverse Fourier transform can be applied and the original image can be recovered.12 Radon Transform The Radon transform is an integral transfor m that converts a function f(x,y) into a set of projections p(r, ). Mathematically it can be represented as dy dx y x r y x f r p ) sin cos ( ) ( ) ( (2-3)

PAGE 28

13 where is the Dirac delta function. This process is graphically represented in Figure 2-4. Each projection p, rotated at an angle is a collection of all of the line integrals perpendicular to the projecti on at all distances r from the origin. The complete Radon transform is a collection of th e projections for all angles. Noise Power Spectrum The NPS shows the ability of an imaging system to process noise as a function of spatial frequency. The term power spectrum is so mewhat of a misnomer in this context. It is a carryover from electrical engi neering. If x(t) is the volta ge across (or current through) a one-ohm resistor, the expectation value of the squared modulus of x(t) is the average power dissipated. Since the NPS is simply th e squared modulus of the Fourier transform of an image function, the term power spectrum is used.13,14 The NPS can be thought of as the variance associated with a particular frequency component of an image.4 For a flat-field image, the NPS is the variance associated with a particular frequency component of the noise in that image. In most cases, noise may be represented, or approximated, as a stationary Gaussian random process with zero-mean. Therefore, all of the relevant statistica l properties will be contained in the power spectrum. Unfortunately, exact determinati on of a power spectrum would require a perfectly measured, indefinitely long pi ece of a random function, and would require indefinitely detailed computations.13 We are therefore reduced to approximating the NPS. The power spectrum of a stationary random process can be directly estimated through the periodogram, which is given by 2 0 2) ( 1 ) (N x f jdx e x I N f NPS (2-4)

PAGE 29

14 for a continuous function I(x) of length N.13,15 With CR, we are dealing with a discrete function I(x) that is represented by the pixel values. The NPS estimation then becomes 2 1 0 2) ( 1 N K N k n je x k I x x N x N n NPS (2-5) where the term x, the pixel size, inside the square d modulus symbol arises from the definition of the discrete Fourier transform (DFT).16 The 1/N term is due to Parsevals theorem that states for discrete functions, th e relationship between power as computed in the spatial domain and as computed in the frequency domain is given by 1 0 2 1 0 2) ( 1 ) (N N N kn H N k h (2-6) where H(n) is the Fourier transform of h(k).16 Equation 2-5 can then be simplified to read 2)] ( [ ) ( x I DFT N x f NPS (2-7). This can be easily expanded to two dimensions as follows, 2)] ( [ ) ( y x I DFT N N y x v u NPSy x (2-8) where represents an ensemble average over many NPS calculations from small regions of interest (ROIs) in the same image. In order to use Equation 2-8 to calculate th e NPS, the data need to be prepared in order to obtain the most accurate estimate of the NPS. Due to the nature of this application of the NPS, only one image is available to calculate each NPS. Therefore, a tradeoff must be made between the number of NPS calculations that are averaged and the size of the ROI for each calculation. U tilizing the center 1024 1024 region of each

PAGE 30

15 image to avoid nonuniformities near the edges makes sixty-four 128 128 ROIs available for the calculation of the NPS.4 In order to reduce the effect on the low frequency components of the NPS due to structure in the flat field imag e, the variation in exposure ac ross the image due to the heel effect must be removed. In order to accomplish this, a surface is fit to the data and must be subtracted from the flat-field image.4,5,17 The final data preparation step is to mini mize the effect of having only a finite data set to perform the NPS calculation. A finite data set, mathematica lly, is a rectangular truncation function multiplied by what would be the infinite data set.15 This causes a function of the form sin( f )/ f to be superimposed on the NPS in the frequency domain. In order to reduce this effect, the data must be truncated with a we ighting function that slowly goes to zero at the Nyquist freque ncy. A function of the following form cX x x h2 cos 2 1 2 1 ) ( cX x 0 (2-9), called the Hanning function,16 multiplied by the data set before the Fourier transform is calculated, will accomplish this goal.5,15 It is important that the Hanning function be normalized so the mean-square value of H(x) is equal to unity in order to preserve the magnitude of the NPS.2 In order to calculate the DQE, the one-d imensional NPS must be calculated from the two-dimensional NPS. The one-dimensional NPS can be calculated from a thick slice comprising four lines of data immediately adjacent to the frequency axes for the scan ( u ) and sub-scan ( v ) directions, respectively. Each data point is assigned a frequency value of 2 2v u and then binned. Each bin is averag ed to produce the one-dimensional NPS.4

PAGE 31

16 Modulation Transfer Function The MTF represents the maximum ability of an imaging system to transfer subject contrast, to the final image, as a function of spatial frequency. There are multiple methods for determining the MTF of an imaging system.4,18-21 The two most common are the use of a sharp attenuating edge or an attenuating material with a narrow slit cut through the material (see Figur e 2-5). Since the images of these devices are already digitized, the pixel values can be easily samp led in the direction of the arrows in Figure 2-5. If these pixel values ar e plotted, the sharp edge produces the system edge spread function (ESF) and the narrow slit produces th e system line spread function (LSF). The LSF can also be derived from the ESF. The L SF of a system is the first derivative of the ESF. The MTF is then the modulus of the DFT of the LSF normalized to the value at zero frequency. Since the image is comprised of discrete da ta, a correction needs to be made for the effect of finite differentiation with the edge method. The correction is made in the spatial frequency domain by multiplying the MTF by cf f 2 sinc 1 Function Correction (2-10) where f is frequency and fc is the Nyquist frequency.19 In order to sample the MTF at frequencie s higher than the Nyquist frequency, the slit or edge must be angled with respect to the pixel matrix. The ideal angle is between one and six degrees. Before the pixel values can be sampled, the edge angle must be determined. This can be done using the Radon transform. The distance of each pixel from the edge, along a line perpendicular to the edge must also be determined. This is done be

PAGE 32

17 transforming the coordinates of the pixels into a coordinate system that is rotated by the edge angle. The two coordinate systems are related by the following relationships: sin cos y x xrotated (2-11) sin cos x y yrotated (2-12). The ESF data can then be binned into intervals smaller than the pixel size.20,22 This allows the calculation of the pre-sampled MTF. The MTF calculation also suffers from th e same problem of rectangular data truncation as the NPS calculation. Therefor e, the Hanning function is multiplied by the data set before the Fourier transform is calcu lated. In this case it is important that the Hanning function be normalized so the mean valu e of H(x) is equal to unity in order to preserve the magnitude of the MTF. Detective Quantum Efficiency The DQE is a quantity used to describe the overall signal-to-noise ratio (SNR) performance of a system. Currently, the DQE has become the standard by which digital x-ray imaging systems are measured in the re search environment. It is essentially the ratio of the output SNR square d to the input SNR squared. The DQE can be calculated from the MTF and the NPS as follows: ) ( ) ( ) (2f NPS Q f MTF k f DQE (2-13) where k corrects for the gain of the imag ing system and Q is the number of photons incident on the image receptor (in photons/mm2) used to generate the flat-field image.2,5,23

PAGE 33

18 Exposure The quantities k and Q combined have units of photons/mm2mR. The determination of this quantity is related to the calculation of exposur e (X) for a specified x-ray spectrum. Exposure is defined as the abso lute value of the tota l charge of one sign produced in air when all the el ectrons liberated by photon inte ractions in a given mass of air are completely stopped in air. Exposure can be calculated from the following equation air air E enW e X (2-14) where is the energy fluence of the x-ray spectrum, (en/ )E,air is the energy dependant mass energy-absorption coefficient for air and (e /W) is a constant of 33.97 J/C that is related to the amount of energy required to create an ion pair in air.24 Contrast-Detail Analysis Visual Perception The NPS, MTF and DQE are indicators of the performance of an imaging system up to the point of storage of the final image. In medical imaging, a radiologists perception of the displayed visual inform ation is used to make a diagnosis. The perception of visual information consists of three sequential steps: detection, recognition and interpretation. Contrast-d etail analysis focuses on the task of detection. Various models have been developed to describe how observers detect visual signals in images. One approach is the use of a signal-to-noise model. The signal-to-noise model describes the ability of an observer to detect simple visual signals embedded in a noisy background. The model characterizes an ideal observer who detects signals w ith a certain likelihood if their amplitude differs from the background by a set threshold.25 Hendee et al states, Wagner and Brown26 have suggested that the performa nce of an imaging system can be

PAGE 34

19 characterized quantitatively by de scribing the ability of an ideal observer to detect simple visual signals provided by the system.(25, p. 295) There are two aspects to th e detection of visual info rmation, visual acuity and contrast discrimination. Hendee et al .defines visual acuity as th e ability of an observer to extract detailed information from a small region of an image. This is the ability to detect high spatial frequency signals. Visual acuity is often measured using the Snellen eye test chart. This chart has black letters of vary ing sizes on a white background. If the grayscale on the chart were to be reversed, the ability of observers to recognize the letters from a distance is impaired. Deficiencies in visual acuity can be corrected with eyeglasses. Hendee et al then define contrast discrimination as the ability of the visual system to distinguish differences in bright ness in the image. This is th e ability to detect middle and low spatial frequency signals. Deficiencies in contrast discrimination suggest a neuroopthalmological cause. Hendee et al conclude that, Contrast discrimination is probably a more critical feature than visual acuity in determining how well the average person sees.(25, p. 297) Contrast-to-Noise Ratio The signal in traditional screen film ra diography can be represented by a quantity called radiographic contrast. Ra diographic contrast is the difference between the average optical densities of the region where the object of interest providing the signal is located in the background. Background SignalOD OD Contrast ic Radiograph (2-15) With the use of an inherently digital imaging modality such as CR, the imaging system software often performs a series of proces sing steps before the image is viewed. One

PAGE 35

20 common form of processing is th e subtraction of a constant fro m all of the pixel values in the image. This can lead to problems if traditional notions of contrast, such as subject contrast that is described in the next secti on (see Equation 2-19), are used. If the constant that is subtracted is equal to the average pixel value represented by N1, the subject contrast becomes infinite. Therefore, a more meaningful and frequently used measure of contrast is the contra st-to-noise ratio (CNR) Background Signal CNR (2-16) where is the standard deviation of the pixel values in the background.10 Contrast-Detail Phantoms Many different forms of contrast-detail pha ntoms have been used to evaluate the performance of imaging systems.6,27-29 While the designs may differ, all have one thing in common: the presence of circular objects of varying diameter and varying levels of subject contrast. Subject contrast is the inherent difference in the x-ray attenuating properties between two regions in an object. If the number of x-rays reaching the imaging plate directly below the two regions of interest are N1 and N2 1 10 1 xe N N (2-17) 2 20 2 xe N N (2-18) then subject contrast is defined as 1 2 1N N N Contrast Subject (2-19). While all contrast-detail phantoms have th is design property in common there are two distinct ways in which subj ect contrast can be designed in the phantom. The first is to start with a phantom of uniform thickness and drill circular holes of varying depth. An

PAGE 36

21 example of this type of phantom is the UF Radiology phantom (see Figure 2-6).27 A radiograph of this phantom will produce an im age with dark signals (circular objects) on a light background. The second method is to in sert circular objects with varying x-ray attenuating properties. An example of this type of phantom is the Leeds test object TO.10i (see Figure 2-7). A radiograph of this phantom will produce an image with light signals on a dark background. Observer Perception Study The first task in contrast-detail analysis is to use human observers to evaluate images of contrast-detail phantoms. The imag es are acquired under the conditions that are to be compared or evaluated. The observer is then asked to evaluate, or score, the images. Depending on the design of the phantom, the observer is either asked to identify the presence or absence of the object (or signal) in a specific region of the image or identify the number of objects that are visible. The de finition of what constitutes a present or visible object is dependent upon the design of the study. The observer is instructed on these criteria before the images are evaluated. Even with the most explicit instructions, there is still an amount of subjectivity in each observers interpretation of those instructions. Therein lies the inherent qualitative nature in this type of study. CNR Calculations In an attempt to speed up the evaluation of quality assurance phantoms, several researchers have developed computer algorith ms to identify, localize and mathematically evaluate objects and ROIs in those phantoms.30-32 These completely automated algorithms depend either on some initial prepar ation of the digital image or an a priori knowledge of i Leeds Test Objects Ltd, Wetherby Road, Boroughbridge, North Yorkshire, YO51 9UY, UK

PAGE 37

22 the size of some or all of the objects in th e image. The methods utilized involved the convolution of a mask of the object of interest with part or all of th e original image. For the circular objects, the maxi mum value of the resulting convolution would be the point of maximum correlation between the mask and the object (i.e., the center of the object). These methods are extremely sensitive to the size of the object of intere st in the image. If the mask is not exactly the same size as the ob ject, a precise center can not be located. In order to make an algorithm more broadly applicable to geometrical magnification or phantom rotation a manual localization procedur e can be utilized. If two points in each image are manually identified, all objects a nd ROIs can subsequently be successfully located. Once the positions of the objects in the image have been successfully located, mathematical operations such as the ca lculation of CNR can be performed. Measurement of Dose There are three aspects to the determination of the radiation dose a patient will receive from a specific type of radiographic examination. First, an anthropomorphic phantom that simulates the anatomy of the human body is constructed. This is done so that the attenuated and scatte red x-ray spectrum produced th rough interacti ons with the patients internal organs is reproduced. Secondly, the amount of absorbed x-ray energy to those organs must be measured. This requires the use of very small radiation detectors (or dosimeters) so that the x-ray spectrum in the anthropomorphic phantom is not appreciably disturbed. Lastly, the amount of measured energy deposition in the dosimeter must be converted to the amount of energy that would have been deposited in human tissues, which can be used to calculate a dosimetric quantity.

PAGE 38

23 Anthropomorphic Phantom Mathematical models have been devel oped to simulate the anatomy of human beings. The most widely used of th ese models was developed by Cristy.33 These models are based on the size of the average person of varying age and gender. Bower34 produced a physical phantom based on the Cris ty and Eckerman one-year old pediatric model. It consists of thre e physical regions, head, trunk a nd legs. The head region is modeled after the Bouchet and Bolch model.35 The phantom is composed of three tissue substitute types: soft tissue, bone and lung. The composition of these tissue substitutes is explained in detail by Bower.34 For organs used in the calcul ation of effective dose that are internal, small holes are drilled in the pha ntom so a small dosimeter can be placed at the centroid of the organs. A radiograph of the anthropomorphic phantoms trunk can be seen in Figure 2-8. Dosimeter The radiation dosimeter needed to meas ure the amount of energy deposited in the organs of a pediatric patient must be very small. Not only does a dosimeter have to be small to measure the absorbed dose to pe diatric organs, the measurement of total accumulated dose requires the use of miniature dosimeters that will not perturb the radiation field. The implicit size restricti on narrows the choice of conventional detectors to thermoluminescent dosimeters (TLD s) or semiconductor diode detectors.36 Accurate surface dose measurements can be made with parallel plate ionization chambers, but because of their size are not suitable for internal organ dose measurements.37 The TLD measures cumulative dose, but the reading pro cedure is destructive in the sense that the stored signal is lost after reading. Although small diodes provide dose rate information, which can be integrated to yield the total dose for any time duration, the signal due to

PAGE 39

24 total dose is not retained in the device and any signal lost can not be recovered.36 An alternative option for absorbed dose measur ements is the use of metal oxide-silicon semiconductor field effect transistors (MO SFETs) as radiation dosimeters. The MOSFET detector has a very small size, with co mmercially available systems having active volumes typically 400 um x 500 um x 100 um.38,39 The MOSFET detector also offers the unique advantage of permanently storing the total cumulative dose signal, which can be read at any time in a nondestructive manner. This ensures that the dose information will not be lost during the readout procedure. In addition, the acquired dose signal is dose rate independent.40 These factors combine to make the MOSFET an ideal detector for absorbed dose measurements in an anthropomorphic phantom. The basic MOSFET structure is shown in Fi gure 2-9. This type of MOSFET is a pchannel MOSFET that is built on a negatively doped (n-type) silicon substrate. Two of the terminals of the MOSFET called the source and the drain are situated on top of a positively doped (p-type) silicon region. The th ird terminal shown is the gate. Underneath the gate is an insulating silicon dioxide laye r and under this oxide layer is the n-silicon substrate. The region of the substrate immedi ately below the oxide layer is known as the inversion layer or channel region. When a suffi ciently negative voltage is applied to the gate, with reference to the substrate, a signi ficant number of minor ity carriers (holes in this case) will be attracted to the oxide-silic on surface from both the bulk of the silicon and the source and drain regions. Once a su fficient number of holes have accumulated there, a conduction channel is formed, allowing an appreciabl e amount of current to flow between the source and the drain. The gate voltage needed to allow a predetermined current flow is defined as the threshold voltage (Vth).40,41

PAGE 40

25 During irradiation, a positive voltage is appl ied to the gate. Electron-hole pairs are generated within the silicon dioxide by the incident ionizi ng radiation. The electrons quickly move toward the positively biased gate while the holes migrate toward the SiSiO2 interface. When the holes ge t close to the interface, some of them are captured in long-term trapping sites. These trapped posit ive charges cause a negative shift in Vth. The magnitude of this shift is proportional to th e radiation dose deposite d in the oxide layer.40 The use of a single MOSFET as a dosimeter does have some limitations. A 1C change in ambient temperature can shift Vth by as much as 4-5 mV. Also, the response of a single MOSFET detector (Vth) as a function of accumu lated dose will exhibit a nonlinear region at high dose levels.40 Commercially available MOSFET dosimeters consist of two identical MOSFETs fabricat ed on the same silicon chip. The two MOSFETs are operated at two different positiv e gate biases during irradiation and the difference between the threshold voltage shifts of the two MOSFETS is representative of the absorbed dose. Since the response to temperature of each individual MOSFET is identical, this type of construction renders minimal temperature effects. This type of construction also reduces the nonlin ear response at high dose levels.40 Many investigators36,38,39,42,43 have reported an angular dependence in the sensitivity of the MOSFET dosimeter. Pomije et al.43 investigated this behavior for diagnostic exposure levels from 60 kVp to 120 kVp and fo und that a significan t reduction in the response of the MOSFET occurs when the epoxy bubble faces away from the x-ray source. The black epoxy bubble covering the ac tive MOSFET can be seen in Figure 2-10 and 2-11. These figures picture a commerc ially available MOSFET dosimetry system manufactured by Thomson and Nielsen Electronics Ltd. This effect can be minimized if

PAGE 41

26 care is taken to orient the MOSFET with the epoxy bubble facing th e x-ray source during dosimetric measurements. Determination of Effective Dose The currently accepted dosimetric quantity linking absorbed dose and the probability of stochastic radiation in duced effects is effective dose (E).44 Effective dose is calculated from the following equation R T R R T TD w w E, (2-20) where R signifies the type of radiation, T signifies the type of tissue exposed, wR and wT are the radiation and tissue weigh ting factors, respectively, and DT,R is the absorbed dose to a particular tissue T from radiation type R. In order to calculate the effective dose to a patient, the absorbed dose to the organs listed in Table 2-1 must be measured. The first step in measuring absorbed dos e using the MOSFET as a dosimeter is to calibrate the system. This procedure is described by Bower and Hintenlang.42 This will result in calibration factors that will be able to conver t the MOSFET reading to exposure in Roentgens (R). After the exposure of the phantom has been measured with the dosimetry system, those measurements must be converted to absorbed dose. An exposure of one R produces an absorbed dose of 0.876 rads in air. Therefore, the absorbed dose to tissue can be calculated as follows,34 V CF Doxide en tissue en air en oxide en tissue 876 0 (2-21) where en is the average mass energy absorption coefficient for the specified material in cm2/g, CF is the calibration factor in R/ mV, and V is the MOSFET reading in mV.

PAGE 42

27 Figure 2-1. The process of photos timulable luminescence in a PSP. Figure 2-2. Optical spectra used in CR (adapted from Bushberg).10 Conduction Band Valance Band Eu+2 Eu+3F F+ eeX-ray Red Laser Blue to Green Photon

PAGE 43

28 Figure 2-3. Representation of two images (a and b) and their associated Fourier transforms (c and d). Used with the permission of Dr. Stephen Lehar.12 Figure 2-4. Graphical representa tion of the Radon transform. x y r p

PAGE 44

29 Figure 2-5. Representation of an ideal edge prof ile and an ideal slit. If the pixel values are sampled in the direction of the arrows the ESF and LSF can be determined. Figure 2-6. Radiograph of the UF radiology phantom. ESF LSF Ideal edge profile Ideal slit

PAGE 45

30 Figure 2-7. Radiograph of the TO.10 phantom. Figure 2-8. Radiograph of Bo wers anthropomorphic phantom. Figure 2-9. Cross-section of a p-t ype MOSFET (adapted from Zeghbroeck).41

PAGE 46

31 Figure 2-10. MOSFET dosimetry system ma nufactured by Thomson and Neilson.ii The active MOSFETs are encased in the black epoxy bubbles at the tips of the long brown strips at the bo ttom of the figure. The bias supply is the small box labeled A in the right center of the figur e while the readout device is at the top of the figure. Figure 2-11. Figure showing actual size of the active region of the MOSFET dosimeter. The black rectangular region (3 mm in width) contains the MOSFETs. ii Thomson and Nielsen Electronics Ltd, 25B Northside Road, Ottowa, ON, Canada K2H 8S1

PAGE 47

32 Table 2-1. ICRP Publicati on 60 tissue weighting factors.44 Tissue or Organ Tissue weighting factor Gonads 0.2 Bone marrow (red) 0.12 Colon 0.12 Lung 0.12 Stomach 0.12 Bladder 0.05 Breast 0.05 Liver 0.05 Esophagus 0.05 Thyroid 0.05 Skin 0.01 Bone Surface 0.01 Remainder 0.05

PAGE 48

33 CHAPTER 3 LITERATURE REVIEW Noise Power Spectrum The NPS has been calculated by many rese archers for a wide variety of imaging systems and image receptors. The choice of normalization of the NPS varies slightly from researcher to researcher but the underlyi ng methodology is consistent (the squared modulus of the Fourier transfor m of a flat-field image). In all cases, the researchers are attempting to determine the noise properties of the imaging system, or systems, in question. Flynn and Samei2 determined the two-dimensional NPS of a CR system utilizing two different imaging plate resolutions. Th e measurements were made by exposing the imaging plate to a spatially uniform x-ra y beam while simultaneously measuring the exposure level to the imaging plate. One-hundr ed and forty-four subregions in a 12 12 array were used in the NPS calculation. For th e low resolution mode with pixels of 200 m, the subregions were 128 128. For the hi gh resolution mode w ith pixels of 100 m, the subregions were 256 256. Even though the x-ray field is assumed to be uniform, in reality slight variations exist across the image receptor due to the heel effect. The exposure measurement and the mean signal leve l in different regions of the image were used to correct for large-scale nonuniformities (e.g. the heel effect). The data in each subregion was truncated with the Hanning f unction before the two-dimensional Fourier analysis generated the NPS. The subregions were averaged to produce the final NPS. The

PAGE 49

34 one-dimensional NPS was estimated by averaging the central row or column and rows or columns from the two-dimensional NPS. Fetterly and Hangiandreou5 calculated the NPS of a Lumisys ACR-2000 single plate desktop CR reader with a method similar to Flynn and Samei.2 The primary differences in their methodology were the size of the subregions us ed to calculate the NPS and the subtraction of a planar fit fr om each subregion before calculation of the NPS. Fetterly and Hangiandreou utilized 225 100 100 ROIs from each uniform field image. This selection was stated to be a matter of convenience. The resultant NPS measurements were judged to contain suffien t detail and the frequency increments were equal to those of their MTF measurements. Dobbins et al .4 measured the NPS of four gene rations of CR imaging plates by acquiring a flat-field image with 0.5-mm Cu added filtration. The two-dimensional NPS was computed directly. The center 1024 1024 portion of a single image for each plate was used to calculate the NPS in order to avoid nonuniformities near the edges. This portion of the image was subdivided into 64 128 128 ROIs. They observed that this size for the ROIs was the smallest size that coul d be used without appreciably changing the shape of the average NPS curve near zero fre quency. In order to remove the background trends created by the heel effect, a planar ramp was fitted to the data of each ROI and subtracted from the ROI before the Fourier an alysis. All of the NPS calculations for each ROI were averaged to generate the final NPS. The method used by Flynn and Samei2 to generate the one-dimensional NPS from th e two-dimensional NPS was found to be insufficient by Dobbins et al .4 Instead, the pixels of a thick slice on either side of the axes, excluding the axes, were binned to generate the one-dimensional NPS. The

PAGE 50

35 frequency value assigned to each pixel was com puted as the radial distance of that pixel from the zero-frequency pixel. Modulation Transfer Function Multiple methods for determining the MTF are presented in th e literature. The methods vary from the use of a radio-opaque square, a thin slit cut into a radio-opaque material, a small pin-hole cut into a radio-opaqu e material and a thin wire. In all cases an image is acquired of one of the test devices listed above and a LSF is generated. This LSF is then used to calculate th e MTF through the use of the F ourier transform. Each method is usually given a name that represents th e test device used. For example, the edge method would utilize the radio-opaque square an d the edges of that square are used in the calculation of the MTF. Fetterly and Hangiandreou5 used an angled edge techni que to determine the MTF of a Lumisys ACR-2000 single plate desktop CR re ader. The edge device was custom made and consisted of a 250-m thic k square of lead laminated between two sheets of acrylic, each 1 mm in thickness. A single image was us ed to evaluate the MTF in both the scan and subscan directions. Fetterly and Hangia ndreou processed the storage phosphors in a dimly lit room with only the monitor of the Lumisys control workstation providing light. This leads to the conclusion that the edge images were acquired with the bare imaging plate and it was not in a normal cassette. The an gle the edge device made with respect to the pixel matrix was calculated to be betw een two and five degrees for all images. A 4cm 8-cm region of the image encompassing th e edge was extracted and the pixel values were averaged in bins 0.2 times the pixel size. This produced an over-sampled ESF. The resulting LSF was zero-padded so that the freq uency values of the MTF matched those of the NPS calculated in the same study. The MTF was then corrected for finite element

PAGE 51

36 differentiation as described by Cunningham and Fenster.19 The formula used to calculate this correction is given in Equation 2-10. Cunningham and Fenster state that the MTF resulting from finite differentiation of the E SF contains an error that remains unimportant only if the sampling rate is approximately four times the Nyquist frequency or greater. Samei and Flynn2 constructed a sharp, attenuatin g edge device and designed a procedure for the measurement of the MT F of a digital storage phosphor radiography system. The edge device was constructed from a 250-m thick lead foil laminated between two 1-mm-thick sheets of acrylic. Im ages of the edge device were acquired with the presence of an additional three-mm of aluminum filtration between the x-ray source and the image receptor. The edge device was angled between one and six degrees with respect to the pixel matrix. An 8-cm 8-cm region containing only the edge was extracted from the image. In order to dete rmine the precise edge angle, the data was processed with the following steps; the gray-s cale image is converted to a binary image through a thresholding procedure, a binary line along the edge is produced through the application of a gradient operation and the edge angle is then determined with the use of the Hough transform. The data were then placed in bins with sub-pixel bin widths of 0.1 times the pixel width. The resulting LSF wa s then truncated with a Hanning filter to eliminate the high-frequency content of the measurement not associated with the edge. The pre-sampled MTF was calculated through F ourier analysis of the LSF. This method was compared to a slit method similar to Fujita et al.21 It was determined that the edge method provided similar results to the slit method. The use of a narrow slit to determine the MTF of a CR imaging system is described by Fujita et al.21 This method was also employed by Dobbins et al.4 Images of a 0.01-mm

PAGE 52

37 slit that was angled less than 2 degrees with respect to the pixel matrix were acquired. The slit was oriented almost perpendicular to the scanning direction. Four rows of data, each five pixels in width, were used in th e determination of the finely sampled LSF. The data was binned using the simple geometrical relationships according to their relative positions. The Fourier transform of this LSF results in the presampled MTF. Detective Quantum Efficiency The majority of the researchers discu ssed above that measured an MTF also measured an NPS in order to calculate a DQE. The basic definition of the DQE is consistent throughout the literature.2,4,5,10,23 In order to calculate the DQE the ideal SNR must be determined. This is the SNR of th e x-ray beam incident upon the image receptor. The ideal SNR is defined as the incident number of x-ray quanta.6,10,23 To calculate this, the x-ray spectrum of the imaging syst em must be determined. Bradford et al. 17 eliminated the need for this calculation by using the same image acquisition conditions as Dobbins et al .4 and utilizing the alrea dy published values. Launders et al.23 used computer simulations based on published me thods and tables. Flynn and Samei2 used a semiempirical x-ray spectra model based on work by Storm.45 Dobbins et al.4 used the ideal SNR value provided by the ma nufacturer. Fetterly and Hangiandreou5 estimated the x-ray spectra used experimentally with the tungsten anode spectral model using interpolating polynomials (TASMIP) pr ogram developed by Boone and Seibert.46 Fetterly and Hangiandreou5 investigated the effects of different x-ray spectra on the DQE of a CR system. In order to obtain di fferent quality x-ray spectra the beam was filtered with various thicknesses of a patient equivalent phantom (PEP), aluminum and copper. The PEP was constructed from 1 mm of type 1100 aluminum secured between two 2.48-cm thick pieces of acrylic. Depending on the peak tube potential up to six PEP

PAGE 53

38 phantoms were used. In addition, up to 80 mm of aluminum and 8.1 mm of copper were utilized. X-ray beams between 70 and 120 kVp were investigated. The term that corrects for the system gain in the cal culation of the DQE was elim inated by converting the image pixel values to exposure values before the calculation of the NPS. This methodology was also followed by Flynn and Samei.2 Contrast-Detail Analysis Lu et al.29 utilized a custom made contrast detail phantom to compare computed radiography (Kodak CR 400) and film-screen combination (Speed 400) systems in regards to patient dose, technique settings a nd contrast-detail detect ability. The phantom was constructed from a 26.5-cm 26.5-cm Lu cite sheet that was 2.5-cm thick. Two hundred and twenty-five holes (a rranged in a 15 15 square pa ttern) were drilled in the Lucite ranging from 0.3 to 8 mm in both diamet er and depth. The phantom was imaged in an under-table bucky with the use of a 10: 1 grid. The phantom was placed on top of additional Lucite sheets to simulate tissues th at generate scattered radiation. The Lucite and the phantom were placed on the examina tion table. Varying amounts of scattering media were used and the source to image di stance (SID) was adjusted so the geometric magnification of the phantom in the image rema ined constant for different thicknesses of scattering media. The CR images were printed onto film for a hard copy reading comparison with the images acquired directly on film. In order to compare the use of CR with a higher peak-tube potential x-ray beam, the soft copy CR images were read and the data manipulation tools of the diagnostic workstation were employed. Four physicists evaluated the images by scoring the threshold target depth for each object size. An object detection ratio was calculated by dividing th e number of detected objects by the total number of objects in the phantom. Lu et al.29 concluded that usi ng a higher peak-tube

PAGE 54

39 potential setting and additional aluminum filtration would reduce the patient entrance skin dose without compromising the contrast-d etail detectability, which was compensated by the contrast manipulation on soft-copy workstations. Padgett and Kotre6 utilized the Leeds TO20 test obj ect to evaluate the effect of dead pixels on image quality in direct digi tal radiography based on selenium detectors. The Leeds TO20 test object creates contrast by using x-ray attenuating objects so the objects in the image are light compared to th e background. Images of the contrast-detail phantom were obtained from a Hologic Direct Ray EPEX system with the anti-scatter grid in place. No additional scattering media was placed between the x-ray source and the image receptor. Four experienced observers sc ored the images and the threshold contrast (CT) was determined for each object size. The th reshold contrast was the contrast of the object at the threshold of visual detectabi lity. The degradation in observer performance was found to be similar to the reduction in the relative DQE (DQE normalized to zero frequency) with the loss of active pixels. The loss of active pixels was simulated in the images prior to scoring. Aufrichtig47 and Aufrichtig and Xue28 investigated the low contrast threshold detectability for an amorphous silicon x-ray detector designed for digital radiography and a standard thoracic screen-film combination. Th e contrast-detail phantom utilized in this study is the commercially avai lable CDRAD contrast-detai l phantom manufactured by Nuclear Associates, Carle Place, NY. The phant om is constructed from a 26.5-cm 26.5cm Plexiglass plate that is 1-cm thick. The Plexiglass plate is divi ded into 255 separate square regions (15 rows 15 columns) of equal size. There are holes drilled in the plexiglass creating objects that vary logarithmically in diameter and depth from 0.3 to 8.0

PAGE 55

40 mm. The first three rows of square regions, containing objects 5.0 mm and larger, have one object in the center of each square. The re maining rows have two identical objects in each square, one in the center and the other in one of the four corners. The design of these rows allows for the use of a four-altern ative forced choice (4-AFC) experimental methodology. The observer is forced to choose in which corner th e object is located. This type of study design can redu ce the subjectivity of the obser vers detection thresholds. Twelve images of the CDRAD phantom we re sequentially acquired for each image receptor. The images were acquired to simula te a clinical chest exam; the image receptor was in an upright bucky, an an ti-scatter grid was utilized and a 12.7-cm acrylic absorber was placed between the CDRAD phantom and the x-ray source. The images acquired with the flat-panel detector were printed to film with a laser imager for hard-copy evaluation. In order to utilize the 4-AFC met hodology, the first three rows of the phantom were not evaluated by the observers. Six obser vers with normal or corrected vision were used. A signal detection mode l that is beyond the scope of this discussion was used to calculate the threshold contrast level wher e a 75 percent probabil ity of correct object identification was reached. Rong et al.48 compared the low-contrast perf ormance of an amorphous siliconcesium iodide based flat-panel digital chest radiography system to those of a screen-film and a CR system by measuring their contrast -detail curves. The CDRAD contrast-detail phantom was also used in this study. Images were acquired with the CDRAD phantom in the center of the x-ray field against the ches t detector-grid assembly and a 0.5-mm thick copper plate was placed at the tube output to simulate patient attenuation. All images were processed and printed according to their respective clinical protocols. The images

PAGE 56

41 were evaluated by three physicists, a graduate student, and an undergraduate student. Rong et al. state that, previous stud ies have shown that for images of a contrast-detail phantom there were no significant differences in observer responses between radiologists and nonradiologists and that there was no noticeable improvement in the readers' performance with increased experience.(48, pp.2330-2331) The minimum detectable object contrast was then determined for each object size. A procedure to correct for the false identification of an object in one of the four corners on a subregion, supplied by the manufacturer, was applied to the raw scoring data before development of the contrastdetail curves. Rong et al. concluded that the flat-panel system demonstrated a significantly better low-contrast performance than the screen-film or CR systems as compared by the contrast-detail curves. MOSFET Dosimetry Bower and Hintenlang42 evaluated the characteristics of a patient dose verification system that uses high sensitivity MOSFET dosimeters that were developed for low dose measurements. The dosimeters were evaluated at clinical diagnostic energy levels and were found to perform well at these energies. The system was found to have a sensitivity that was precise enough for many physics applic ations. It was also determined that the size of the dosimeters would not interfer e with diagnostic image quality. Bower and Hintenlang concluded that, The dosimeters have a good angular response, a very linear response with dose and are extremely small. These characteristics make them good candidates for use in a variety of dosimetry applications.(42, p.204)

PAGE 57

42 CHAPTER 4 EXPERIMENTAL EQUIPMENT Imaging System All images utilized in this research we re acquired in the Radiology Departments Examination Room 1 in Shands Teaching Hospital at the University of Florida. The x-ray source is a Dunlee Duratron x-ray tube with a Picker VPE 3-phase generator. The x-ray tube insert has a target angle of 13.5 degr ees with a 0.6-mm and a 1.2-mm focus. The inherent filtration of the tube -collimator assembly is 2.35-mm aluminum equivalent. The system has a half-value layer of 2.73 mm of aluminum at 60 kVp. The table assembly was manufactured by Picker and contains a re ciprocating anti-scatter grid. The grid is 36.2 cm 48.0 cm with a grid ratio of 12:1 and 103 lines per inch. The grid has a focal distance of 36 to 40 inches. The CR system is the Agfa** ADC system with a Compact plate reader and the MD30 code15 imaging plates. The 24-cm 30cm high-resolution imaging plates were used. The plate-reader combination generate s 8.8 pixels per mm in the final image. Software There were various commercially available computer packages used in the completion of this work. Matlab, a high-le vel technical compu ting language produced Dunlee, 555 North Commerce Street, Aurora, IL 60504 Philips Medical Systems, 22100 Bothell Everett Highway, P.O. Box 3003, Bothell, WA 98041-3003 ** Agfa Corporation, 100 Challenger Road, Ridgefield Park, NJ 07660

PAGE 58

43 by The Mathworks, Inc., was used in the calculation of the image quality metrics investigated in this research. TA SMIP, developed by Boone and Seibert46, was used to generate the x-ray spectra needed in the ev aluation of the image quality metrics. The General Monte Carlo N-Particle Transport Code Version 4C (MCNP4C), produced by the Diagnostics Applications Group at Los Alamos National Laboratory, was used to verify input parameters to TASMIP and simu late additional x-ray spectra for different experimental setups. All statistical anal ysis was performed using SPSS for Windows, Release 11.5.0, which is produced by SPSS Inc. The images that were retrieved from th e PACS at Shands hospital are in DICOM format. Matlab was not able to read the images directly from the PACS. The program ImageJ, developed by Wayne Rasband at the Na tional Institutes of Health was used to make the images readable by Matlab. This was accomplished by saving the DICOM images as text images. NPS C MTF C and DQE C Acrylic was used to simulate patient scatte r in the acquisition of the images used to calculate the NPSC while maintaining a flat field. Tw o blocks of acrylic, each 40 cm 32 cm 2.5 cm were utilized. In addition to th e acrylic, a 254-m thick tungsten square (10 10 cm) was used as the edge device in the measurement of the MTFC. This device was manufactured by Electronic Space Products International.*** The edges of the tungsten square were manufactured to be straight with a precision of 50 m. In order to reproduce The Mathworks, Inc., 3 Apple Hill Drive, Natick, MA 01760 Los Alamos National Laboratory, P.O. Box 1663, Los Alamos, NM 87545 SPSS Inc., 233 South Wacker Drive, 11th Floor, Chicago, IL 60606 *** Electronic Space Products International, 1050 Benson Way, Ashland, OR 97520

PAGE 59

44 the angle of the edge, a device was custom ma de in-house to position the tungsten square. Exposure measurements were made with the Radcal MDH (model 1015) ion chamber. The standard probe was used for table top m easurements and the pancake probe was used for measurements in the bucky. The genera tor ripple was measured with a Kiethley 35080A kVp voltage divider, with the 50-kV 150-kV pack, and a Tektronix THS720A 100 MHz oscilloscope. Clinical Contrast Detail Score In addition to the acrylic used to simula te patient scatter, two contrast-detail phantoms were used in this study. The fi rst is the UF Radiology phantom shown in Figure 2-6. This phantom was constructed from a 20 cm 15 cm 2.5 cm piece of Lucite.27 The phantom contains ni ne object sizes ranging from 12.7 mm to 0.8 mm in diameter. For each object size, seven contrast steps are created from drilled holes ranging from 10.0 mm to 0.25 mm in depth. The object s are arranged in a square pattern. As pictured, decreasing object size is arranged in columns and decreasing contrast steps are arranged in rows. The second is the Leeds te st object TO.10 shown in Figure 2-7. This phantom contains twelve object sizes with each size having nine contrast steps. The six largest object sizes are arrange d in a circular pattern around the periphery of the phantom while the six smallest object sizes are arrang ed in rows in the phantoms center. The objects range in diameter from 11.1 mm to 0. 25 mm. The contrast steps for each object size and the relative subject cont rast between the different object sizes are not the same as Radcal Corporation, 426 West Duarte Road, Monrovia, CA 91016 Cardinal Health Radiation Management Services 6045 Cochran Road Cl eveland, OH 44139-3303 Tektronix, Inc, 14200 SW Karl Braun Drive, P. O. Box 500, Beaverton, OR 97077

PAGE 60

45 in the UF Radiology phantom. The nominal cont rast values, as liste d in the instruction manual, range from 0.012 to 0.93 but are not specified for a specific energy.49 Dosimetry In order to measure the effective dose fo r different radiographic examinations, an anthropomorphic phantom of a one-year old, produced by Bower34, and the MOSFET AutoSense Dose Verification System, manufactured by Thomson and Nielsen Electronics Ltd, were utilized. The anthr opomorphic phantom, previously mentioned in Chapter 2, is shown in Figure 4-1. The arms are contained in the el liptical trunk while the head and legs are not permanently attached. The trunk of the phantom is 21.5 cm in length and 13.2 cm in width. Figure 4-2 and Figure 4-3 show the top and bottom of the phantoms trunk respectively. Figure 4-4 show s the bottom of the phantoms head. The small holes (3 mm in diameter) provide access to the body organs of interest for insertion of the MOSFET dosimeters. The large hole in th e head is for the spine, which protrudes from the top of the trunk. The high sens itivity MOSFET dosimeters, TN-1002RDI, and the dual bias supply, TN-RD-22, were utilized in this study. The dual bias supply allows for either high sensitivity or standard bias in one piece of equipment. Figure 4-1. One-year old Bower st ylized anthropomorphic phantom.

PAGE 61

46 Figure 4-2. View of the top of the phan toms trunk showing the MOSFET access ports and the spine (top-center). Figure 4-3. View of the bottom of the phantoms trunk.

PAGE 62

47 Figure 4-4. View of the bottom of the phantom s head. The large hole on the left is for insertion of the spine.

PAGE 63

48 CHAPTER 5 EXPERIMENTAL METHODOLOGY The majority of the research objectives stat ed in Chapter 1 deal with comparing the NPSC, MTFC, DQEC and CDSC from images acquired and/ or presented under varying conditions. In order to make this comparison quantitative, a single metric was defined for each image quality parameter. Since the NPSC, MTFC and DQEC vary over a range of spatial frequencies, the integral under these cu rves, up to the Nyquist frequency, was used for comparisons. These integral values were labeled INPSC, IMTFC and IDQEC respectively, where the I sta nds for integral in each case. This methodology has been utilized by previous researchers.50 The total number of visible objects in an image of the contrast detail phantom (either perceived by a human observer or inferred from CNRT values calculated from the obser ver studies) was used for the CDSC. Image Acquisition Experimental Setup All images were acquired with a standard experimental setup that mimics clinical practice. The clinical exam ination setup for a chest/abdomen/pelvis (CAP) exam of a pediatric patient of similar size to the an thropomorphic phantom was chosen. This setup can be seen in Figure 5-1. The edge test device and the TO.10 phantom were imaged on the exam table between two 2.5-cm blocks of acrylic. Flat-field images were acquired with only the two blocks of acrylic in the x-ray field. Since the UF Radiology phantom was constructed from a 2.5-cm block of Luc ite it was imaged on the exam table, the drilled holes facing up, with only one 2.5-cm block of acryl ic on top of the phantom. The

PAGE 64

49 test devices were placed betw een the two blocks of acrylic to simulate the evaluation of the image quality metrics at a location in side the patient. The two contrast-detail phantoms were oriented so that the portion of each phantom with the smallest objects and the lowest subject contrast we re in the portion of the image with the highest exposure. The imaging cassette was placed in the bucky and the SID was set at 40 inches. The automatic collimation was used in all cases except when imaging the anthropomorphic phantom. In those cases the x-ray field was tightly collimated on the phantoms trunk. Finally, the small focal spot (0.6 mm) was used for all images. Acquisition Parameters The medical physics staff at Shands Hosp ital has conducted extensive studies to optimize the technique settings for various exams with the Agfa CR imaging system. The x-ray generator technique se ttings (peak-tube pot ential and current-t ime product) that would normally be used for a CAP of the anthropomorphic phantom are 60 kVp and 1.0 mAs. The image quality metrics were evaluated with images acquired around this exposure level. In addition to technique se ttings, there are many ways the raw image data can be collected and processed before it is viewed on the QA station. These settings are selected by the user before the image plate is scanned. The two most important parameters for this research were the exposure class and the proc essing options associated with a specific type of exam. The exposure class sets the ga in of the photomultiplier tube in the plate reader. This setting is chosen based on the exposure to the imaging plate. The CR imaging system at Shands is calibrated so th at an image acquired in the bucky with the automatic exposure control (AEC) is sca nned with exposure class 400. The exposure class is analogous to the traditional speed of a screen-film imaging system. If the

PAGE 65

50 exposure to the plate is doubled, the exposure cl ass needs to be reduced by half. This will produce the desired optical de nsity, or brightness, in the final image. The average brightness level over the entire image is s hown at the QA station and is called the exposure level. The desired exposure level is 2.2. There are three levels of processing opti ons to describe a specific type of radiographic procedure. Unless otherwise stated, the following processing options were used for all images in this work. The primary and secondary processing option used was system diagnosis. These options allowed for the minimum amount of processing available with this imaging system. The tertia ry processing options used were either flatfield or full range. The flat-field option wa s used for the determination of the NPSC and the full range option was used for all other images. In order to minimize any additional variations in the acquisition pa rameters being evaluated, a single imaging plate was used for the acquisition of all images that were directly compared (unless otherwise noted). Image Retrieval After the images were acquired, the consis tency of the radiation output from the xray tube, and therefore the leve l of exposure to the imaging plate, was verified for each image acquired by recording the exposure leve l indicator at the QA station. The images were then sent to the PACS. Most of the imag es, except those utilized in observer studies, were retrieved from the PACS via CD. These im ages were then transferred to a personal computer (PC) for processing. Clinical Noise Power Spectrum Variation with Peak-Tube Pote ntial and Current-Time Product In order to determine the NPSC, flat-field images with 5 cm of added acrylic (see Figure 5-2) were acquired and mathematically evaluated the same way as a traditional

PAGE 66

51 NPS calculation. This calculation was detaile d in Chapter 2 and a Matlab algorithm, or m-file, was developed to determine the INPSC. This m-file and an explanation of its functionality can be seen in Appendix A. To compare the variation of the INPSC with current-time product and peak-tube potential, images were acquired for each of the technique settings listed in Tables 5-1 and Ta ble 5-2. The last column in each table gives the exposure class used to scan the imagi ng plate. This was based on an AEC exposure with the CPS setup. For the peak-tube potenti al variation study, the current-time product was adjusted to maintain as constant an e xposure level as possible to the imaging plate. The exposure level was measured in th e bucky with the MDH ion chamber. The measured exposure level for each set of tec hnique factors is listed in Table 5-2. Variation with Processing Option In order to quantify the effect of other processing options, flat-field images were acquired at 60 kVp and 1.0 mAs as previously described but two different exam types were selected to process the raw image data The processing options compared are listed in Table 5-3. Clinical Modulation Transfer Function Edge Angle Verification The edge alignment device was constructe d to produce an edge angle of three degrees with respect to the pixel matrix (s ee Figure 5-3). The fo llowing setup procedure was tested to ensure reproducibility of the edge angle. The first block of acrylic was placed on the exam table so the long edge wa s aligned with the line down the center of the table. The alignment device, fitting snugl y over the edge of the acrylic, was put in place. The hinged arm (the right arm in Figure 5-3) was set to the correct angle by a small peg. The edge device was placed on the acryli c and aligned with the right arm. The right

PAGE 67

52 edge of the tungsten square was secured to the acrylic with tape. The alignment device was removed and the second block of acrylic was placed on top of the first. This procedure ensured the left and upper edges of the tungsten square (as shown in Figure 53) were near the center of the image so one image could be used to calculate the MTFC in both the scan and subscan di rections (see Figure 5-4). The imaging plate does not fit tightly in the imaging cassette (see Figure 5-5), so even if the alignment device was exactly repr oducible, slight vari ations in the plate position could change the edge angle with re spect to the pixel matrix. Two series of images were acquired to investigate the reprodu cibility of the edge device setup. In the first series, the setup was repeated for each im age. In the second series, the edge device was set up once and multiple images were taken. These images were mathematically evaluated the same way as the traditional MTF calculation detailed in Chapter 2. A Matlab m-file was developed to determine the IMTFC. This m-file and an explanation of its functionality can be seen in Appendix B. Variation with Peak-Tube Potential, Cu rrent-Time Product and Processing Option To compare the variation of the IMTFC, images of the edge device were acquired with the same technique settings and pro cessing options as listed above for the INPSC. The images were then evaluated with the IMTFC Matlab m-file. Clinical Detective Quantum Efficiency Once the NPSC and MTFC had been determined for a given set of acquisition parameters, the DQEC could be determined. Before this calculation could take place, the parameters k and Q in Equation 2-13 were determined. This was done through the simulation of the x-ray spectrum at the imag ing plate. The x-ray spectrum output from the x-ray tube housing was simulated using TASMIP and MCNP4C. This procedure will

PAGE 68

53 be explained later in this chapter. The details of this procedure are not necessary for this discussion and it is assumed the output spec trum is known. A simulation was performed in MCNP4C based on the physical setup of the NPSC measurement. The MCNP4C input file for the simulation of the x-ray spectrum in the bucky at the imaging plate is shown in Appendix D. This simulation was performed twice. The first was with a thin sheet of lead simulating the thickness of the grid in the xray field. The second was without the lead in the x-ray field. A weighted average was perf ormed on the two resulting spectra to get the simulated spectrum at the imaging plate. Th e weights were based on the proportion of the x-ray field obstructed by the thickness (t) of the lead strips in the grid to the gaps (g) between the strips (see Figure 5-6). The gr id ratio is defined as h divided by g. After the x-ray spectrum incident on the imaging plate was determined the value of Q (photonsmm-2mR-1) was calculated. The x-ray spectrum from MCNP4C was normalized so that when multiplied by a constant would give a true spectrum of photons/mm2. Before the exposure in mR could be calculated from equation 2-14, the energy dependant mass-energy absorption coefficients (en/ ) for air were calculated from published values (see Table 5-4).24 The simulated spectrum was calculated at one keV intervals, but en/ values are only published at a few energies between 10 and 80 keV. The intermediate values were calculat ed by fitting several curves to the known data. The energy range between 0 keV and 80 ke V was subdivided into five regions; 010, 10-20, 20-40 40-60 and 50-80. The en/ values for the 10-20 keV range were fit with a power series of the form 1738 3 610 2 ) ( E Een (5-1). The en/ values for the 20-40 keV range were fit with a power series of the form

PAGE 69

54 9433 2 610 5 ) ( E Een (5-2). The en/ values for the 40-60 keV range were fit with a polynomial of the form 3326 0 865 9 5 80 ) (2 E E Een (5-3). The en/ values for the 50-80 keV range were fit with a polynomial of the form 1611 0 576 3 333 23 ) (2 E E Een (5-4). This last energy range overlapped with the previous one so that three en/ values were available for the curve fitting but E quation 5-4 was only used to calculate the en/ values above 60 keV. The R2 values describing how well the above equations fit the data are 1.0, 0.9992, 1.0 and 1.0 respectively, w ith 1.0 being the best possible fit. The en/ values for the 0-10 keV range were determin ed from linear back-i nterpolation using the known en/ value for 10 keV and the calculated value for 9 keV. Since two equations were fit to each known data point except at 10 keV and 60 keV, the published values were used at those energies. The multiplicative constant mentioned previously was changed until the sum of the exposure (in mR) for each energy bin in the spectrum was one. The exposure for each energy was calculated using E quation 2-14. The photon spectrum was then summed over all energies to give a to tal number of photons per mm2 to produce one mR at the imaging plate. This value was then used in the calculation of the DQEC based on the measured exposure level in the bucky for a gi ven set of technique factors. The parameter k, representing the gain of the imaging system for each set of acquisition parameters, was determined from the average pixel values from the center

PAGE 70

55 portion of the flat-field images used in the determination of the INPSC. A ROI about the size of the MDH pancake probe used to meas ure the in bucky exposure was extracted and averaged from each flat-field image. These values were then averaged for each set of acquisition parameters. These average pixel va lues were then normalized to one set of acquisition parameters. This gave a relative system gain which included not only the change in technique factors, but also the exposure class. Clinical Contrast Detail Score Observer Study Images were acquired of the two contrast detail phantoms with 5 cm of total acrylic scattering media. The images were acquired at 60 kVp with the acquisition parameters listed in Table 5-5 and Table 5-6. The expos ure classes were base d on an AEC exposure of each phantom setup and adjusted accordin gly for the change in current-time product. As previously mentioned, the system-diagnos is/full-range processi ng options were used. Two sets of these images were sent to the PACS. The first set was the normal images and the second was a set of those same images with the grayscales inverted. All of the images were then available for scoring at a diagnosti c workstation in the pe diatric reading room at Shands. Thirteen human observers were utilized to score the images of the contrast-detail phantoms. There were four thirdor fourth -year radiology residents, four experienced medical physicists and five medical physics grad uate students. All images were evaluated on the same diagnostic viewing workstation. All of the images of one phantom were evaluated in a single viewing session. Theref ore, each observer had two viewing sessions during this study. The images were presen ted to each observer in the same randomly selected order based on the current-time produc t acquisition value. For each current-time

PAGE 71

56 product value four images were read in a sp ecific order. The normal image was scored first without the use of any of the image processing tools av ailable at the workstation. Next, the grayscale inverted image was scored without the use of any processing tools. The normal image was then scored again with the use of the window and level controls. Finally, the grayscale inverted image was scored with the use of the window and level controls. Each observer was allowed to get familiar with the func tioning of the window and level controls at the workstation with a test image of the phantom before the scoring began. Each observer was also given the follo wing instructions before scoring began. 1. Identify the number of objects you can see of each object size. An object is considered visible if it is an entire circle with a clear circular outline. 2. Score each row with your first impression. Do not spend more than a few seconds on each row. 3. Operate the window and level controls as you see fit to best score the image. 4. Do not score past an object that is not visible. Once all of the images were scored, one of th e images the viewer had already scored was presented again as a check of observer reproducibility. CNR Calculations The observer study produced an average threshold object number for each object size for the four image scoring conditions. In order to calculate the CNR associated with this threshold of visibility, a Matlab m-file was developed for each phantom to calculate the CNR of each object. This m-file and an expl anation of its functiona lity can be seen in Appendix C. Since both m-files function esse ntially the same and differ only by input parameters, only the TO.10 m-file is presented.

PAGE 72

57 Variation with Processing Option The m-file in Appendix C can also automati cally score images of the contrast-detail phantoms used in this study. This algorith ms automatic scoring ability is based on the combination of the CNR calculations and th e observer data. The reasoning behind this procedure will be presented in the next chapter. This ability allowed the scoring of images acquired with the acquisition paramete rs listed in Table 5-5 and Table 5-6 for each of the processing options listed in row tw o and three of Table 5-3 without the use of human observers. Anthropomorphic Phantom Viewer Study In order to evaluate the image acquisiti on parameters needed to produce an image of minimum acceptable quality, a series of im ages were acquired of the anthropomorphic phantom. The phantom was placed on the ex am table and the imaging cassette was placed in the bucky. No additional scattering material was placed in the x-ray field. The images were acquired with the acquisition pa rameters listed in Table 5-6 (based on an AEC exposure of the phantom) for both of th e processing options lis ted in rows one and two of Table 5-3. The images were then sent to the PACS for retrieval at the same diagnostic viewing station used in the cont rast-detail observer study. Since Shands has already undergone a procedure to optimize their technique fact ors, the technique factors recommended by the technologist to image th e anthropomorphic phantom were initially assumed to produce an image of minimal accep table quality (60 kVp and 1.0 mAs). In order to verify that assumption the radiology re sidents were asked to rate all other images to this standard image. In a process similar to Rill,51 the images were presented one at a time against the standard image and the radiol ogist was asked to score the new image on a five-point scale (see Table 5-7). An image w ith a lower exposure level than the standard

PAGE 73

58 image with an average rating of three or hi gher will replace the standard image as the image of minimum acceptable quality. Dosimetry Four MOSFET dosimeters were used in the measurement of absorbed dose to the individual organs of the anthropomorphic pha ntom. This was done because only four of the dosimeters had a total accumulated exposur e resulting in less than 10,000 mV. Under this level of exposure, the dosimeters do not n eed to be recalibrated. Three measurements were made at each of the relevant organ sites needed to calculate the effective dose. Each measurement was performed with the same phys ical setup used in the anthropomorphic phantom viewer study with techniqu e settings of 60 kVp and 300 mAs. In order to perform the calculation of effective dose using Equation 2-20, the average mass energy absorption coefficients we re needed. This involved determining the average energy of the x-ray beam. An accura te method for determining diagnostic x-ray spectra is through the use of TASMIP. Th e TASMIP algorithm requires three input parameters; peak tube potential, generato r ripple and aluminum -equivalent-inherent filtration. The peak tube potential, the effective kV and the generator ripple were physically measured with an oscilloscope The inherent filtration provided by the manufacturer was verified through the measur ement of the x-ray units aluminum halfvalue layer (HVL) and the use of MCNP4C. Th e MCNP4C input file that performs this simulation is shown in Appendix D. The half-value layer was m easured with the use of the MDH ion chamber, a test stand designed specifically for the HVL measurement and varying thicknesses of aluminum. An initial exposure measurement wa s performed without any aluminum in the x-ray field. Varying thicknesse s of aluminum were added and an exposure measurement

PAGE 74

59 was taken for each aluminum thickness until th e exposure level was less than one-half the value of the initial exposure measurement. This data was then used to determine the thickness of aluminum that would be needed to reduce the exposure to one half of the initial exposure measurement. This procedur e was repeated for 50, 60, 70 and 80 kVp. Starting with the manufacturers stated inhe rent filtration as an input to TASMIP, a spectrum was generated and used in a simula tion of the x-ray system. The simulation was performed twice. The first was with the meas ured amount of aluminum, determined from the half-value layer measurement, in the x-ray field. The second was without the aluminum in the x-ray field. The exposure to a simulated detector was calculated for each simulation. When the first simulation pr oduce a reduction in the exposure to the simulated detector of one-half the value of th e second simulation, the inherent filtration parameter used as the input to TASMIP was considered correct. This did not occur with the first x-ray spectrum so the inherent filtration valu e was changed until the proper exposure reduction was seen. The mean energy was calculated from this spectrum. The effective dose was then calcu lated from the measured exposure level and linearly scaled with current-time product to get the effective doses at the dia gnostic exposure levels used in the evaluation of the image quality metrics.

PAGE 75

60 Figure 5-1. Image acquisition setup. Figure 5-2. Image acquisition se tup for the flat-field images. The two 2.5-cm blocks of acrylic are placed on the exam table and th e x-ray tube is set at an SID of 40 inches.

PAGE 76

61 Figure 5-3. Device to ensure pr oper angulation of edge device. Figure 5-4. Final setup of the edge device for the determination of the MTFC.

PAGE 77

62 Figure 5-5. The Agfa imaging plate inside the open imaging cassette. Figure 5-6. Idealized graphical representation of the grid. Table 5-1. Technique factors for th e current-time produc t variation study. kVp mAs Exposure Class 60 0.4 800 60 0.8 800 60 1.0 800 60 1.2 800 60 2.0 800 60 3.2 400 t g h

PAGE 78

63 Table 5-2. Technique factors for the peak-tube potential variation study. kVp mAs Exposure Class 50 12 200 60 5.6 200 70 3.0 200 80 2.1 200 Table 5-3. Processing options compared. Processing Option Level Primary Secondary Tertiary Exposure Class System Diagnosis System Diagnosis Flat Field 800 Pediatric Chest Chest PA 800 Pediatric Upper Extremity Hand AP 800 Table 5-4. Published mass energy-a bsorption coefficients for air. Energy (keV) (en/ )E,air (1/cm) 10 4.61 15 1.27 20 0.511 30 0.148 40 0.0668 50 0.0406 60 0.0305 80 0.0243 Table 5-5. Technique factors for the TO.10 phantom. kVp mAs Exposure Class 60 0.4 800 60 0.8 800 60 1.0 800 60 1.2 800 60 2.0 800 60 3.2 400

PAGE 79

64 Table 5-6. Technique factors for the UF Radiology Phantom. kVp mAs Exposure Class 60 0.4 800 60 0.8 800 60 1.0 800 60 1.2 800 60 2.0 400 60 3.2 200 Table 5-7. Five-point scale for s ubjective image quality evaluation. Score Image Quality Compared to Standard Image 1 Much Worse 2 Worse 3 Same 4 Better 5 Much Better

PAGE 80

65 CHAPTER 6 RESULTS AND DISCUSSION NPS C The full two-dimensional NPSC was calculated directly from the flat-field images in this study. The pixel values in the raw im age data were used directly and were not converted to exposure values before calculation of the NPSC. Unlike the MTFC, which must be calculated in the laser scan a nd subscan directions separately, the twodimensional NPSC can be calculated from a single image and the one-dimensional NPSC in the scan and subscan dire ctions is then calculated from the two-dimensional NPSC as previously described. The INPSC is the integral under the full two-dimensional NPSC. Even though the full two-dimensional NPSC has the positive, as well as the redundant negative frequency valu es present, the INPSC was defined in this way for ease of evaluation. Variation with Current-Time Product The variation of the INPSC with current-time product is shown in Figure 6-1. Five images at each current-time product we re used to calcula te an average INPSC. The INPSC decreased as the current-time product was in creased. A one-way analysis of variance (ANOVA) was performed to compare the means and check for statistically significant differences. The ANOVA p-value was 0.000 so differences exist among the means. Additionally, the Levene statistic was used to test for homogeneity of the variances. The p-value was 0.000 so the assumption of equa l variances was not used in the post-hoc tests. Three multiple comparison post-hoc test s that do not assume equal variances were

PAGE 81

66 performed. The results of the post-hoc te sts are shown in Appendix E (Table E-1). Tamhanes T2 is a conservative pairwise comp arisons test based on a t test. Dunnetts T3 is a pairwise comparisons test based on the Studentized maximum modulus. Lastly, Games-Howell is a more liberal comparison test. The INPSC decreases with increasing current-time product. This was the expected result since the NPSC is a measure of the noise level in an image and as the current-time product is increased more photons go into the formation of the image and the quantum noise level is reduced. The differences in the INPSC were statistically significant for all current-time product comparisons. The INPSCs determined at 0.4 mAs had the largest variance. The average values and thei r associated standard deviations ( ) are shown in Table 6-1. As the current-time product is increased, the magnitude of change in the INPSC for equal increases in current-time product does not remain constant. For example, the difference in the average INPSC for a doubling of the current-time product as determined from 0.4 mAs and 0.8 mAs is 2.222-6 while the difference is 9.497-7 between 1.0 mAs and 2.0 mAs. This shows that the INPSC is more sensitive to changes in exposure level at lower exposure levels. This was the ex pected result since th e CR imaging plate is a non-linear photon detection system. Variation with Peak-Tube Potential The variation of the INPSC with peak-tube potential is shown in Figure 6-2. Five images at each peak-tube potential we re used to calculate an average INPSC. The currenttime product was adjusted in the acquisition of the flat-field images used to test the INPSC variation with peak-tube potential in the attempt to maintain a constant exposure level of 1.0 mR (3.876 C/kg). The mA and time were adjusted independently at each

PAGE 82

67 peak-tube potential to mainta in the exposure level as clos e to 1.0 mR (3.876 C/kg) as possible. The INPSC decreased slightly as the peak-tube potential was increased from 50 kVp to 60 kVp, then remained statistically constant up to 80 kVp. The ANOVA p-value was 0.000 so significant differences exist among the means (in this case the INPSCs determined at 50 kVp and 60 kVp statistically differed). The Levene p-value was 0.000 so the assumption of equal variances was not us ed in the post-hoc tests. The results of the post-hoc tests are shown in Table E-2. Variation with Processing Option The variation of the INPSC with the three different processing options previously discussed is shown in Figure 6-3. Ten im ages processed with each processing option were used to calculate an average INPSC. Ten images were used instead of five because of the increased variance in the INPSC when the hand AP processing algorithm was used. The INPSC decreased with the chest PA processi ng option and increased with the hand AP processing option. The ANOVA p-value wa s 0.000 so statistical differences exist among the means. The Levene statistic p-va lue was 0.001 so the assumption of equal variances was not used in the post-hoc tests. The results of the posthoc tests are shown in Table E-3. The slight decrease in the INPSC when the chest PA processing option was used indicated that the algorithm performs a smoothing operation on the raw image data. The significant increase in the INPSC when the hand AP processing option was used indicated that an edge enhan cement process was performed. MTF C Two methods of determining the MTFC were initially performed; the use of a narrow slit and the use of a sharp attenuating ed ge. Initial images of a 20-m slit at the diagnostic exposure levels inves tigated in this work did not produce a viable data set to

PAGE 83

68 calculate the LSF. With the low exposure le vels and the presence of simulated patient scattering media the slit was barely visibl e in the image. There was not enough data present in the image to generate a usable LSF and the data that was available was extremely noisy. The images of the 250-m th ick tungsten edge device were easily used to generate an ESF. Therefore, the investigations of the MTFC that follow were all accomplished with the edge method. Scan Versus Subscan Direction The Agfa Compact CR reader scans the imag ing plate along its longest dimension. Since a single image is used to calculate the MTFC in both the scan and subscan directions, these metrics are determined us ing two perpendicular edges of the tungsten square. To ensure that both edges produced the same results, several images were acquired with both edges in the scan and s ubscan direction. The images were acquired with 60 kVp and 1.0 mAs. It should be noted th at the Matlab m-file used to calculate the IMTFC was designed to evaluate the subscan MTFC using a vertical edge that is parallel to the scan direction (see Figure B1). In order to evaluate the MTFC in the scan direction the image must be rotated 90-degrees counter -clockwise so that the top edge of the tungsten square in Figure B-1 is properly oriented for evaluation. The average IMTFC from five images for both edges in the scan a nd subscan directions are listed in Table 6-4. A Students t-test was performed on the data and it was determined that both edges functioned equally well in both directions. IMTF C Reproducibility There are two main sources of variability in the IMTFC calculation. The first is due to the manual selection of th e usable edge data from the edge image. This caused a variation in the value of the IMTFC for multiple evaluations of the same image. The

PAGE 84

69 second is the difference in the IMTFC as calculated from multiple images acquired with the same acquisition parameters. Both of these sources of variability were evaluated. The IMTFC calculated five times from a single im age acquired with 60 kVp and 1.0 mAs is shown in Table 6-5. This evaluation s howed a maximum variability of 0.006 mm-1. The IMTFC calculated once from five separate images is shown in Table 6-6. This evaluation showed a maximum variability of 0.017 mm-1. Edge Angle Reproducibility The actual angle the edge device made w ith respect to the pixel matrix was calculated by the Matlab IMTFC algorithm. The target edge angle was three degrees. The actual angle of the edge in the images for bot h repeated edge setups and repeated images with a single setup are listed in Table 6-7. The average for the repeated setups was 2.8 degrees with an overall variab ility of 0.61 degrees. The average for repeated images of the same setup was 2.75 degrees with an ove rall variability of 0. 50 degrees. Combining these results, the 95-percent confidence interval for the edge angle used to determine the IMTFC is 2.79 .49 degrees. The edge angle was the same for the scan and subscan direction in each case. Number of ESF Data Points The number of data points selected about th e edge for the ESF has an effect on the magnitude of the IMTFC. This is primarily due to incr eased exposure just under the edge device from scattered x rays. This phenome non causes a shoulder on the attenuated side of the ESF that reduces the sharpness of th e edge response (see Figure 6-4). If more pixels are used for the ESF, more of the s houlder is incorporated and the magnitude of the IMTFC is decreased.

PAGE 85

70 Due to the way the FFT is calculated, the number of data points in the final MTFC is one half the number of data points in the LSF. Since the LSF is zero padded to 256 data points before the calc ulation of the MTFC, there are 128 data points in the MTFC independent of the length of the ESF. Ther efore, the frequency resolution of the MTFC appears to be independent of the ESF length. This is not the case since the length of the LSF is artificially increased to 256 data points. This process is analogous to adding additional data points to the MTFC by interpolating between th e values that would be in the MTFC based on the true length of the ESF. The IMTFC for ESFs of varying lengths are shown in Table 6-8 and Table 6-9. An im age of the edge device acquired with 60 kVp and 1.0 mAs was used to calculate the IMTFCs in Table 6-8 and Table 6-9. There is approximately a 20percent reduction in the IMTFC when 250 data points are used compared to the use of 25 data poi nts in both the scan and subscan directions. This is not a large decrease and since the shoul der on the attenuated side of the ESF is a direct result of the clinical nature of the e dge image acquisition, it was decided that some of the shoulder should be included in the clinical edge response. In addition, the smaller number of data points used does produce a smoother MTFC (see Figure 6-5), but at the cost of frequency resolution. Since the MTFC is used to compare different images, the choice of data length is somewhat arbitr ary as long as the same methodology is used for each calculation. Therefore, in order to see the effect of the ESF shoulder 128 data points were selected for furt her calculations of the IMTFC. Variation with Current-Time Product The variation of the IMTFC with current-time product is shown in Figure 6-6 and Figure 6-7. Five images at each current-time product were used to calculate an average IMTFC. The IMTFC appears to decrease with increasing curre nt-time product. The

PAGE 86

71 ANOVA p-value was 0.002 for the scan directio n and 0.000 for the subscan direction so differences appear to exist among the means. The Levene statistic p-value was 0.001 for the scan direction and 0.029 for the subscan direction so the assumption of equal variances was not used in the post-hoc tests. The results of the posthoc tests are shown in Table E-4 and Table E-5. Even though the ANOVA had a p-value of 0.002 for the scan direction, when the post-hoc tests were pe rformed and the observed significance level was adjusted for the fact that multiple comparisons were made, the IMTFC means were not statistically different. The IMTFCs for the subscan direction showed statistically significant differences between those determin ed at low current-time product values and those determined at high current-time product values. A difference was not seen between the IMTFCs determined from 0.4 to 1.0 mAs and those determined from 1.0 to 3.2 mAs. The IMTFCs determined at 0.4 mAs showed th e largest variation. The average values and their associated standard deviat ions are shown in Tabl e 6-10. This larger variation, as well as the overall larger av erage magnitude, can be attributed to the increased noise level in the edge imaged acquired at 0.4 mAs. The increased noise causes increases in the high freque ncy components of the MTFC. As the current-time product was increased and the noise level in the edge image decreased, the MTFC became more stable. This behavior can be seen in Figure 6-8 and Figure 6-9. If the MTFC is averaged over many images these variations star t to get averaged out and the MTFC starts to stabilize (see Figure 6-10). Even tho ugh these increases in the high frequency components are reduced when the MTFCs are averaged before integration, there are still slight increases in the average MTFC determined at lower cu rrent-time product values (see Figure 6-11). Since the ma ximum variation of the IMTFC is approximately 0.04

PAGE 87

72 mm-1 and the reproducibility of the IMTFC calculation is on the order of 0.02 mm-1, these small increases in the high fr equency components of the MTFC from increased noise in the edge image account for the increase in the IMTFC at low current-t ime product values. Variation with Peak-Tube Potential The variation of the IMTFC with peak-tube potential is shown in Figure 6-12 and Figure 6-13. Five images at each current-time product were used to calculate an average IMTFC. The IMTFC showed a definite increase with peak-tube potential. The ANOVA pvalue was 0.000 for both the scan and subscan directions so differences exist among the means. The Levene statistic p-value was 0. 230 for the scan direction and 0.564 for the subscan direction so the assumption of equal variances was assumed in the post-hoc tests. Three multiple comparison post-hoc tests that assume equal variances were performed. The results of the post-hoc tests are shown in Table E-6 and Table E-7. Both Tukeys honestly significant difference test (HSD) and the Bonferroni test are multiple comparisons tests that adjust the observed si gnificant level for the fact that multiple comparisons are made. Tukeys HSD is based on the Studentized range statistic while the Bonferroni test is based on St udents t statistic. The least significant difference (LSD) test is a pairwise multiple comparison test that is equivalent to multiple individual t tests between all pairs of groups. Th e LSD test does not adjust the observed significance level for multiple comparisons. All of the comparis ons showed statistical differences in the average IMTFCs at all peak-tube potentia ls in both directions. The increase in the IMTFC with peak-tube potential was a direct result of the shape of the ESF. It was previously noted that the ESFs at 60 kVp have a shoulder on the attenuating side of the edge as a result of scatter. As the peak-tube potential is increased, the scattering of photons under the edge device is decreased and the shoulder on the ESF

PAGE 88

73 is reduced (see Figure 6-14). This leads to an increase in the ma gnitude of the MTFC and as a result an increase in the IMTFC. Using the MTFC as a measure of image quality could lead to the conclusion that the higher the peak-tube potential used the better the image quality. This is not true for all applic ations because more factors than just the ability of the system to transfer subject cont rast need to be considered. If an imaging system had a perfect MTFC (a value of one for all spatia l frequencies) but such a high peak-tube potential was used that the object being imaged presented very little subject contrast, the image would still be of poor quality. So even though the IMTFC increases with increasing peak-tube potential, more f actors need to be considered before a statement can be made regarding the quality an image would have if the same acquisition parameters were used. Variation with Processing Option The variation of the IMTFC with the three different pr ocessing options is shown in Figure 6-15 and Figure 6-16. Five images pr ocessed with each processing option were used to calculate an average IMTFC. The IMTFC increased when both the chest PA and the hand AP processing algorithms were appl ied to the raw image data. The ANOVA pvalue was 0.000 for both the scan and subscan directions so differences exist among the means. The Levene statistic p-value was 0. 019 for the scan direction and 0.178 for the subscan direction so the assumption of equal variances was not used in the post-hoc tests for the scan direction but was used for the s ubscan direction. The re sults of the post-hoc tests are shown in Table E8 and Table E-9. The IMTFC means were statistically different for all processing options. The increase in the IMTFC for the chest PA and the hand AP processing options was directly attributed to the enhancement of the ESF (see Figure 617). The shoulder on the attenuated side of the edge is almost completely eliminated,

PAGE 89

74 which lead to an increase in the magnitude of the MTFC and therefore an increase in the IMTFC. This occurs because both processing op tions perform a smoothing of the bright areas (those that would represent bone in a c linical image) after the edge enhancement. This leads to not only a better edge response but a flattening of the data under the edge device which is a bright area in the image similar to a bone. The hand AP processing option was so effective with th is enhancement that the MTFC has a value greater than one for the low frequency components in the scan direction (see Figure 618). This behavior was not noticed in the subscan direction becau se the edge response was not reproduced as well in that direction, but the IMTFC is still increased with the use of the hand AP processing option. This enhancement of the MTFC does have its limitations. In the case of the hand AP processing option the noise level in the image is increased. If the noise level in the original image data is sufficiently low, this noise increase may be tolerable in a clinical setting. For the chest PA processing option, the increase in noise over the entire image caused by the edge enhancement is corrected by a smoothing operation which leads to an overall noise reduction. Dynamic Image Manipulation The calculation of the INPSC and the IMTFC in the previous two sections was performed on the raw, or digitally processed, image data before any changes were made by the viewing software or the viewer. The m onitors of the diagnostic display stations used in the Radiology Departme nt at Shands are capable of displaying 4096 gray scale levels so a 12-bit image would be accurately represented. Therefore, the raw image data used for calculations in this study should be represented by the pixel brightness values displayed on the monitor. Unfortunately, the viewing software automatically adjusts the window and level settings on the image to optim ize the display. The level is simply the

PAGE 90

75 midpoint of the pixel values displayed. The wi ndow is the range of pi xel values that are displayed around the level. This automatic ad justment changes the image data that the viewer sees so the di splayed image is no longe r the same as the raw image data. Once the image is displayed, the viewer also has th e ability to dynamically change the window and level settings. This causes a continuous change in the displayed image data. These changes have an effect on the INPSC and the IMTFC, but it is not practical to capture the displayed image data to perform calculations. In order to demonstrat e the range of this effect, the window and level of a flat-field and an edge image (acquired at 60 kVp and 1.0 mAs) were altered with the ImageJ program. The INPSC and the IMTFC were then calculated (see Table 6-11 and Figure 6-19) from these mani pulated images. ImageJ only allows the user to save an image that has ha d the window and level ad justed if it is an 8bit grayscale image. Since the images acquire d from PACS are 12-bit images, they were first converted to 8-bit images before the adjustments were made. The window and level were adjusted one at a time in both the posit ive and negative directions to the full extent allowed by the ImageJ program. After each adjustment the image was saved as a text image so it could be read by Matlab. As the window was increased to its maxi mum slope, the flat-field image became an almost uniform image with all pixel values set to 255 except a few around the periphery where the imaging plate was outside of the x-ray field during image acquisition. This reduced the INPSC to nearly zero. The reason the INPSC was not exactly zero is discussed in Appendix A. The increase in window cause d the edge image to resemble the perfect binary image discussed in Appendix B. All of the pixel values outside of the edge device were set to 255 while all of the pixel values under the edge device were set to zero. This

PAGE 91

76 caused a dramatic increase in the IMTFC. It should be noted that this is the most extreme case and in clinical practice as the window is increased, so is the perceived noise level. The noise level begins to vanish as the window approaches its maximum slope, but so does any subject contrast. Th erefore, once again more factors than just the MTFC need to be considered when assessing the quality of an image. As the window was decreased to its minimum slope, the pixels in the flat-field image were set to a narrow range of values in the middle of the grayscale range This caused a reduction in the INPSC but was not as drastic an effect as the window increase. Th e window decrease had a similar effect on the edge image except the pixels were set to two narrow ranges of values corresponding to those pixels outside of the e dge device and those under the ed ge device respectively. This did not have an effect on the IMTFC because the relative edge response remained unchanged. As the level was increased to its maximu m value, the flat-field image became a uniform image as almost all of the pixel values increased bey ond the usable 8-bit grayscale level of 255. This cau sed the same effect on the INPSC as the increase in window. This increase in leve l caused the majority of the pixels outside of the edge device to increase in value beyond the usable gr ayscale level and were therefore set to a value of 255. This flattened the tail of th e ESF on the non-attenuated side of the edge resulting in a small increase in the IMTFC. As the level was decreased to its minimum value, the pixel values in the flat-field image were a ll reduced by a cons tant but none of the pixel values decreased below zero so the INPSC was not affected. This decrease in level caused the majority of th e pixel values under the edge device to decrease below the minimum grayscale level and therefore be set to zero. This significantly flattened out the

PAGE 92

77 shoulder on the attenuated side of the ESF and caused an increase in the IMTFC. The MTFCs for these changes in window and level are shown in Figure 6-19. It should be noted that the increase in level affected th e non-attenuated side of the ESF because the raw image data from the PACS has an inverted grayscale compared to the images viewed on the display stations. Th e raw image data has high pixel values for high radiation exposure and low pixel values fo r low radiation exposure. This is exactly opposite from images acquired on film where a high radiation exposure leads to a darker area on the film. This increased exposure re sults in less light being transmitted through the film causing the perception of a low numbe r. In order to make the raw image data in CR to appear like a traditional film, the grayscale is inverted before display. For this reason, the ESFs shown in this chapter have been grayscale inverted for display purposes so they appear in the traditionally accepted way. DQE C The DQEC was calculated from the NPSC and MTFC data previously shown in this chapter. In order to calculate the DQEC, the parameters k and Q in Equation 2-13 were calculated. The value of k for this application of the DQEC was determined to be unity. Since the raw image data is corrected fo r background trends by subtracting a fitted surface from the original data be fore the calculation of the NPSC, the mean pixel value of each ROI used in the NPSC calculation is nearly zero. This procedure corrected for the total gain of the imaging system including any change in exposure cla ss. The effect of the system gain on the image noise content is not accounted for by the background trend correction. A correction for this was not desire d because the altered noi se content is what is initially displayed on the monitor as 12-b it brightness levels during image viewing. In order to calculate Q, the input parameters for TASMIP were determined. The half-value

PAGE 93

78 layer (HVL) was determined to be 0.2733 mm of Al at 60 kVp. Simulating the HVL measurement in MCNP4C the x-ray tube housing inherent filtration was determined to be 2.35-mm Al-equivalent. The genera tor ripple was determined to be approximately five percent across the peak-tube potential range of interest. Using the x-ray spectrum generated by TASMIP, the flat-field setup was simulated in MCNP4C and the parameter Q was calculated for each peak-tube potential (see Table 6-12). The parameter Q and was then normalized by the exposure measured in the bucky for each set of acquisition parameters (see Table 6-13 and Table 6-14). Variation with Current-Time Product The variation of the IDQEC with current-time product is shown in Figure 6-20 and Figure 6-21. The IDQEC showed the same trend with current-time product as the INPSC. This decrease in the average IDQEC with current-time product is due to the parameter Q in Equation 2-13. The NPSC decreases with currenttime product but the Q value increases at a greater rate causing a decrease in the IDQEC. Physically this was determined to show that as the current-tim e product is increased th e efficiency of the imaging system per unit photon decreases. The ANOVA p-value was 0.000 for both the scan and subscan directions so differences exist among the means. The Levene statistic pvalue was 0.098 for the scan direction a nd 0.152 for the subscan direction so the assumption of equal variances was assumed in the post-hoc tests for both directions. The results of the post-hoc tests are shown in Table E-10 and Ta ble E-11. All of the average IDQECs showed statistically signifi cant differences for two of the post-hoc tests in the scan direction and all of the post-hoc tests in the subscan di rection. The Bonferroni test showed statistically equivalent IDQECs determined at 0.8, 1.0 and 1.2 mAs.

PAGE 94

79 Variation with Peak-Tube Potential The variation of the IDQEC with peak-tube potential is shown in Figure 6-22 and Figure 6-23. The IDQEC showed a decreasing trend with increasing peak-tube potential. The ANOVA p-value was 0.000 for both the scan and subscan directions so differences exist among the means. The Levene statistic pvalue was 0.470 for the scan direction and 0.885 for the subscan direction so the assump tion of equal variances was assumed in the post-hoc tests for both directions. The results of the post-hoc tests are shown in Table E12 and Table E-13. The decrease in the average IDQEC was statistically significant for at least two of the post-hoc tests in the scan direction for all increases in peak-tube potential. The only statistically significant di fference in the average subscan IDQEC with a change in the peak-tube potential was for 80 kVp. The average subscan IDQEC determined from lower peak-tube potentials were not statistically different as determined by the post-hoc tests that adjust the error rate for multiple comparisons. This decrease in the IDQEC with increasing peak-tube potential was once again driven by the parameter Q but the MTFC also added to the decrease. Since the exposure level was kept constant for all of the peak-tube potentials, the NPSCs were almost the same. Both the MTFC and the Q value increased with in creasing peak-tube potential, this lead to the decrease in the IDQEC with increasing peak-tube potential. Variation with Processing Option The variation of the IDQEC with processing option is shown in Figure 6-24 and Figure 6-25. The average IDQEC increased with the chest PA processing option but did not change with the hand AP processing optio n in both the scan and subscan directions. The ANOVA p-value was 0.000 for both the scan and subscan directions so differences exist among the means. The Levene statistic pvalue was 0.026 for the scan direction and

PAGE 95

80 0.028 for the subscan direction so the assump tion of equal variances was not assumed in the post-hoc tests for both directions. The resu lts of the post-hoc tests are shown in Table E-14 and Table E-15. All three post-hoc tests show ed a statistically sign ificant increase in the average IDQEC with the use of the chest PA processing option but no statistical change occurred with the use of the hand AP processing option. The use of the integral of a function as a means of comparing two functions does have its limitations. As the NPSC and MTFC changed with different acquisition parameters (current-time product and peak-tub e potential), the direction of the change (positive or negative) was consistent for all spatial frequencies. The IDQEC did not show this same behavior with a change in the processing option. The DQEC in the subscan direction for all three pr ocessing options is shown in Figure 6-26. The IDQEC for the fullrange and the hand AP processing options are statistically the same but the DQECs have dramatically different shapes. This shows that the processing options are truly application specific and the magnitude of the DQEC they produce at any particular spatial frequency value would depend on what is being imaged. If small objects are to be imaged, the ideal DQEC would have a large magnitude at high spatia l frequencies while in contrast if large objects are to be imaged, the ideal DQEC would have a large magnitude at low spatial frequencies. This shows that the IDQEC could be a misleading i ndication of overall image quality. CDS C Observer Study Each of the contrast-detail phantoms us ed in this study was scored under four different viewing conditions: the standard imag e without the use of window and level, the standard image with the use of window and level, and the inverted image under the same

PAGE 96

81 two conditions. The total number of visible objects, or score, for each phantom under all viewing conditions is shown in Table 6-15 and Table 6-16. These are the combined results of all observers. A one-way ANOVA was performed on the total score across all current-time product values and viewing condi tions for each phantom to determine if there was a difference in the performance of the observers based on specialty. Of the three groups of observers, the radiology resident s and the students performed statistically equivalent for both phantoms. The medical physicists performe d statistically equivalent to the other groups for the UF Radiology phant om but scored statistically higher by 7.5 percent for the TO.10 phantom (the p-valu e was 0.001). This is attributed to the subspecialties within the medical physicists gr oup. Despite the same viewing instructions, the nuclear medicine physicist and the mammo graphy specialist scor ed higher than the other medical physicists. Theref ore, the total scores shown in Table 6-15 and Table 6-16 are representative of an average observer across several radiological specialties. Since the number of visible objects changes with current-time product, the ANOVA was repeated for the mean total score of each group across all viewing conditions at each current-time product independ ently. This was done to ensure that the increased variance caused by the total scor e changing with current -time product did not affect the results. The p-values for this analysis are shown in Table 6-17. The only difference found between the groups was with the TO.10 phantom at 2.0 mAs. Post-hoc tests showed that only the medical physicists and the ra diology residents scoring of the contrast-detail images statis tically differed but the conc lusion regard ing inter-group performance was verified.

PAGE 97

82 To determine if there were any differen ces in the scores due to the different viewing conditions, the same type of ANOVA was performed for the different viewing conditions as was done for the observer groups. The p-values for the mean total score for each viewing condition across all groups and current-time product values was 0.549 and 0.098 for the TO.10 and the UF Radiology phantom, respectively. The p-values for the ANOVAs performed at each current-time produc t independently are shown in Table 618. This analysis showed that there was no statistically significant difference in observer performance for the diffe rent viewing conditions. CNR T Determination The purpose for determining the CNRT for each object size was to automate the scoring of the contrast-detai l phantoms. Since there was no statistical difference in observer performance for the four viewing condi tions the observer scores for the standard image with the use of window and level were used for all subsequent analysis. This viewing condition was chosen because it best represents the true clinical viewing conditions. Utilizing the Matlab m-file show n in Appendix C the CNR for all of the objects in each phantom was calculated for each image used in the observer study. It was then necessary to determine the object that wa s considered the thresh old of visibility for each object size, or row, across all observ ers. Since each observer counted the total number of visible objects in each row, these values were averaged to determine the object that the average observer would consider the threshold of visibility. A non integer value was allowed to add precision to the calculation of the CNRT. These average values are listed in Table 6-19 a nd Table 6-20. The CNRT was then calculated for this threshold object by linearly interpolating between the CNRs for the two objects on either side if the average threshold object.

PAGE 98

83 The CNRTs for each phantom are shown in Figure 6-27 and Figure 6-28. The CNRT for each row is nearly independent of the exposure level to the image plate especially for the UF Radiology phantom. Th is behavior became less pronounced as the object size decreased. It should be noted that CNRTs were not calculated for all of the rows in each phantom. The last three rows (smallest three object sizes) of the TO.10 phantom and the last row of the UF Radi ology phantom were excluded from the CNRT analysis. This was done because the small size of the objects produced only a few pixels in the image. The small number of pixels led to large relative errors in the reproducibility of the CNR calculations. The manual image registration procedure disc ussed previously and incorporated in the Matlab m-file shown in Appendix C does ha ve a slight disadvantage to a completely automated procedure. Depending on the exact location of the user input from the image during the registration procedure (see Appendi x C), the CNR calculations for each object can vary. Even with this limitation the manua l registration procedure produced better results than a completely automated proce dure. If a completely automated procedure produced inadequate localization of the objec ts there was no way to correct the problem. The m-file produced the same CNR values each time but they were incorrect. The m-file using the manual registration procedure is capable of exactly reproducing the CNR calculations for multiple analyses of the sa me image if the registration inputs are identical. This ability to exactly reproduce th e CNR calculations was difficult due to the slight variations of the user input duri ng the registration procedure so there were variations in the reproducibility of the CNR calculations. These variations were less than five percent for the larger obj ects but increased to as much as 30 percent for the small

PAGE 99

84 objects. To limit the effect of these varia tions some object sizes were excluded and the CNR calculations were performed five times on each image and the results were averaged. The CNRT determined for each object size was then averaged over all currenttime product values to determine the CNRT to be used in the automated scoring algorithm described below. The CNRT for each object size and each phantom is shown in Table 621. Automated Phantom Scori ng The Average Observer Once a CNRT was determined for each object si ze both contrast-detail phantoms could be scored without the use of observers and a CDSC for the average observer determined. In order to increase the accura cy of the auto-scoring process the CDSC is an interpolated value and not a whole number. This procedure is explained in Appendix C. The total number of visible objects for the ob ject sizes used in this evaluation and the CDSC for all current-time product values for bot h phantoms is shown in Table 6-22. The CDSCs in Table 6-22 were calculated from the sa me images used in the observer study. The auto-scoring algorithm performs very well except at the lowest current-time product, the 0.4 mAs level. This is due to the fact that the CNRT is not exactly independent of exposure level. Since the CNRTs calculated from the images acquired at 0.4 mAs were slightly below the average CNRT for all object sizes and the CNR decreases with decreasing subject contrast, the use of a slightly higher CNRT than was actually measured from the observer data caused fewer objects to be scored as visible. This caused the CDSC to be smaller than the observers score at this current-time product level. Automated scoring reproducibility Due to the variations that can occur in th e calculation of the CNR for each object in an image, the automated scoring algorithms used to calculate the CDSC (see Appendix C)

PAGE 100

85 were tested for reproducibility. A single image of each phantom was acquired at each of the current-time product values s hown in Table 6-22 and five CDSCs were determined from each image. A one-way ANOVA was performed on the results to see if this variability affects the ability of the scoring algorithm to detect small changes in exposure level. The p-value was 0.000 for the analysis of both phantoms so there were statistical differences in the mean CDSC determined at different current-time product values. Posthoc tests showed that all mean CDSCs were statistically different for all current-time product comparisons with a p-value of 0.000 in all cases. This showed that the variation in the calculation of the CNR does not affect the performance of the scoring algorithm. Variation with current-time product Like the NPSC, the CDSC is capable of distinguishi ng between small changes in exposure levels. Five images of the TO.10 pha ntom were acquired wi th the current-time product levels listed in Table 5-5. The variation of the average CDSC with current-time product determined from these im ages, as well as the associat ed standard deviations, is shown in Table 6-23. The ANOVA p-value wa s 0.000 so significant differences exist among the means. The Levene p-value was 0. 227 so the assumption of equal variances was used in the post-hoc tests. The results of the post-hoc tests are shown in Table E-16. All of the CDSCs statistically differed for all curr ent-time product comparisons. Since the CDSC differentiated a 0.2 mAs change with the TO.10 phantom, the CDSC variation with current-time product for the UF Radiology phant om was done only at 0.8, 1.0, 1.2 and 2.0 mAs. These results are also shown in Table 6-23. The ANOVA p-value was 0.000 so significant differences exist among the means. The Levene p-value was 0.408 so the assumption of equal variances was used in the post-hoc tests. The re sults of the post-hoc tests are shown in Tabl e E-17. All of the CDSCs statistically differed for all current-time

PAGE 101

86 product comparisons. With both phantoms, the CDSC increases with increasing currenttime product and 0.2 mAs changes are discernable. Variation with peak-tube potential The variation of the CDSC with peak-tube potential was not characterized since the primary goal of the development of the ideal observer was to determine if small changes in current-time product could be identified. Since this work is specifically designed for pediatric radiology all of the contrast detail analysis was performed with a fixed peaktube potential of 60 kVp. Variation with processing option The effect of processing option on the CDSC was determined by evaluating five images of each phantom acquired with the th ree previously discussed processing options with 60 kVp and 1.0 mAs. These results ar e shown in Figure 6-29 and Figure 6-30. The ANOVA p-value was 0.000 for both phantoms so significant differences exist among the means. The Levene p-value was 0.406 for the TO.10 phantom and 0.141 with the UF Radiology phantom so the assumption of equal variances was used in all post-hoc tests. The results of the post-hoc tests are shown in Table E-18 and Table E-19. All of the CDSCs statistically differed for all processing option comparisons. For both phantoms the hand AP processing option increased the CDSC but the chest PA processing option increased the contrast de tectability the greatest. Automated Phantom Scoring The Ideal Observer In the previous section the CDSC was based on an observer study and was intended to emulate the average observer. During this research a question was posed regarding the mathematical detectability of an object vers us a human observers ability to detect an object. To determine if an object can be mathem atically detected, if the location of that

PAGE 102

87 object is known, the ideal observer was de veloped. The only difference between the average observer algorithm in Appendi x C and the ideal obs erver is the CNRT. For the ideal observer, a constant CNRT (the same for all object size s) was calculated. This CNRT was calculated by performing repeated CN R calculations of two adjacent background regions where no object was located from im ages of both phantoms at all current-time product values investigated to this point. These CNRs are plotted in Figure 6-31. The CNR values were found to be normally distri buted about zero. Sin ce an object can only be present with a posit ive CNR (as defined in this study), the CNRT was defined as the upper limit of the 95-percent c onfidence interval about the m ean of zero. This value was determined to be 0.172. Using this value as the CNRT instead of the CNRTs determined from the viewer study, the images used in the variation w ith current-time product section were scored with the ideal observer algorithm. The result s for the TO.10 phantom are shown in Figure 6-32. At 2.0 mAs and above all of the objects in rows scored by the ideal observer were visible. The results for the UF Radiology pha ntom are shown in Figure 6-33. Not all of the objects were visible for the id eal observer at 3.2 mAs but the CDSC still appeared to be approaching its maximum value with incr easing current-time product. This showed that the UF Radiology phantom could be usef ul over a wider range of exposure levels before all objects become visible. The increase in the CDSC for the ideal observer showed that some objects can be mathematically detected that are not visibl e to a human observer. This could be of clinical relevance but in this application the location of each object is precisely known.

PAGE 103

88 Without this a priori knowledge of not only the location, but the si ze of the object this mathematical detection might not be possible. Anthropomorhic Phantom Evaluation The images of the anthropomorphic phan tom were evaluated by the four radiology residents who participated in the contrast de tail observer study. Two sets of images were acquired; the first using the system diagnos is/full range processi ng option and the second using the upper extremity/full range processing option. All im ages were acquired with 60 kVp. The base image (the one all others we re compared too) was acquired with 60 kVp and 1.0 mAs with the full range processing optio n. The results of this study are shown in Table 6-24. Under the rules defined in Chap ter 5, the use of the chest PA processing option would allow a reduction of 0.6 mAs to obtain an image of minimum acceptable quality. Even though the images acquired with 0.4 and 0.8 mAs and processed with the chest PA processing option received an av erage score of three (having the same perceived image quality as the reference image), the radiology residents were split as to whether these images were better or wors e. This shows the subjectivity of what constitutes a good image. The images proces sed with the chest PA processing option produced much sharper transitions across th e bony structures. When the two residents that rated these two images worse than the reference image were asked why they made this decision (after the viewing session was co mpleted), they both stated that the images processed with the chest PA processing option appeared to be noisier and preferred the smoother images of the full-range processi ng option. Therefore, this study is not sufficient to define a differe nt set of acquisition paramete rs to define an image of minimum acceptable quality based on the full range processing option. It does show that using the same technique factors (60 kVp a nd 1.0 mAs in this instance), all of the

PAGE 104

89 radiology residents noticed an improvement in image quality with the use of the chest PA processing option. This demonstrates the benefit of a truly digital imaging modality such as CR. Effective Dose Calculation One of the limitations of the MOSFET dosimetry system utilized in this study is its sensitivity. The MOSFET dosimeters are not sensitive enough to measure the exposure at the diagnostic exposure levels investigated The effective dose to the anthropomorphic phantom for a CAP exam performed at 60 kVp and the range of current-time product values previously stated was indirectly m easured. The effective dose was measured at 250 mAs and assuming the dose to be linear with current-time pr oduct, the dose was scaled to the desired current-time product. Th e absorbed doses to the organs listed in Table 2-1 were calculated with the method de scribed in Chapter 2 with the exception of the bone marrow and the bone surface. Using a method similar to Johnson3, the absorbed dose from each exposed bone site available for measurement within the anthropomorphic phantom (see Table 6-25) was multiplied by a bone marrow fraction (BMF) or a bone surface fraction (BSF) then summed to get the absorbed dose to the exposed bone marrow and bone surface, respectively. The remaining information needed to comp lete the effective dose calculation was the mass energy-absorption coefficients (en/ ) for the x-ray beam. The average energy of the x-ray beam used in this study was determin ed to be 37.5 keV with the use of TASMIP and MCNP4C as previously described. The en/ values for the three tissue types used in the construction of the anthropomorphic pha ntom are shown in Table 6-26. Multiple dosimetric measurements were made and averag ed for each organ in order to increase the accuracy of the dose measurement. The averag e voltage reading (R) for each organ site,

PAGE 105

90 the MOSFET calibration factor (CF), the appropriate en/ ratio and the tissue weighting factor ( wT) is shown in Table 6-27. The same info rmation used in the calculation of the bone marrow and bone surface absorbed doses is shown in Table 6-28 and Table 6-29. The effective doses for the exposure levels i nvestigated in this research are shown in Table 6-30. 0.00E+00 1.00E-06 2.00E-06 3.00E-06 4.00E-06 5.00E-06 0.40.811.223.2 Current-Time Product (mAs)Average INPSC Figure 6-1. Variation of the INPSC with current-time product at 60 kVp. 0.00E+00 5.00E-08 1.00E-07 1.50E-07 2.00E-07 2.50E-07 3.00E-07 3.50E-07 50607080 Peak-Tube Potential (kVp)Average INPSC Figure 6-2. Variation of the INPSC with peak-tube potential fo r a constant exposure level.

PAGE 106

91 0.00E+00 5.00E-07 1.00E-06 1.50E-06 2.00E-06 2.50E-06 3.00E-06 3.50E-06 Flat FieldChest PAHand AP Processing OptionAverage INPSC Figure 6-3. Variation of the INPSC with processing option. 0.4 0.6 0.8 1 1.2 1.4 1.6 050100150200 Bin NumberRelative Pixel Value Scan Subscan Figure 6-4. Relative pixel valu es across the edge from an image acquired at 60 kVp and 1.0 mAs.

PAGE 107

92 0 0.2 0.4 0.6 0.8 1 01234 Spatial Frequency (mm-1)MTFC 25 Data Points 128 Data Points Figure 6-5. MTFC for 25 and 128 data points in the scan direction from an edge image acquired at 60 kVp and 1.0 mAs. 0 0.2 0.4 0.6 0.8 1 1.2 1.4 0.40.811.223.2 Current-Time Product (mAs)Average IMTFC (mm-1) Figure 6-6. Variation of the IMTFC in the scan direction with current-time product.

PAGE 108

93 0 0.2 0.4 0.6 0.8 1 1.2 0.40.811.223.2 Current-Time Product (mAs)Average IMTFC (mm-1) Figure 6-7. Variation of the IMTFC in the subscan direction with current-time product. 0 0.2 0.4 0.6 0.8 1 01234 Spatial Frequency (mm-1)MTFC Figure 6-8. Five MTFC curves in the scan direction fr om five separate edge images acquired at 0.4 mAs.

PAGE 109

94 0 0.2 0.4 0.6 0.8 1 01234 Spatial Frequency (mm-1)MTFC Figure 6-9. Five MTFC curves in the scan direction fr om five separate edge images acquired at 3.2 mAs. 0 0.2 0.4 0.6 0.8 1 01234 Spatial Frequency (mm-1)MTFC MTFc 1 MTFc 2 Figure 6-10. Smoothing effect of averaging the MTFC before integration. MTFC 1 is the MTFC from a single image. MTFC 2 is an average MTFC from five images all acquired at 0.4 mAs.

PAGE 110

95 0 0.2 0.4 0.6 0.8 1 01234 Spatial Frequency (mm-1)MTFC 0.4 mAs 3.2 mAs Figure 6-11. Average MTFC in the scan direction calculate d from five images acquired at both 0.4 mAs and 3.2 mAs. 0 0.2 0.4 0.6 0.8 1 1.2 1.4 50607080 Peak-Tube Potential (kVp)Average IMTFC (mm-1) Figure 6-12. Variation of the IMTFC in the scan direction with peak-tube potential for a constant exposure level.

PAGE 111

96 0 0.2 0.4 0.6 0.8 1 1.2 1.4 50607080 Peak-Tube Potential (kVp)Average IMTFC (mm-1) Figure 6-13. Variation of the IMTFC in the subscan direction w ith peak-tube potential for a constant exposure level. 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 050100150200 Bin NumberRelative Pixel Value 50 kVp Scan 50 kVp Subscan 80 kVp Scan 80 kVp Subscan Figure 6-14. Change in the ESF with peak-tube potential.

PAGE 112

97 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 Full RangeChest PAHand AP Processing OptionAverage IMTFC (mm-1) Figure 6-15. Variation of the IMTFC in the scan direction with processing option. 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 Full RangeChest PAHand AP Processing OptionAverage IMTFC (mm-1) Figure 6-16. Variation of the IMTFC in the subscan direction with processing option.

PAGE 113

98 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 050100150200 Bin NumberRelative Pixel Value Full Range Chest PA Hand AP Figure 6-17. ESFs for the three proce ssing options in the scan direction. 0 0.2 0.4 0.6 0.8 1 1.2 01234 Spatial Frequency (mm-1)MTFC Full Range Hand AP Figure 6-18. MTF C for the full-range and the hand AP processing options in the scan direction.

PAGE 114

99 0 0.2 0.4 0.6 0.8 1 01234 Spatial Frequency (mm-1)MTFC Standard Window + Window Level + Level Figure 6-19. The MTFC for different window and level se ttings in the scan direction. 0 50 100 150 200 250 300 350 400 450 0.40.811.223.2 Current-Time Product (mAs)Average IDQEC (mm-1) Figure 6-20. Variation of the IDQEC with current-time product in the scan direction.

PAGE 115

100 0 50 100 150 200 250 300 350 0.40.811.223.2 Current-Time Product (mAs)IDQEC (mm-1) Figure 6-21. Variation of the IDQEC with current-time product in the subscan direction. 0 50 100 150 200 250 300 50607080 Peak-Tube Potential (kVp)Average IDQEC (mm-1) Figure 6-22. Variation of the IDQEC with peak-tube potential in the scan direction.

PAGE 116

101 0 50 100 150 200 250 50607080 Peak-Tube Potential (kVp)Average IDQEC (mm-1) Figure 6-23. Variation of the IDQEC with peak-tube potential in the subscan direction. 0 100 200 300 400 500 600 Full RangeChest PAHand AP Processing OptionAverage IDQEC (mm-1) Figure 6-24. Variation of the IDQEC in the scan direction with processing option.

PAGE 117

102 0 50 100 150 200 250 300 350 400 450 Full RangeChest PAHand AP Processing OptionAverage IDQEC (mm-1) Figure 6-25. Variation of the IDQEC in the subscan direction with processing option. 0 50 100 150 200 250 300 350 01234 Spatial Frequency (mm-1)DQEC Full Range Chest PA Hand AP Figure 6-26. Variation of the DQEC in the subscan direction with processing option.

PAGE 118

103 0 0.5 1 1.5 2 2.5 3 3.5 4 0.40.811.223.2 mAsCNRT Row A Row B Row C Row D Row E Row F Row G Row H Row J Figure 6-27. CNRT for each object size as a function of acquisition current-time product for the TO.10 phantom. 0 0.5 1 1.5 2 2.5 3 3.5 4 0.40.811.223.2 mAsCNRT Row A Row B Row C Row D Row E Row F Row G Row H Figure 6-28. CNRT for each object size as a function of acquisition current-time product for the UF Radiology phantom.

PAGE 119

104 56 57 58 59 60 61 62 63 Full RangeChest PAHand AP Processing OptionCDSC Figure 6-29. Variation of the CDSC with processing option for the T0.10 phantom. 31 32 33 34 35 36 Full RangeChest PAHand AP Processing OptionCDSC Figure 6-30. Variation of the CDSC with processing option for the UF Radiology phantom.

PAGE 120

105 0 5 10 15 20 25 -0.16-0.0800.080.16 CNRCNR Frequency Figure 6-31. Repeated CNR calculations of two adjacent background regions. 76 77 78 79 80 81 82 0.40.811.223.2 Current-Time Product (mAs)CDSC Figure 6-32. The CDSC for the ideal observer and the TO.10 phantom.

PAGE 121

106 38 40 42 44 46 48 50 52 0.40.811.223.2 Current-Time Product (mAs)CDSC Figure 6-33. The CDSC for the ideal observer an d the UF Radiology phantom. Table 6-1. Average INPSCs and their associated standard deviations. mAs Average INPSC 0.4 4.397E-06 2.662E-08 0.8 2.175E-06 2.490E-08 1.0 1.716E-06 2.556E-08 1.2 1.353E-06 1.920E-08 2.0 7.663E-07 2.099E-09 3.2 4.831E-07 3.011E-09 Table 6-2. Average INPSCs and their associated standard deviations. kVp Average INPSC 50 3.252E-07 4.539E-09 60 2.871E-07 2.658E-09 70 2.856E-07 8.771E-09 80 2.846E-07 1.109E-08 Table 6-3. Average INPSCs and their associated standard deviations. Processing Option Average INPSC System Diagnosis/Flat Field 1.696E-06 5.261E-8 Chest/Chest PA 1.362E-06 7.322E-8 Upper Extremity/Hand AP 3.068E-06 2.485E-7

PAGE 122

107 Table 6-4. Edge comparison for the IMTFC (mm-1) in the scan and subscan directions. Orientation Edge 1 Edge 2 p-value Scan 1.2316 1.0944 0.372 Subscan 1.2370 1.1038 0.132 Table 6-5. Calculation based variability in the IMTFC. IMTFC (mm-1) Evaluation Scan Subscan 1 1.2337 1.0960 2 1.2337 1.0919 3 1.2321 1.0924 4 1.2293 1.0976 5 1.2287 1.0919 Table 6-6. Image based variability in the IMTFC. IMTFC (mm-1) Image Scan Subscan 1 1.2413 1.0994 2 1.2350 1.0985 3 1.2360 1.0919 4 1.2282 1.0892 5 1.2448 1.0931 Table 6-7. Edge angle in degrees for repeat ed setups of the edge device and repeated images of the same setup. Repeated Setups Same Setup 2.49 2.70 3.30 2.70 2.50 2.80 2.90 3.00 3.10 2.80 2.70 2.50

PAGE 123

108 Table 6-8. IMTFCs for various ESF lengths in the scan direction. Number of ESF Data Points IMTFC (mm-1) 25 1.5175 50 1.3428 100 1.2562 150 1.2117 250 1.1574 Table 6-9. IMTFCs for various ESF lengths in the subscan direction. Number of ESF Data Points IMTFC (mm-1) 25 1.3365 50 1.2089 100 1.1180 150 1.0756 250 1.0374 Table 6-10. Average IMTFCs and their associated standard deviations. Average IMTFC (mm-1) mAs Scan Subscan Scan Subscan 0.4 1.27006 1.13638 0.026643 0.020392 0.8 1.24306 1.10602 0.016477 0.003792 1.0 1.23378 1.09542 0.008067 0.009759 1.2 1.23800 1.09142 0.006496 0.008100 2.0 1.22678 1.08640 0.013779 0.008329 3.2 1.23396 1.08494 0.006065 0.006593 Table 6-11. The INPSC and IMTFC value for different window and level settings. IMTFC (mm-1) Image Adjustment INPSC Scan Direction Standard 5.7631E-08 1.2414 Window Increased 5.9711E-37 4.1171 Window Decreased 5.0270E-10 1.2493 Brightness Increased 5.9411E-37 1.7753 Brightness Decreased 5.7631E-08 2.2583

PAGE 124

109 Table 6-12. The value of Q for each peak-tube potential. kVp Q (photonsmm-2mR-1) 50 1.5518E5 60 1.8253E5 70 2.0325E5 80 2.5305E5 Table 6-13. In bucky exposure meas urements at 60 kVp for the DQEC variation with current-time product study. mAs Exposure (mR) 0.4 0.044 0.8 0.106 1.0 0.140 1.2 0.182 2.0 0.334 3.2 0.570 Table 6-14. In bucky exposure measurements for the DQEC variation with peak-tube potential study. kVp Exposure (mR) 50 1.03 60 1.04 70 1.04 80 1.02 Table 6-15. Number of visible objects in the TO.10 phantom. Viewing Conditions mAs Standard Standard W/L Inverted Inverted W/L 0.4 55.92 62.15 57.54 60.23 0.8 65.92 67.23 64.08 66.08 1.0 71.15 71.77 71.00 72.62 1.2 73.23 75.69 73.00 74.85 2.0 81.23 82.92 81.15 81.69 3.2 87.31 89.46 87.38 88.46

PAGE 125

110 Table 6-16. Number of visible objects in the UF Radiology phantom. Viewing Conditions mAs Standard Standard W/L Inverted Inverted W/L 0.4 30.33 32.42 32.50 31.42 0.8 33.42 34.92 33.50 34.50 1.0 36.42 37.50 35.92 36.92 1.2 34.83 36.25 35.17 35.92 2.0 37.33 39.42 37.50 38.58 3.2 39.67 41.67 39.25 41.25 Table 6-17. ANOVA p-values for total score comparisons by observer group. mAs TO.10 Phantom UF Radiology Phantom 0.4 0.100 0.675 0.8 0.101 0.914 1.0 0.056 0.534 1.2 0.064 0.753 2.0 0.003 0.249 3.2 0.114 0.442 Table 6-18. ANOVA p-values for total sc ore comparisons by viewing condition. mAs TO.10 Phantom UF Radiology Phantom 0.4 0.233 0.412 0.8 0.860 0.759 1.0 0.952 0.824 1.2 0.935 0.680 2.0 0.969 0.272 3.2 0.844 0.181 Table 6-19. Average threshold object for the TO.10 phantom. Row mAs A B C D E F G H J K L M 0.4 7.00 6.15 5.85 6.38 6.00 6.08 6.62 5.92 4.85 4.69 1.92 0.69 0.8 7.15 6.38 6.46 5.85 6.08 5.38 7.08 7.08 5.62 5.00 3.38 1.77 1.0 7.46 7.15 6.62 6.92 6.08 6.00 7.38 7.15 5.15 5.38 4.46 2.00 1.2 7.54 7.62 6.54 7.46 6.85 6.69 7.38 7.62 6.15 6.62 4.85 0.38 2.0 8.15 7.92 7.85 7.92 7.15 7.00 7.62 7.46 6.85 6.69 5.31 3.00 3.2 8.46 8.54 7.92 8.54 8.69 8.23 8.62 7.62 6.92 8.08 4.69 3.15

PAGE 126

111 Table 6-20. Average threshold objec t for the UF Radiology phantom. Row mAs A B C D E F G H J 0.4 4.75 4.67 3.92 4.25 3.58 3.58 3.17 2.50 2.00 0.8 4.83 4.33 4.17 4.17 4.00 3.92 3.67 3.17 2.67 1.0 5.58 4.50 4.42 4.42 4.33 4.17 4.00 3.33 2.75 1.2 5.00 4.50 4.25 4.33 4.50 4.25 3.83 3.08 2.50 2.0 5.33 5.00 5.00 4.75 4.33 4.08 4.00 3.50 3.42 3.2 6.08 5.08 5.00 5.00 4.75 5.00 4.17 3.75 2.83 Table 6-21. CNRT for each object size and each phantom. CNRT Object Size (Row) TO.10 Phantom UF Phantom Row A 0.6573 0.7307 Row B 0.7402 0.9643 Row C 0.8012 0.9877 Row D 1.0395 1.0851 Row E 1.3596 1.3425 Row F 1.3786 1.4021 Row G 2.0702 1.7583 Row H 2.2272 2.4761 Row J 2.6991 N/A Table 6-22. Observer study and CDSC comparison. Total Observer Score CDSC mAs TO.10 Phantom UF Phantom TO.10 Phantom UF Phantom 0.4 54.85 30.42 48.76 27.07 0.8 57.08 32.25 56.92 32.29 1.0 59.92 34.75 60.75 33.60 1.2 63.85 33.75 62.36 34.50 2.0 67.92 36.00 68.40 37.15 3.2 73.54 38.83 73.21 39.07

PAGE 127

112 Table 6-23. Variation of the CDSC with current-time product. Average CDSC mAs TO.10 Phantom UF Phantom 0.4 48.95 1.07 0.8 56.20 1.25 31.77 0.39 1.0 59.70 0.69 32.87 0.36 1.2 62.32 0.53 33.90 0.30 2.0 67.21 1.08 36.71 0.16 3.2 72.02 0.89 Table 6-24. Radiology reside nts evaluation of anthropomorphic phantom images. Radiology Resident Processing Option mAs 1 2 3 4 Average Full Range 0.4 2 3 3 2 2.50 Full Range 0.8 2 2 2 3 2.25 Full Range 1.0 Full Range 1.2 3 3 3 2 2.75 Full Range 2.0 3 4 3 4 3.50 Full Range 3.2 2 4 3 5 3.50 Chest PA 0.4 4 2 4 2 3.00 Chest PA 0.8 4 2 4 2 3.00 Chest PA 1.0 5 3 4 3 3.75 Chest PA 1.2 4 4 4 3 3.75 Chest PA 2.0 5 3 5 3 4.00 Chest PA 3.2 5 4 5 4 4.50 Table 6-25. The exposed bone sites a nd their associated BMFs and BSFs. Bone Site BMF BSF Right Arm 0.0779 0.0980 Left Arm 0.0779 0.0980 Pelvis 0.2191 0.2191 Spine 0.1588 0.1850 Skull 0.1470 0.1660 Table 6-26. Mass energy-apsorp tion coefficients for the f our tissue types at 35 keV. Tissue Type en/ (cm2/g) Air 0.0871 Soft Tissue 0.0898 Bone 0.3003 Lung 0.9213

PAGE 128

113 Table 6-27. Calculation of th e Effective dose at 60 kVp. Organ R (mV) CF (mR/mV) en/ Ratio wT E (mSv) Gonads 12.67 29.82 1.0307 0.20 0.6821 Colon 15.33 29.38 1.0307 0.12 0.4881 Lung 24.00 33.99 1.0577 0.12 0.9070 Stomach 16.67 33.99 1.0307 0.12 0.6138 Bladder 18.00 31.72 1.0307 0.05 0.2578 Breast 25.33 29.82 1.0307 0.05 0.3410 Liver 16.00 31.72 1.0307 0.05 0.2291 Esophagus 9.67 29.82 1.0307 0.05 0.1301 Thyroid 22.33 29.38 1.0307 0.05 0.2962 Skin 25.33 29.82 1.0307 0.01 0.0682 Bone marrow (red) See Table 6-28 0.12 0.5311 Bone Surface See Table 6-29 0.01 0.0500 Remainder 6.44 31.22 1.0307 0.05 0.0907 Total = 4.6853 Table 6-28. Calculation of the ab sorbed dose to the bone marrow. Bone Site R (mV) CF (mR/mV) en/ Ratio BMF E (mSv) Right Arm 15.33 31.72 3.4472 0.0779 1.1441 Left Arm 15.33 31.72 3.4472 0.0779 1.1441 Pelvis 4.67 33.99 3.4472 0.2191 1.0495 Spine 4.00 29.82 3.4472 0.1588 0.5720 Skull 3.67 31.72 3.4472 0.1470 0.5163 Total = 4.4260 Table 6-29. Calculation of the ab sorbed dose to the bone surface. Bone Site R (mV) CF (mR/mV) en/ Ratio BMF E (mSv) Right Arm 15.33 31.72 3.4472 0.0980 1.4393 Left Arm 15.33 31.72 3.4472 0.0980 1.4393 Pelvis 4.67 33.99 3.4472 0.2191 1.0495 Spine 4.00 29.82 3.4472 0.1850 0.6664 Skull 3.67 31.72 3.4472 0.1160 0.4074 Total = 5.0019 Table 6-30. Effective doses for clinical pediatric current-time product levels. mAs E (mSv) 0.4 0.0750 0.8 0.01499 1.0 0.01874 1.2 0.02249 2.0 0.03748 3.2 0.05997

PAGE 129

114 CHAPTER 7 CONCLUSIONS AND FUTURE WORK The primary objective of this research was to investigate the feasibility of developing a quantitative metric that can be us ed to assess image qua lity and therefore set acquisition parameters for pe diatric CR exams without the use of complicated observer studies. In order to reach this objective, a series of intermediate objectives were established as steps that would lead to the completion of the primar y objective. Each of these intermediate objectives were detailed in the first chapter and are reiterated below with the conclusions reached from each objective immediately following. Objective A. To determine if the traditional methods of calculating NPS, MTF and DQE can be applied to images acquired in a non-traditional manner at diagnostic exposure levels. The traditional methods of determining the NPS were easily adapted to the calculation of the NPSC. The addition of simulated patien t scattering media and the use of low exposure levels did not change the mathem atical procedures used in the calculation of the NPS and were reproduced to calculate the NPSC. Of the two traditional methods of determining the MTF, only the edge method was adaptable to the calculation of the MTFC. Due to the low exposures the slit method did not provide an adequate data set for the calculation of the MTFC. The mathematical procedures used in the edge method did not need altering in the calculation of the MTFC. The only alteration of the traditional methods was the introduction of simulated pa tient scatter and placi ng the image receptor in the under-table bucky for image acquisition.

PAGE 130

115 Since the DQE is a combination of the NPS and the MTF, the DQEC was easily determined using the traditional definition of the DQE. The only change in the definition of the DQEC from the DQE is the point in the imaging chain were the parameter is applicable. The DQE, through the application of the parameters k and Q in equation 2-13, is a measure of the SNR properties of the imaging system at the image receptor. By defining the parameter k as unity due to the background trend correction, the DQEC is a measure of the SNR properties of the final image as viewed by a human observer before the image is altered with active image manipulation tool s (e.g. window and level). Objective B. To quantify the variation of the NPSC and MTFC with current-time product and peak-tube potential. In order to readily compare the NPSC and the MTFC as determined from images acquired under differing acquis ition parameters the integral of these functions up to the Nyquist frequency were used. The INPSC was found to distinguish small changes in exposure levels. As the current-time product was increased for a constant peak-tube potential, the INPSC decreased as expected, but when the exposure level was kept relatively constant the INPSC did not change with peak-tub e potential above 60 kVp. The IMTFC was not able to detect changes in curr ent-time product for a constant peak-tube potential over the range of current-time product values inve stigated in this work, but when the exposure level was kept relatively constant the IMTFC increased with increasing peak-tube potential. This behavior was primarily due to the method used to determine the MTFC. With the adaptation of the traditional edge method by introducing simulated patient scatter and the acquisition of the edge images in the under-table bucky, the scattering of photons produced a shoulder on the attenuated side of the ESF. This

PAGE 131

116 caused a decrease in the MTFC that would not occur with the traditional MTF. As the peak-tube potential was increased, the scat tering decreased and the magnitude of the MTFC increased leading to an increase in the IMTFC. The use of the INPSC or the IMTFC alone as a measure of image quality was insufficient to identify changes in both current-time product and peak-tube potential. These two metrics were then combined to produce the DQEC. As the current-time product was increased for a constant peak-tube potential, the IDQEC decreased. This behavior is not intuitive and is a result of th e choice of normalization of the DQEC. The parameter Q in Equation 2-13 was the dominati ng factor that caused the IDQEC to decrease with increasing current-time product. This behavior is consistent with the traditional DQE as presented by other researchers.4,5 This decrease with in creasing current-time product brings into question the clini cal applicability of the IDQEC as a measure of image quality. Traditionally the DQE has been used to compare different aspects of radiographic imaging systems or the overall performance of the entire imaging chain. A system with a higher DQE is considered to be the better imaging system. In this case, the IDQEC leads to the conclusion that as the current-time pr oduct is increased over the range of currenttime product values investigated in this st udy, the image quality would decrease when in reality the opposite is true. When the exposure level was ke pt relatively constant, the IDQEC decreased as the peak-tube potential was increased. The discussion above regarding the trend of the IDQEC with current-time product is not applicable since the optimum peak-tube potential, both from the clinical image quality and patient dose perspectives, depends on the type of radi ographic study being done. The fact that the IDQEC was statistically different at all peak -tube potential values only in the scan

PAGE 132

117 direction showed that even though the IDQEC is not an ideal indicator of clinical image quality, the IDQEC in the scan direction was the only parameter capable of identifying changes both in current-time product and peak-tube potential. Objective C. To quantify the effect of a ny computer processing of the raw data before the image is viewed on the evaluations of the NPSC, MTFC and the CDSC. The ability to digitally manipulate the raw image data before it is viewed by a human observer further confuses th e notion of using the metrics INPSC, IMTFC and IDQEC as indicators of clinical image quality. When the ch est PA processing option is used, the INPSC decreased, the IMTFC increased and the IDQEC increased. All of these are desirable and lead to the conclusion th at an image processed with the chest PA processing option would be supe rior in quality to an imag e processed with the system diagnosis processing options. The confusion occurs when the hand AP processing option is used. The INPSC increased, which is not desirable, and the IMTFC increased even more than with the chest PA processing option, but the IDQEC did not statistically change. In this case, the IDQEC would indicate that the system di agnosis processing options and the hand AP processing option would produce an imag e of equal quality when in reality they would visually appear dramatica lly different. The use of the IDQEC as a measure of clinical image quality can be further questioned if the shapes of the DQEC curves are examined. Since the DQEC is a function of spatial frequency, the magnitude of the DQEC at any particular spatial frequency value w ould depend on what is being imaged. If small objects are to be imaged, the ideal DQEC would have a large ma gnitude at high spatial frequencies while in contrast if large objects are to be imaged, the ideal DQEC would have a large magnitude at low spatial frequencies. So even though the IDQEC was not

PAGE 133

118 statistically different between the system di agnosis processing options and the hand AP processing options, the DQEC curves had dramatically different shapes. The hand AP processing option produced a DQEC that has a larger magnitude at higher spatial frequencies which makes clinical sense since that processing option is designed to image the small bones in a hand. These results show ed that the processi ng options are truly application specific. In order to produce an image with the optimum noise properties or the optimum subject-contrast transfer propert ies, the ability to digitally manipulate the raw image data allows for application specific alteration of these properties. The questions regarding the use of the IDQEC as a measure of clinical image quality were confirmed when the variation of the CDSC with the different processing options was investigated. The chest PA pro cessing option produced an image with the best overall contrast de tectability which is in agreement with the IDQEC results. In contrast, the hand AP proce ssing option, which was supposed to produce an image of equal quality with that of the system diagnos is processing options, showed an increase in contrast detectability. The application specific nature of the processing options also raises a question about the app licability of the CDSC as a broad measure of clinical image quality. Even though the chest AP processing option produces a higher CDSC than the hand AP processing option, an image of the same physical object might not necessarily be of better quality if the chest PA pro cessing option is applie d. A radiograph of a pediatric hand, if processed with the ches t PA processing option, might not have the desired detail of the small bones due to the smoothing operation that is performed. This examination specific application of digital pr ocessing makes the identification of a broad indicator of image quality not possible.

PAGE 134

119 Objective D. To quantify the effect of gr ayscale inversion on th e evaluation of the CDSC. Through the observer studies, it was determin ed that the contrast detectability of the human observers was not affected by th e inversion of an images grayscale. Objective E. To determine if the calculation of the NPSC and the MTFC with the original image data is directly applicable, as a measure of image quality, to an image displayed on the radiologists m onitor that has been manipulated. The artificial adjustments of the window and level settings on the images before the calculation of the NPSC and the MTFC had a profound effect on the magnitude of the INPSC and the IMTFC. These artificial adjustments simulate what the effect would be of capturing the image data from the viewing station and then performing the calculations. The determination of the NPSC and the MTFC from the raw image data gives a metric that is representative of the traditional mean ing of the NPS and the MTF, as system performance parameters. If the manipulated im age data is used, the calculation the NPSC and the MTFC are more representative of an image NPS or MTF (i.e., a parameter specifically tailored to the image being displayed). Since these two methods of determining the NPSC and the MTFC produce dramatically different results, it was concluded that the calculation of the NPSC and the MTFC with the original image data is not directly applicable, as a measure of image quality, to an image displayed on the radiologists monitor that has been manipulated. Objective F. To quantify the effect of this dynamic image manipulation on the determination of the CDSC through observer studies.

PAGE 135

120 The artificial adjustment of the window and level settings was not necessary in quantifying the effect of thos e manipulation tools on the CDSC. The observer study was designed to answer this question directly When the observer study was conducted using a clinical diagnostic viewing station in the radiology department, the use of the window and level controls did not statistically change the contrast detectability performance of the human observers. The display monitor was suffi ciently capable of displaying all available information in the images of the contrast-d etail phantoms without the need for digital manipulation. Objective G. To correlate the newly developed image quality metrics with a qualitative evaluation of anthropomorphic pha ntom images and identify the radiation dose associated with the exam parame ters used to obtain those images. The qualitative evaluation of the anthropomorphic phant om failed to identify a different set of image acquisition parameters that would produce an image of minimum acceptable quality than those already utilized by the Radiology Department at Shands Hospital. The currently used clinical acquis ition parameters for a CAP exam of a patient the size of the anthropomorphic phantom are 60 kVp, 1.0 mAs and the appropriate pediatric chest processing option. The metrics investigated in this study determined from the appropriate images acquired under these acquisition parameters are shown in Table 71. Due to previously mentioned reasons, th ese values can not be applied as overall measures of the quality of an image produced with these acquisition parameters and are just shown for completeness. Even though a single quantitativ e metric to evaluate the ov erall quality of an image was not identified, the contrast-detail obser ver study revealed important information

PAGE 136

121 about how the noise content in an image affect s the low-contrast detectability of different sized objects. Since the CNRT for each object size in the contrast-detail phantoms was almost independent of the exposure le vel, this identified the minimum CNRT that would be necessary for an object of that size to be v isible in a clinical im age. As the size of the objects decreased in each cont rast-detail phantom, the CNRT increased. This showed that the contrast required for an obj ect to be visible was greate r for smaller objects since the noise level was approximately constant th roughout the image. This behavior was consistent for each phantom but when the size of the objects and the CNRTs for those objects were compared between the two phant oms, the opposite behavior was seen. For a given object size a higher CNRT was determined for the UF phantom. The best example of this was comparing the objects in row E of both phantoms. The objects are 26 pixels in diameter in the TO.10 phantom and are 47 pixe ls in diameter in the UF phantom, but the CNRT was almost identical for that object si ze in each phantom. The sizes of the objects in each row for both phantoms as determin ed from the phantom characterization radiographs are shown in Table 7-2. The actual object sizes in the images used in the observer study were larger due to geometric magnification but the relative sizes were the same. These, when compared with Table 6-21, further show this tr end in object size and CNRT for the inter-phantom comparisons. Ther e are two plausible reasons for this behavior. The first involves th e limited number of contrast steps in the UF phantom as compared to the TO.10 phantom. There is a smaller difference between the subject contrast of adjacent objects in each row in the TO.10 phantom than in the UF phantom. Since the objects are not of the same size in each phantom this might be causing this unexpected result. The second involves the regular pattern of the objects in the contrast-

PAGE 137

122 detail phantoms. This regul ar pattern can influence th e observers threshold of detectability since they know precisely wher e each object is located. In order to further investigate these findings, the construction of two new contrast-detail phantoms and a new observer study is proposed below. The new contrast-detail phantoms will allo w the use of the 4-AFC experimental methodology. One phantom will utilize holes of varying depths in acrylic to produce the subject contrast while the second will utiliz e varying thicknesses of aluminum embedded in acrylic. The physical dimensions of the phantom will allow its use with a 24-cm 30cm imaging plate with the addition of up to 10 cm of simulated patient scatter at a maximum SID of 72 inches. Both phantoms will consist of a tray constructed from 1-mm thick sheets of acrylic with internal dimensi ons of 20 cm 25 cm 1 cm. The objects, of varying size and subject contrast will be located in one of the four corners of 2.5-cm 2.5-cm 1-cm acrylic blocks. These acrylic blocks will be removable so they may be randomized for each experiment. There will be eight object sizes and nine contrast steps for each object size. In order to use the phantom for a 4-AFC observer study the small acrylic blocks will need to be arranged in th e tray. It is proposed that the top row consist of eight acrylic blocks that will have each object size at the maxi mum subject contrast located in the center of the block. The bloc ks will be arranged so that objects of decreasing size move from left to right. Belo w each of these blocks the blocks with the objects of the same size will be placed in a column in a random order. The phantom will then be placed between the desired amount of simulated patient sca tter. Each object size and the depth of the hole or the thickness of lead to produce the same subject contrasts (at a 60-kVp peak-tube potential) are sh own in Table 7-3 and Table 7-4.

PAGE 138

123 An observer study conducted with both of these phantoms should identify the reason for the increased CNRT for larger objects in the UF Radiology phantom. Since a 4-AFC experimental methodology will eliminate any bias on the observers threshold of detectability, a more accurate CNRT for each object size should result. Even though the results of this type of study would not be a pplicable as a direct measure of image quality they would be useful in identi fying the noise properties that an image could have for the identification of circular objects of different sizes. Table 7-1. Metrics evaluated at 60 kVp, 1.0 mAs with the chest PA processing option. INPSC 1.362E-6 Scan (mm-1) Subscan (mm-1) IMTFC 1.421 1.378 IDQEC 481.4 404.9 TO.10 Phantom UF Phantom CDSC 62.35 35.16 Table 7-2. Object diameters in image pixe ls for each row of both contrast-detail phantoms. Row TO.10 Phantom UF Phantom A 100 120 B 73 89 C 50 74 D 38 62 E 26 47 F 18 33 G 15 25 H 12 18 J 8 N/A

PAGE 139

124 Table 7-3. Physical dimensions in mm of the objects in the proposed contrast-detail phantom using drilled holes for subject contrast. Object Size 2.00 3.21 4.43 5.64 6.86 8.07 9.29 10.50 Depth 1 10.00 7.50 7.50 7.50 7.50 7.50 5.00 5.00 Depth 2 8.81 6.59 6.59 6.59 6.59 6.59 4.41 4.41 Depth 3 7.63 5.69 5.69 5.69 5.69 5.69 3.81 3.81 Depth 4 6.44 4.78 4.78 4.78 4.78 4.78 3.22 3.22 Depth 5 5.25 3.88 3.88 3.88 3.88 3.88 2.63 2.63 Depth 6 4.06 2.97 2.97 2.97 2.97 2.97 2.03 2.03 Depth 7 2.88 2.06 2.06 2.06 2.06 2.06 1.44 1.44 Depth 8 1.69 1.16 1.16 1.16 1.16 1.16 0.84 0.84 Depth 9 0.50 0.25 0.25 0.25 0.25 0.25 0.25 0.25 Table 7-4. Physical dimensions in mm of the objects in the proposed contrast-detail phantom using lead disks for subject contrast. Object Size 2.00 3.21 4.43 5.64 6.86 8.07 9.29 10.50 Thickness 1 1.63 1.22 1.22 1.22 1.22 1.22 0.81 0.81 Thickness 2 1.43 1.07 1.07 1.07 1.07 1.07 0.72 0.72 Thickness 3 1.24 0.92 0.92 0.92 0.92 0.92 0.62 0.62 Thickness 4 1.05 0.78 0.78 0.78 0.78 0.78 0.52 0.52 Thickness 5 0.85 0.63 0.63 0.63 0.63 0.63 0.43 0.43 Thickness 6 0.66 0.48 0.48 0.48 0.48 0.48 0.33 0.33 Thickness 7 0.47 0.34 0.34 0.34 0.34 0.34 0.23 0.23 Thickness 8 0.27 0.19 0.19 0.19 0.19 0.19 0.14 0.14 Thickness 9 0.08 0.04 0.04 0.04 0.04 0.04 0.04 0.04

PAGE 140

125 APPENDIX A INPSC MATLAB M-FILE AND DESCRIPTION This appendix lists the Matlab m-f ile developed to calculate the INPSC. Following the m-file is a description of how the code functions as well as the built in Matlab functions that were used. Fina lly, the code is tested with known inputs and the outputs are examined to ensure proper functionality. INPS C MATLAB m-file % This code calculates the INPSC from M^2 different NxN regions of the image % and averages them. The data is smoot hed with a Hanning truncation function % before the FFT is performed clear all; close all; % Prompt user for file name prompt = {'Enter file name with extension:'}; dlgTitle = 'File Name Input'; lineNo = 1; phantom_image_filename = char(inp utdlg(prompt,dlgTitle,lineNo)); % Read in image data phantom_image = dlmread(phantom_image_filename); % Locate center of image and select usable data N = 128; % Length of data in x direction (number of columns) M = 8; % Number of square data repetitions in x and y direction center_x = round(size(p hantom_image,2)/2); center_y = round(size(p hantom_image,1)/2); usable_data = double(phantom_imag e(center_y N*M/2:center_y +... N*M/2 1,center_x N *M/2:center_x + N*M/2 1)); % Perform abs(fft(ROI))^2 and write to a matrix r1 = 1; r2 = N; c1 = 1; c2 = N;

PAGE 141

126 for n = 1:M, for m = 1:M, roi_rect = usable_data(r1:r2,c1:c2); roi_fit = heel_effect(roi_rect,N ); % Subtract out the background trends ROI_fft = hanning_fft2(roi_fit,N); % Apply a Hanning filter and FFT2 ABS_ROI = abs(ROI_fft)./N^2; ABS_SQ_ROI = ABS_ROI.^2; USABLE_DATA(r1:r2,c1:c2) = ABS_SQ_ROI; r1 = r2+1; r2 = r1+N-1; end r1 = 1; r2 = N; c1 = c2+1; c2 = c1+N-1; end % Calculate FFT Ensamble Average r1 = 1; r2 = N; c1 = 1; c2 = N; ENS_SUM = zeros(N,N); for n = 1:M, for m = 1:M, ENS_SUM = ENS_ SUM + USABLE_DATA(r1:r2,c1:c2); r1 = r2+1; r2 = r1+N-1; end r1 = 1; r2 = N; c1 = c2+1; c2 = c1+N-1; end ENS_AVE = ENS_SUM./(M^2); % Calculate NPSC Nx = N; Ny = N; delta_x = 1/8.8; delta_y = 1/8.8; NPSC_2D = ENS_AVE./(Nx*Ny).*(delta_x*delta_y); % % Determine 1D NPSC from 2D data %

PAGE 142

127 % Calculate 1D frequency values f_1D = 0:1/(N*delta_x):1/( 2*delta_x)-(1/(N*delta_x)); % Calculate 2D frequency values for m = 1:4, for n = 1:N/2, f_2D(m,n) = sqrt(f_1D(m+1)^2 + f_1D(n)^2); end end f_2D(5:5,1:N/2) = f_2D(4:4,1:N/2); f_2D(6:6,1:N/2) = f_2D(3:3,1:N/2); f_2D(7:7,1:N/2) = f_2D(2:2,1:N/2); f_2D(8:8,1:N/2) = f_2D(1:1,1:N/2); %Extract 1D data DATA_V_2D = NPSC_2D(2:5,1:N/2); DATA_V_2D(5:8,1:end) = N PSC_2D(end-3:end,1:N/2); DATA_U_2D = NPSC_2D(1:N/2,2:5)'; DATA_U_2D(5:8,1:end) = NPSC _2D(1:N/2,end-3:end)'; % Bin data in delta f width bins warning off MATLAB:divideByZero hits = 0; NPSC_1D_V_temp = zeros(1,N/2); NPSC_1D_U_temp = zeros(1,N/2); for c = 1:N/2-1, for m = 1:8, for n = 1:N/2, if f_2D(m,n ) >= f_1D(c) & f_2D(m,n) < f_1D(c+1) NPSC_1D_V_tem p(c) = NPSC_1D_V_temp(c) + DATA_V_2D(m,n); NPSC_1D_U_tem p(c) = NPSC_1D_U_temp(c) + DATA_U_2D(m,n); hits = hits + 1; end end end NPSC_1D_V(c) = NPSC_1D_V_temp(c)/hits; NPSC_1D_U(c) = NPSC_1D_U_temp(c)/hits; hits = 0; end % Assign frequency components greater than f_1D(128) to NPSC_1D(128) hits = 0; for m = 1:8, for n = 1:N/2, if f_2D(m,n) >= f_1D(N/2)

PAGE 143

128 NPSC_1D_V_temp(N/2) = NPSC_1D_V_temp(N/2) + DATA_V_2D(m,n); NPSC_1D_U_temp(N/2) = NPSC_1D_U_temp(N/2) + DATA_U_2D(m,n); hits = hits + 1; end end end NPSC_1D_V(N/2) = NPSC _1D_V_temp(N/2)/hits; NPSC_1D_U(N/2) = NPSC _1D_U_temp(N/2)/hits; subplot(1,2,1), plot(f_1D (5:end),NPSC_1D_V(5:end)) title({'One-Dimensional NPSC','i n the Sub-Scan Direction'}) xlabel('1/mm') ylabel('NPSC Magnitude') subplot(1,2,2), plot(f_1D (5:end),NPSC_1D_U(5:end)) title({'One-Dimensional NPSC', 'in the Scan Direction'}) xlabel('1/mm') ylabel('NPSC Magnitude') % Calculate total power INPSC = trapz(trapz(NPSC_2D))*(1/( N*delta_x))*(1/(N*delta_y)); % Scale, Shift and Image the 2D NPSC SCALED_NPSC_2D = NPSC_2D; SCALED_NPSC_2D(1,1) = NPSC_2D(5,5); SCALED_NPSC_2D = l og(SCALED_NPSC_2D); SCALED_SHIFTED_NPSC_2D = ff tshift(SCALED_NPSC_2D); figure imagesc(SCALED_SHIFTED_NPSC_2D) colormap(gray) axis image off title({'Two-Dimensional NPSC for File: phantom_image_filename}) text(1,6,{'Total Power = num2 str(total_power)},' Color',[1 1 1]) % Write NPSC data to a file NPSC_f = f_1D; dlmwrite('NPSC_V.txt', NPSC_1D_V, '\t') dlmwrite('NPSC_U.txt', NPSC_1D_U, '\t') dlmwrite('NPSC_f.txt', f_1D, '\t') Description of MATLAB m-file and Built in MATLAB Functions The first line of code illustrates how comme nts are inserted into an m-file. Any text following a % is automatically considered a comment. Then to initialize the system, the commands of clear all and clos e all are given. This closes all variables in the active memory and all open figure windows. This is a good practice when running multiple m-

PAGE 144

129 files consecutively so that variables that already exist do not corrupt new variables created by the next m-file. The next section uses the prompt command to prompt the user for the filename of the edge image to be analyzed. The image is then displayed with the imagesc command. This is the only portion of this m-file that prompts the user for information. The next section locates the center of the image using the size command and extracts the desired, or usable, data for the NPSC calculation. The size command returns the number of rows and columns in the image. The amount of data selected depends on the size of each ROI and the numbe r of regions desired for each image. As previously described, there are 64 128 128 ROIs used for each image. Most of the calculations performed in the rest of this code require double pr ecision numbers so the usable data is converted from unsigned 16-bit integers. The next section performs the fast Fourie r transform (FFT) on each individual ROI. Since a Hanning truncation function is used the built in two-dimensional FFT command (fft2) in Matlab could not be utilized. Ther efore, a special function was developed to apply the truncation function and perform th e FFT. Before the FFT can be performed, low frequency background trends caused by the heel effect need to be removed. The functions designed to subtract the backgr ound trends and perform the two-dimensional FFT are described later in this appendix. In order to get the desired DFT from the FFT, the FFT must be divided by the number of pixe ls used in the calculation. The magnitude of the DFT is then calculated. This procedure is repeated M2 times. The next two sections average all of the DFT calcu lations and compute the NPSC.

PAGE 145

130 In order to calculate the DQEC, it is necessary to determine the one-dimensional NPSC from the two-dimensional NPSC. Frequency values are assigned to the elements in the two-dimensional NPSC, adjacent to the frequency axes based on their linear distance from the zero frequency element. The positive frequency components of the twodimensional NPSC are then extracted and binned. During the binning procedure, the command warning off MATLAB:di videByZero is used to prevent the error message that will occur when the zero freque ncy bin is averaged. Since the NPSC data used is adjacent to the zero-frequency axis, there will be no NPSC values with frequency components small enough for the first bin. Next, all NPSC values with frequency components greater than the la st one-dimensional frequency value are put in the same bin. The one-dimensional NPSC bins are then averaged and plotted for the user to see. The integral under the NPSC is then calculated up to the system Nyquist frequency (4.4 cycles/mm) using the trapz comma nd. This command uses the trapezoidal approximation to the integral and is the onl y means available for integration in Matlab without curve-f itting the NPSC data. The remainder of the code simply plots the NPSC for the user and writes the NPSC and frequency data to a tabdelimited text file using the dlmwrite command. Background Subtraction Function function output_roi = he el_effect(input_roi,N) % Subtract out the backgr ound trends from a matrix % Get a sampling of averaged data points from input_roi c = 1; r1 = 1; r2 = 11; c1 = 1; c2 = 11; for n = 1:2,

PAGE 146

131 for m = 1:2, roi = input_roi(r1:r2,c1:c2); data_points(c) = mean(mean(roi)); c = c + 1; r1 = r2 + N 21; r2 = r1 + 10; end r1 = 1; r2 = 11; c1 = c2 + N 21; c2 = c1 + 10; end x = [1 1 N N]; y = [1 N 1 N]; % Fit data xi = 1:N; yi = 1:N; [Xi,Yi] = meshgrid(xi,yi); heel = griddata(x,y,data_points,Xi,Yi,'linear'); output_roi = input _roi heel; Description of MATLAB functi on to Subtract Background Trends Functions in Matlab have a unique structur e from a standard m-file. They can be used when a specific task needs to be perf ormed and you do not want to put the code in the main m-file. They act lik e subroutines in more tradit ional programming languages. The first line of this code shows how a func tion is identified. The inputs to the function are in the parentheses an d the output is to the left of the equal sign. The two main commands in this function are meshgrid and griddata. The first section of the code gets an average value fr om the four corners of the input ROI. These average values are then assigned as the repres entative values of the four corners of the input ROI. The meshgrid command then pr oduces the coordinate s where the griddata command will fit a surface based on known values on that surface. Since the four corners of the surface are the only input s, griddata returns a plane. This fitted plane is then subtracted from the original input data and the background subtracted ROI is passed back

PAGE 147

132 to the INPSC m-file. An important point to note is that variables in functions like this only produce local variables. Therefore, variables produced will not corrupt already existing variables in the main m-file. Two-Dimensional Fast Fourier Transform Function function OUTPUT_ROI = hanning_fft2(input_roi,N) % Apply truncation function to the data array and take the 2D FFT x = 0:1:N-1; Hanning_row = (0.5-0.5*cos(2*pi*x/(N))).*1.63 3; % Normalized so that the mean % squared value is one y = x'; Hanning_column = Hanning_row'; for row = 1:N, input_roi(row:row,1:N) = input_roi(row:row,1:N).*Hanning_row; INPUT_ROI(row:row,1:N) = fft(input_roi(row:row,1:N),N,2); end for column = 1:N, INPUT_ROI(1:N,column:column) = INPUT_ROI(1:N,column:column).*Hanning_column; OUTPUT_ROI(1:N,column:column) = fft(INPUT_ROI(1:N,column:column),N,1); End Description of MATLAB Function to Perform the Fast Fourier Transform The built in Matlab function, fft2, to compute the two-dimensional FFT could not be used because of the Hanning truncation wi ndow. Therefore, the two-dimensional FFT was computed by using the fft command on each row and each column in the input ROI. This procedure is described by Brigham.15 First, each row is isolated, the Hanning window applied and then the FFT computed. This procedure is then repeated on the resulting columns. INPS C Code Verification Two known inputs are used to verify the functionality of the NPSC m-file. The first is a perfect flat field of all zeros and the second is a perfect flat field with all pixels

PAGE 148

133 having a value of 2,000. This value was chos en because it is approximately the average pixel value for a flat-field image acquired at 60 kV and 1.0 mAs. The image of all zeros returned a NPSC, both one and two-dimensional of zero. The image with a constant value of 2,000 did not return a perfect resu lt of zero but a two-dimensional NPSC integral of approximately 1-34. This is a result of the backgr ound trend removal. With all zeros, the griddata command fits a plane with a c onstant value of zero. But with the second image, the griddata command introduces an error on the order of 10-34. Given the order of magnitude of an actual NPSC, this is an acceptable error.

PAGE 149

134 APPENDIX B IMTFC MATLAB M-FILE AND DESCRIPTION This appendix lists the Matlab m-file (or code) developed to calculate the IMTFC. Following the m-file is a description of how the code functions as well as the built in Matlab functions that were used. Finally, the code is te sted with known inputs and the outputs are examined to ensure proper functionality. IMTF C MATLAB m-File % Matlab code to calculate the IMTFC on an edge image clear all; close all; % Prompt user for file name prompt = {'Enter file name with extension:'}; dlgTitle = 'File Name Input'; lineNo = 1; phantom_image_filename = char(inp utdlg(prompt,dlgTitle,lineNo)); % Read in and display image data phantom_image = dlmread(phantom_image_filename); imagesc(phantom_image); colormap('gray') axis image title('Click on top of data to be analyzed on the edge') % Select data to find the edge [top_edge_x1,top_edge_y1] = ginput(1); top_edge_x = round(top_edge_x1); top_edge_y = round(top_edge_y1); % Define usable data usable_data = phantom_image(top_edge_y:top_edge_y+511,... top_edge_x-140:top_edge_x+140); % Create binary edge image [edge_image,thresh] = edge (usable_data,'sobel');

PAGE 150

135 % Detect line theta = 0:0.1:179; [R,xp] = radon(edge_image,theta); [x t] = find(R==max(max(R))); % Determine edge angle edge_angle = theta(t)*pi/180; % Determine the location of each pixel along the rotated x' axis with the % origin at 1,1 [r c] = size(usable_data); for row = 1:r, for column = 1:c, x_prime(row,column) = (column 1) *cos(edge_angle) + (1 row)*sin(edge_angle); end end % Bin data along x' axis bin_size = 0.5; first_bin = x_prime(r,1); number_of_bins = ceil((x_prime(1, c) x_prime(r,1))/bin_size); bin_data = zeros(1,number_of_bins); bin_hits = zeros(1,number_of_bins); usable_data = double(usable_data); % Convert uint16 data to double precision for b = 1:number_of_bins, bin_value = first_bin + (b 1)*bin_size; bin_ul = bin_va lue + bin_size/2; bin_ll = bin_value bin_size/2; for row = r:-1:1, for column = 1:c, if x_prime(row,colu mn) >= bin_ll & x_prime(row,column) < bin_ul bin_data(b) = bin_data(b) + usable _data(row,column); bin_hits(b) = bin_hits(b) + 1; end end end end ESF_temp = bin_data./bin_hits; % Find max slope of ESF_temp to determine where the edge is located plot(ESF_temp) title('Click on maximum slope') [edge_x,edge_y] = ginput(1); N = 128; ESF = ESF_temp(fix(edge_x) N/2:fix(edge_x) + N/2);

PAGE 151

136 % Compute the LSF pixel_size = 1/8.8; % Pixel size in mm delta_x = pixel_size*bin_size; LSF_rect = diff(ESF)./delta_x; % Apply Hanning truncation function x = 0:1:N-1; Hanning = (0.5-0.5*cos(2*pi*x/(N))).*2; % Normalized so the mean value is 1 LSF = LSF_rect.*Hanning; % Determine the uncorrected MTFC N_MTFC = 256; MTFC_f = 0:1/(N_MTFC*delta_x):1/( 2*delta_x)-(1/(N_MTFC*delta_x)); uncorrected_MTFC = abs( fft(LSF,N_MTFC))./N_MTFC; normalized_uncorrected_MTFC = unco rrected_MTFC(1:N_MTFC/2)./ uncorrected_MTFC(1); % Apply a correction for finite element differentiation fc = 1/(2*delta_x); correction_function(1) = 1; for element = 2:128, theta = (pi*MTFC_f(element))/(2*fc); correction_function(element) = theta/sin(theta); end corrected_MTFC = uncorrected_MTFC(1 :N_MTFC/2).*correction_function; normalized_corrected_MTFC = corre cted_MTFC./corrected_MTFC(1); % Calculate integral under MTFC f_system_Nyquist = 1/(2*pixel_size); f_indices = find(MTFC_f<=f_system_Nyquist); Nyquist_index = f_indices(end); IMTFC = trapz(normalized_corr ected_MTFC(1:Nyquist_index)) *(1/(N_MTFC*delta_x)); % Plot MTFC plot(MTFC_f,normalized_unc orrected_MTFC,'b-', MTFC_f,normalize d_corrected_MTFC,'r+') axis ([0 4.4 0 1]) title({'MTFC for File: phantom_image_filename}) xlabel('1/mm') ylabel('MTFC Magnitude') text(3,0.9,{'Total Modulation = num2str(total _modulation)}) edge_angle_degrees = edge_angle*180/pi; text(3,0.8,{'Edge Angle = num2str(edge_angle_degrees)}) % Write MTFC data to a file

PAGE 152

137 dlmwrite('MTFC.txt', norma lized_corrected_MTFC, '\t') dlmwrite('MTFC_f.txt ', MTFC_f, '\t') Description of MATLAB m-file and Built in MATLAB Functions The first few sections of this m-file accomplish the same things as the INPSC mfile. The system is initialized and the user is prompted for the filename. The first numeric input is from the ginput command. This command provides a crosshair on the active figure (the edge image) and allows the user to select a data point with the mouse. The user is instructed to Click on top of data to be analyzed on the edge. This gives the reference point for the next section that ex tracts a section of da ta containing only the usable edge. The edge used for this m-file is identified in Figure B-1. The next three sections calculate the angle that the edge makes with respect to the pixel matrix. The usable data that was extracte d from the original image is processed with the edge command. This function returns a binary image with the edge identified. The Sobel method of edge detection is used in this case. Once the edge is identified, the radon command performs a Radon transform on the binary image to determine the edge angle. The projection a ngle with maximum intensity is the angle the edge makes with respect to the pixel matr ix. In reality, the radon command returns the angle a new xy coordinate system, where th e new y axis is parallel to the edge, is rotated in the counter-clockwise direction (see Figure B-2) The find command id entifies the pixel of maximum intensity in the Radon transform t hus determining the angle of rotation. Depending on the function, Matlab switches be tween degrees and radians. For example, sine and cosine are calculated using radi ans, while the radon command requires degrees as an input.

PAGE 153

138 The next task is to bin the data along the x axis. The first step in this procedure is to determine the coordinate transform (from xy to xy) of each pixe l in the usable edge data matrix. The relationship between the di stance along the x-axis and the x-axis is shown in Figure B-2. Both of the coordinate systems are centered in the first-row, firstcolumn pixel (1,1) for the calculation in the code. After the x values are determined for each pixel, they are placed into bins one-h alf the unit pixel spacing. The first bin is centered on the first-column, last -row pixel of the usable edge data. The number of bins is determined by the linear distance between this pixel and the last-column, first-row pixel. Each pixel is then checked to see which bin it belongs in and each bin averaged. Now that the super-sampled ESF has been identified, it must be truncated. The ginput command is used to en sure the center of the ESF is on the edge and 128 data points about the edge are extracted. The LSF is computed using the actual bin width and the Hanning truncation functi on is applied (the use of a truncation function made it important that the maximum of the LSF was in the center of the data vector). The LSF is zero padded to 256 data points. This length LSF with each data point representing onehalf of a pixel ensures the fr equency components in the MTFC will match up with the frequency components in the NPSC. The MTFC is then calculated using the fft command. The MTFC is then normalized to the zero -frequency component. Since these calculations were done with unit spacing between pixels, the real spacial frequency values corresponding to th e values of the MTFC need to be determined. These values are from zero to the Nyquist frequency (1/2 x) with spacing of 1/N x where N is the number of data points used calculating the FFT. The next section calculates and applies

PAGE 154

139 the correction for finite element differentia tion. The last sections perform the same functions as those described for the INPSC m-file. IMTF C Code Verification Two known inputs are used to verify the functionality of the IMTFC m-file. The first is a perfect zero-degree binary edge and the second is a three-degree binary edge. For the zero-degree edge, a bin width of one ha s to be used. If a bin width of 0.5 is used, every other bin will have no data poin ts. This input produced a perfect MTFC of one at all frequency values. A bin width of 0.5 was used with the th ree-degree edge. The resulting MTFC is shown in Figure B-3. This is not a perfect MTFC due to the discrete nature of the ESF. When the data is binned, there is a singl e bin blurring of the ESF. This can be seen in Figure B-4. This result is consistent with Samei and Flynn.2 Figure B-1. Input image for the IMTF C m-file.

PAGE 155

140 Figure B-2. Representation of the extracted usable edge data and the coordinate transform. 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 01234 Spacial Frequency (mm-1)CTF Figure B-3. MTF C of an ideal three-degree binary edge with a bin width of 0.5.

PAGE 156

141 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 01234567891011 Bin NumberESF Figure B-4. Binned data for th e three-degree binary edge.

PAGE 157

142 APPENDIX C MATLAB M-FILE FOR CONTRAST-D ETAIL PHANTOM SCORING AND DESCRIPTION This appendix lists the Matlab m-file de veloped to score a co ntrast detail phantom. Since the code for scoring both the TO.10 a nd the UF Radiology phantom are essentially the same, the code for the TO.10 phantom will be used. Following the m-file is a description of how the code f unctions as well as the built in Matlab functions that were used. MATLAB m-file for TO.10 Phantom clear all; close all; % Prompt user for file name prompt = {'Enter file name with extension:'}; dlgTitle = 'File Name Input'; lineNo = 1; phantom_image_filename = char(inp utdlg(prompt,dlgTitle,lineNo)); % Read in and display image data phantom_image = dlmread(phantom_image_filename); imagesc(phantom_image) colormap('gray') axis image title('Click on object 1 in row C') % Read in TO.10 and object position data to10data = dlmread('TO10_data.txt','\t'); position = dlmread('object _position.txt','\t'); threshold_CNR_data = dlmread('th reshold_CNR_data.txt','\t'); % % Establish orientation and size of test object % % Select and display ROI containing M1 and J1

PAGE 158

143 [large_ROI_x1,large_ROI_y1] = ginput(1); large_ROI_x = fix(large_ROI_x1); large_ROI_y = fix(large_ROI_y1); large_ROI = phantom_image(large_ROI_y 200:large_ROI_y + 200,... large_ROI_x 200:large_ROI_x + 200); imagesc(large_ROI); axis image title('Click on object 1 in row J') % Select and display small ROI for J1 [small_ROI_x1,small_ROI_y1] = ginput(1); small_ROI_x = fix(large_ROI_x 200 + small_ROI_x1 1); small_ROI_y = fix(large_ROI_y 200 + small_ROI_y1 1); small_ROI = phantom_image(small_ROI_y 20:small_ROI_y + 20,... small_ROI_x 20:small_ROI_x + 20); imagesc(small_ROI); axis image title('Click on the center of object J1') % Determine x,y coordinate for object J1 [J1_x1,J1_y1] = ginput(1); J1_x = small_ROI_x 20 + J1_x1 1; J1_y = small_ROI_y 20 + J1_y1 1; % Select and display large ROI for M1 imagesc(large_ROI); axis image title('Click on object 1 in row M') [small_ROI_x1,small_ROI_y1] = ginput(1); small_ROI_x = fix(large_ROI_x 200 + small_ROI_x1 1); small_ROI_y = fix(large_ROI_y 200 + small_ROI_y1 1); small_ROI = phantom_image(small_ROI_y 20:small_ROI_y + 20,... small_ROI_x 20:small_ROI_x + 20); imagesc(small_ROI); axis image title('Click on the center of object M1') % Determine x,y coordinate for object M1 [M1_x1,M1_y1] = ginput(1); M1_x = small_ROI_x 20 + M1_x1 1; M1_y = small_ROI_y 20 + M1_y1 1; % Determine scaling factor new_M1_J1_delta_x = M1_x J1_x; new_M1_J1_delta_y = M1_y J1_y; new_M1_to_J1_distance = sqrt(new_M1_J 1_delta_x^2 + new_M1_J1_delta_y^2);

PAGE 159

144 known_M1_J1_delta_x = to10data(1 2,1) to10data(9,1); known_M1_J1_delta_y = to10data(1 2,2) to10data(9,2); known_M1_to_J1_distance = sqrt(known_M1_J1_delta_x^2 + known_M1_J1_delta_y^2); scaling_factor = new_M1_to_J1_di stance/known_M1_to_J1_distance; % Determine the angle of rotation for the M1 to J1 line from the new image new_theta_temp = atan(abs(new_M 1_J1_delta_x)/abs(new_M1_J1_delta_y)); if new_M1_J1_delta_x >= 0 & new_M1_J1_delta_y > 0 new_theta = new_theta_temp; elseif new_M1_J1_delta_x >= 0 & new_M1_J1_delta_y < 0 new_theta = 180*pi/180 new_theta_temp; elseif new_M1_J1_delta_x >= 0 & new_M1_J1_delta_y == 0 new_theta = 90*pi/180; elseif new_M1_J1_delta_x < 0 & new_M1_J1_delta_y < 0 new_theta = 180*pi/180 + new_theta_temp; elseif new_M1_J1_delta_x < 0 & new_M1_J1_delta_y == 0 new_theta = 270*pi/180; elseif new_M1_J1_delta_x < 0 & new_M1_J1_delta_y > 0 new_theta = 360*pi/180 new_theta_temp; end % % Check for the presence of each object % % Define necessary counters object_counter = 1; % % Identify each object in phantom % for row = 1:9, object_data = 1; for object = 1:9, % % Define the object ROI and write data to a vector % % Determine the known r and alpha for the object known_J1_object_delta_x = to 10data(9,1) to10data (row,object_data); known_J1_object_delta_y = to 10data(9,2) to10data(row,object_data+1); known_M1_object_delta_x = to10da ta(12,1) to10data (row,object_data); known_M1_object_delta_y = to10data (12,2) to10data(ro w,object_data+1);

PAGE 160

145 known_J1_to_object_distance = sqrt(known_J1_object _delta_x^2 + known_J1_object_delta_y^2); known_M1_to_object_distance = sqrt(known_M1_object_delta_x^2 + known_M1_object_delta_y^2); if row == 9 & object == 1 known_alpha = 0; elseif row == 12 & object == 1 known_alpha = 0; else known_alpha = acos((known_M1_to_J1_distance^2 + known_M1_to_object_distance^2 known_J1_to_object_distance^2) /(2*known_M1_to_J1_distance*kno wn_M1_to_object_distance)); end % Determine on which side of the M1 to J1 line the object li es and the polar angle % Then calculate the x,y coordinates of the object center if position(row,object) == 0 phi = new_theta + known_alpha; new_M1_to_object_dist ance = known_M1_to_object_dist ance*scaling_factor; object_x = M1_ x new_M1_to_object_distance*sin(phi); object_y = M1_ y new_M1_to_object_distance*cos(phi); elseif position(row,object) == 1 phi = new_theta + 360*pi/180 known_alpha; new_M1_to_object_dist ance = known_M1_to_object_dist ance*scaling_factor; object_x = M1_ x new_M1_to_object_distance*sin(phi); object_y = M1_ y new_M1_to_object_distance*cos(phi); elseif position(row,object) == 9 object_x = J1_x; object_y = J1_y; elseif position(row,object) == 12 object_x = M1_x; object_y = M1_y; end % Define the object ROI known_object_size = to10data(row,3); new_object_size = fix(scaling_factor*known_object_size); ROI_start_y = fi x(object_y new_object_size/2); ROI_end_y = fi x(object_y + new _object_size/2); ROI_start_x = fi x(object_x new_object_size/2); ROI_end_x = fi x(object_x + new _object_size/2); ROI_data = phantom_image(R OI_start_y:ROI_end_y,ROI_start_x:ROI_end_x); % Make ROI data form a circle, not a square, and write data to a vector ROI_element = 1; ROI_data_vector = [];

PAGE 161

146 for ROI_row = 1:new_object_size, for RO I_column = 1:new_object_size, ROI_r = sqrt((ROI_row new_object_size/2)^2+... (ROI_column new_object_size/2)^2); if ROI_r < new_object_size/2 ceil( new_object_size*0.1) ROI_data_v ector(ROI_element) = ROI_da ta(ROI_row,ROI_column); ROI_element = ROI_element + 1; end end end % % Define the background ROIs and write data to a vector % dim_c = new_object_size/2 + 10; dim_in = 4; dim_out = 8; dim_l = new_object_size/2; for ROI_counter = 1:4, background_ROI1 = pha ntom_image(fix(object_y dim_c dim_out):fix(object_y dim_c + dim_in),... fix(obj ect_x dim_l):fix(object_x + dim_l)); background_ROI2 = pha ntom_image(fix(object_y + dim_c dim_in):fix(object_y + dim_c + dim_out),... fix(obj ect_x dim_l):fix(object_x + dim_l)); background_ROI3 = phantom _image(fix(object_y di m_l):fix(object_y + dim_l),fix(object_x dim_c dim_ou t):fix(object_x dim_c + dim_in)); background_ROI4 = phantom _image(fix(object_y di m_l):fix(object_y + dim_l),fix(object_x + dim_c dim_ in):fix(object_x + dim_c + dim_out)); end % Write background data to a vector background_data_vect or1 = mat2vec(b ackground_ROI1)'; background_data_vect or2 = mat2vec(b ackground_ROI2)'; background_data_vect or3 = mat2vec(b ackground_ROI3)'; background_data_vect or4 = mat2vec(b ackground_ROI4)'; background_data_vector = [bac kground_data_vector1 background_data_vector2 ... backgr ound_data_vector3 bac kground_data_vector4]; % % Calculate the CNR of the ROI % ROI_data_vect or = double(ROI_data_vector); background_data_vector = double(backgr ound_data_vector); mean_ROI = mean(ROI_data_vector); mean_background = mean(background_data_vector);

PAGE 162

147 stdev_background = std(background_data_vector); CNR = (mean_ROI mean_background)/stdev_background; CNR_summary(row,object) = abs(CNR); % Determine the number of elements present threshold_CNR = threshold_CNR_data(row); if abs(CNR) > threshold_CNR presence(object_counter) = 1; else presence(object_counter) = 0; end object_counter = object_counter + 1; % Increment counter for main for loop object_data = object_data + 4; end end % Determine how many objects were found and then score each row rounded_score = sum(presence); object_counter = 1; for row = 1:9, object_score = 0; counter_stop = 0; for object = 1:9, if presence( object_counter) == 1 & object < 9 object_score = object_score + 1; elseif presence(object_c ounter) == 1 & object == 9 & counter_stop == 0 row_score(row) = 9; elseif presence(object_counter) == 0 if object_score == 0 row_score(row) = 0; elseif object_score > 0 & counter_stop == 0 interpol ated_object_score = (object_score+1)-... ((CNR_s ummary(row,object)-threshold_CNR_data(row))/... (CNR_s ummary(row,object)-CNR_summary(row,object-1))*... ((object_score+1)-object_score)); row_sc ore(row) = interpolated_object_score; counter_stop = 1; end end object_counter = object_counter + 1; end end row_score(10) = sum(row_score);

PAGE 163

148 % Write CNR summary to a file dlmwrite('CNR_summary.txt', CNR_summary, '\t') dlmwrite('row_score.txt', row_score, '\t') Description of MATLAB m-file and Built in MATLAB Functions As with the other m-files alre ady described, the system is initialized and the user is prompted for a filename. There are three files th at contain data essent ial to the running of this code. The first two files, TO.10_data .txt and object_position.txt, contain data characterizing the locations and sizes of the objects in the phantom. The data in TO.10_data.txt, which was ascertained from a high exposure radiograph of the phantom, can be seen in Table C-1 and Table C-2. The object locations refer to matrix indices and the sizes are in pixels. In any subsequent im age, the objects are located through a manual registration process in which two known points in the phant om are selected. The file object_position.txt relates the relative positi on of the objects in relation to a line connecting the two selected registration points. A one is assigned to objects on the left of the line and a zero to the objects on th e right of the line. The last file, threshold_CNR_data.txt, cont ains the threshold CNRT data determined in the viewer studies. The next few sections determine the orie ntation and size of the phantom in the image. This procedure allows for 360-degree rotation and magnification of the phantom. First, a section of the image containing the tw o registration points is selected and imaged using the ginput command. The user is prompt ed to click on the first object in row C, object C1, with the mouse (the first object is the one with the highest contrast). The user is then instructed through a series of images to click on the center of the first objects in rows M and J. These are the two points used to register the imag e to the known object locations. With two known points in the new phantom image, the amount of

PAGE 164

149 magnification and the angle of rotation can be calculated. The magnification, or scaling factor, is calculated by comparing the known distance between objects M1 and J1 and the new distance between those objects. The angle of rotation is calculated by comparing the difference in the x and y values of objects M1 and J1. All possible iterations are taken into account, which allows for 360-degree rotation. This angle, in Figure C-1, is the amount of rotation in the counter-clockwise di rection of the vector connecting objects M1 to J1 with respect to the pixel matrix. Now that the orientation and magnificati on of the phantom in the new image has been determined, the objects must be located and it must be determined if a viewer would be able to detect each object. The first step for each object in turn, is calculating the distance of that object from object M1. Then, the polar angle a vector connecting object M1 and the current object would make with re spect to the pixel matrix using the image data in Table C-1 is calculat ed. This is done using the law of cosines and is depicted graphically in Figure C-1. This example is for an object loca ted to the left of the line connecting objects J1 and M1. Th e procedure is essentially th e same for an object on the right of the line except = + (360 ). In general, the known phantom configuration is superimposed on the new image (rotated by with object M1 overlapping) and expanded by the scaling factor to identify th e new coordinates of the object. Now that the coordinates of the objects center are known, the data is extracted based on the known size of the object listed in Ta ble C-2. In order to ensure that only data points within the object are used, all pi xel values that are further away from the center of the object than its radius minus 10 percent are excluded. During this exclusion procedure, the data is then written to a vect or. The final piece of information needed to

PAGE 165

150 calculate the CNR is the noise level. This is done using the ba ckground data around the object. Four slices (one on each side of the object) of background data are extracted. Each slice is four pixels wide and the diameter of the object long. This data is written to a vector using the function mat2vec, which can be seen in the next section. This function was obtained from the Internet. The author and URL are in the comment section of the function. Using the object and background data the CNR is then calculated. The last section of the main loop then writes a one to a vector if the object is present, based on the CNRTs from the observer studies, and a zero if the object is not present. Finally, the necessary counters are incremented and th e process is repeated for all objects. The last main section of the code tabulate s the final score of the number of visible objects in the phantom. The numbe r of objects that are deemed visible is calculated in two separate ways. The first is done by si mply summing the vector containing ones for visible objects and zeros for non-visible objects. The second is to linearly interpolate between the CNR of the last visible object and the one immediately after based on the CNRT for that object size. This provides a more refined estimate of the score even though it is not a whole number. The last few lines of the code write the CNR of each object as well as the score for each row to a respective file. MATLAB Function mat2vec function V=mat2vec(M); %---------------------------------------------------------------% mat2vec function convert matrix to vector. % Input : Matrix. % Output : Vector. % Tested : Matlab 5.3 % By : Eran O. Ofek May 2000 % URL : http://wise-obs.tau.ac.il/~eran/matlab.html %---------------------------------------------------------------[SizeI,SizeJ] = size(M);

PAGE 166

151 V = zeros(SizeI.*SizeJ,1); for K=1:1:SizeJ, V((K-1).*SizeI+1:K.*SizeI) = M(:,K); end Code Verification Since it was not practical to generate a perfect input image of the contrast detail phantoms, the code was verified through a seri es of manual checks. It was important to verify two main aspects of the code; object localization and object sc oring. In order to check the accuracy of the objec t localization procedure, a s ection was added to the code to write the coordinates of each objects cent er to a file. Using the program ImageJ to display an image that was evaluated by the code, the coordinates produced by the code were visually verified. One important point n eeds to be made regard ing the pixel indices displayed in ImageJ and the matrix indices us ed by Matlab. The center of the first row, first column pixel is assigned a value of (1,1 ) in Matlab and a valu e of (0,0) in ImageJ. Therefore, a correction of one for each dimension needs to be made to the values displayed in ImageJ so that they match up with the expected values produced in Matlab. This procedure was also repeat ed for the background regions. The second aspect of the code that was veri fied is the interpola tion scoring routine. This was done by comparing the CNR summar y for all of the objects, the vector identifying the presence or absence of each object and the CNRT values from the viewer studies. Using this data, the score for one image was manually calculated and compared to the score produced by the code. The two results were the same. Both of these verification procedures were repeated for the UF phantom.

PAGE 167

152 Figure C-1. Geometrical re lationship used in the lo calization of an object. M1 J1 Current Object b c a x y ac 2 a c b cos2 2 2

PAGE 168

153 Table C-1. The column and row (x,y) indici es for the objects in the TO.10 phantom. Object 1 2 3 Row x y x y x y A 1289.5 753 1498.5 926.5 1593.5 1183.5 B 1211 922 1349 1034 1411 1201 C 1153.5 1048.5 1234.5 1115.5 1271.5 1216.5 D 1143.5 677.5 1043.5 675.5 947 688.5 E 1128.5 769 1047.5 765.5 966.5 775.5 F 1119.5 846 1050.5 843.5 984.5 852.5 G 1056 915 1021.5 938.5 990.5 962 H 1082.5 949.5 1048 973 1015.5 997 J 1107.5 984 1074 1008 1040 1030.5 K 1139 1128 1116.5 1144.5 1093 1161 L 1165.5 1165.5 1143 1182 1120 1199 M 1192 1202 1169 1218 1146.5 1234.5 Object 4 5 6 Row x y x y x y A 1546 1456.5 1371 1666 1117.5 1762.5 B 1379.5 1378.5 1264 1516.5 1097 1577.5 C 1253.5 1322 1184.5 1404.5 1086.5 1439.5 D 827 730.5 742.5 778.5 666.5 841.5 E 866 812 794 853 731.5 906.5 F 900 882 840.5 916 786.5 960.5 G 942.5 995 909 1018 875 1042 H 969 1030 936 1055 903.5 1078 J 995 1064 961 1086.5 929 1110 K 1062.5 1181.5 1040 1198 1015 1217 L 1090 1220 1068 1236 1042 1254 M 1115.5 1256.5 1092.5 1271.5 1068 1290 Object 7 8 9 Row x y x y x y A 847 1715 636 1544 538 1288.5 B 924 1548.5 785 1435 721 1270 C 982.5 1424.5 898 1355.5 861 1257.5 D 589.5 940.5 544 1028.5 513.5 1125.5 E 666.5 989.5 629.5 1063.5 606.5 1144.5 F 732 1029.5 701 1091.5 682.5 1158 G 827 1076.5 794.5 1100.5 759.5 1125 H 854.5 1111.5 822.5 1135.5 787.5 1160.5 J 881 1145 849 1167 813 1193 K 982.5 1238 961 1254.5 939 1270 L 1010.5 1275.5 989.5 1291.5 966 1308 M 1037 1312 1016.5 1326.5 992 1344

PAGE 169

154 Table C-2. Diameter of objects in each row. Row Object Size A 100 B 73 C 50 D 38 E 26 F 18 G 15 H 12 J 8 K 5 L 4 M 3

PAGE 170

155 APPENDIX D MCNP INPUT FILES This appendix lists the MCNP input files used to determine the x-ray spectrum at the table top and at the image receptor. Follo wing each MCNP input file is a description of the various aspects of the i nput file (e.g. input geometry and simulated materials used). The MCNP input files in this appendix are adapted from the work of Pazik.52 MCNP Input File to Determine th e X-Ray Spectrum at the Table Top X-ray spectrum verification C 345****1*********2*********3*********4*********5*********6*********7 C C Cell Cards C Material Density Cell Definition 1 1 -0.001205 1 -2 3 -4 5 -6 $ no aluminum C 1 2 -2.69 1 -2 3 -4 5 -6 $ aluminum shielding C Meter 2 1 -0.001205 -7 C Collimators 3 0 -14 -12 13 (-8:9:-10:11:15:-13) C Cylinder of world 4 1 -0.001205 -18 19 -20 7 #1 #3 5 0 18:20:-19 C Surface Cards C Aluminum shielding 1 px -7 2 px 7 3 py -7 4 py 7 5 pz 0 6 pz 0.1956 $ HVL C Meter 7 sz -31.4325 1.27 C Collimators 8 px -0.3 9 px 0.3 10 py -0.3

PAGE 171

156 11 py 0.3 12 pz 27.19 13 pz 0.5 14 cz 9.99 15 pz 27.18 C Cylinder encircling world 18 cz 10 19 pz -32.8 20 pz 27.2 MODE P SDEF DIR=d1 POS=0 0 27.17 ERG=d2 PAR=2 VEC=0 0 -1 IMP:P 1 1 0 1 0 C Cone biasing with an angle of 0.6208 degrees C This is tight collimation around the detector C -1, Cos of Angle, 1 si1 -1 0.99993674 1 C 0, (1+costheta)/2,(1-costheta)/2 sp1 0 0.99996837 0.0000316297 sb1 0 0 1 # si2 sp2 C KeV Probability 0.000 0.000000e+000 0.001 0.000000e+000 0.002 0.000000e+000 0.003 0.000000e+000 0.004 0.000000e+000 0.005 0.000000e+000 0.006 0.000000e+000 0.007 0.000000e+000 0.008 0.000000e+000 0.009 0.000000e+000 0.010 6.755106e-005 0.011 6.187203e-003 0.012 1.089790e-001 0.013 1.351073e+000 0.014 9.360937e+000 0.015 4.791080e+001 0.016 1.600411e+002 0.017 4.536293e+002 0.018 1.000711e+003 0.019 1.673322e+003 0.020 2.742782e+003 0.021 3.870986e+003 0.022 5.412766e+003 0.023 6.835494e+003

PAGE 172

157 0.024 8.580416e+003 0.025 1.019463e+004 0.026 1.165599e+004 0.027 1.306568e+004 0.028 1.448104e+004 0.029 1.526285e+004 0.030 1.608235e+004 0.031 1.656373e+004 0.032 1.708593e+004 0.033 1.719031e+004 0.034 1.727609e+004 0.035 1.727553e+004 0.036 1.708338e+004 0.037 1.679504e+004 0.038 1.642074e+004 0.039 1.598678e+004 0.040 1.550015e+004 0.041 1.501779e+004 0.042 1.444833e+004 0.043 1.355170e+004 0.044 1.267312e+004 0.045 1.212584e+004 0.046 1.149174e+004 0.047 1.054976e+004 0.048 9.594644e+003 0.049 8.971107e+003 0.050 8.309536e+003 0.051 7.672315e+003 0.052 6.827759e+003 0.053 6.044898e+003 0.054 5.237788e+003 0.055 4.409499e+003 0.056 3.609754e+003 0.057 2.565573e+003 0.058 1.703195e+003 0.059 9.725921e+002 0.060 2.674413e+002 0.061 2.042372e+001 0.062 0.000000e+000 0.063 0.000000e+000 C Talley for Dose in MeV/g F6:P 2 C Material Cards m1 6000 -0.000124 7000 -0.755267 8000 -0.231781 18000 -0.012827 $ dry air m2 13000 1 $ aluminum nps 250000

PAGE 173

158 Description of MCNP Table Top Input File The first line of the input file provides MC NP with a title to use in the output file. The next line shows how a comm ent line is inserted in an in put file. If a capitol C is in the first column, then all the other characters on that line are considered as a comment and not used by MCNP. The comment on the se cond line shows the column number since MCNP interprets input differently depending on which column the input starts on each line. Having the column numbers shown makes it easier to ensure the input is started in the proper column. Since the font used in this text is a scalable font (i.e., a space does not take up the same horizontal distance as a C) the columns do not line up. It is recommended that the reader wishing to crea te an MCNP input file of their own do so using a non-scalable font such as Courier New. The next section of the input file contains the cell cards. A card is simply a line of input that is interpreted sepa rately by MCNP. The cell cards are used to define threedimensional regions of space to be used in the simulation. The first entry in each card is the card number and that numeric value is placed in the first column. Following each call card number is the material that cell is comprised of (defined by an integer), the density of that material and how the cell is defined. The materials used for each cell are defined at the end of the input file in such a way that the integer material numbers correspond to the integer that defines that material at the end of the input file. Each cell is defined by surfaces that are specified in the next sect ion of the input file by corresponding integer values. If the density value is negative the units of g/cm3 are assigned and if the density value is positive the units of 1024 atoms/cm3 are assigned. In order to define a threedimensional cell from a group of surfaces, the relative orientation of each surface must be established. This is accomplished by combin ing the surfaces with logical operators. For

PAGE 174

159 example, if a negative sign is put in front of a surface number, all of the space to the negative side of that surface in a Cartesian coor dinate system is considered part of that cell. This process is continued until the cell is defined. Other logical operators used in this input file are the implied positive sign which has the opposite meaning of the negative sign, the : which is the uni on operator and the # sign which is the complement operator. The next section lists the surface card s where each surface used in the cell definitions is defined. The first column indicates the surface number, the next alphanumeric entry lists the type of surface and th e last entry indicates the location of that surface in the Cartesian coordina te system. The surfaces used in this input file are plane (defined with a p where px is a plane perpendi cular to the x-axis), a cylinder defined by a c and a sphere defined by an s. The geometry defined by the surface and cell cards in this input file are depicted gra phically below in Figure D-1. The next card defines the source and is starte d with the text SDEF in the first four columns. This is followed by what can be considered subcards. The d1 in the DIR subcard specifies the direction bias of the sour ce. In this case it is a cone bias and the size of the cone is specified in the si1 and the sp1 cards. The POS sub card defines the position of the source in Cartesian coordinates. Th e d2 in the ERG subcard specifies the energy distribution of the source as detailed in th e si2 and sp2 cards. The si2 and the sp2 cards are input in a column format. This is acco mplished by placing the # sign in the first column of the card. The last two subcards for the source definition are the PAR and the VEC subcards. The PAR subcard defines the particle type (2 for photons) and the VEC subcard gives a reference vector for the direction of the particles emitted from the source.

PAGE 175

160 The next card (IMP) specifies the importanc e of the particles in each cell. In this case all cells are given an importance of 1 (the highest level) except the collimators and the space outside of the problem geometry. These cells are given an importance of 0 which immediately kills the particle when it en ters those cells. The si and sp cards that were explained above follow on the next seve ral lines and includes the energy spectrum. Immediately following the energy spectrum is a ca rd that specifies an F6 tally. Tallies are the way in which MCNP keeps track of the part icles fates. This F6 tally defines a track length estimate of energy deposition fo r photons in cell two (the detector). The next few cards are the material cards They are defined by an m in the first column and an integer that corresponds to the number used to define the materials in the cell cards. The first entry in a cell card defi nes the element and takes the form ZZZAAA. For photons, the atomic number (A) is zero fo r all elements. The s econd entry defines the relative composition of the elemen ts that make up each material. If a + sign is used the entries represent atomic fractions and if a sign is used the entries represent weight fractions. The final card is an nps card and sp ecifies the number of particles used in the simulation. MCNP Input File to Determine the X-Ray Spectrum at the Image Receptor X-ray spectrum verification C 345****1*********2*********3*********4*********5*********6*********7 C C Cell Cards C Material Density Cell Definition 1 1 -0.001205 1 -2 3 -4 5 -6 C Acrylic 2 3 -1.19 -14 22 -23 C Collimators 3 0 -14 -12 13 (-8:9:-10:11:15:-13) C Cylinder of world 4 1 -0.001205 -18 19 -20 #1 #2 #3 #6 #7 5 0 18:20:-19

PAGE 176

161 C Grid 6 4 -11.33 -14 25 -26 $ with grid C 6 1 -0.001205 -14 25 -26 $ without grid C Imaging plate 7 1 -0.000001 27 -28 29 -30 -31 32 C Surface Cards C Aluminum shielding 1 px -8 2 px 8 3 py -8 4 py 8 5 pz 0 6 pz 0.2733 $ HVL is changed to air for this problem C Collimators 8 px -3.9375 9 px 3.9375 10 py -3.15 11 py 3.15 12 pz 27.19 13 pz 0.5 14 cz 19.3 15 pz 27.18 C Cylinder encircling world 18 cz 19.31 19 pz -74.64 20 pz 27.2 C Acrylic blocks 22 pz -40.64 23 pz -35.56 C Grid 25 pz -69.6579942 26 pz -69.65 C Imaging Plate 27 px -15 28 px 15 29 py -12 30 py 12 31 pz -74.43 32 pz -74.63 MODE P SDEF DIR=d1 POS=0 0 27.17 ERG=d2 PAR=2 VEC=0 0 -1 IMP:P 1 1 0 1 0 1 1 C Cone biasing with an angle of 8.39835 degrees C -1, Cos of Angle, 1

PAGE 177

162 si1 -1 0.98927653 1 C 0, (1+costheta)/2,(1-costheta)/2 sp1 0 0.99463826 0.00536174 sb1 0 0 1 # si2 sp2 C KeV Probability 0.000 0.000000e+000 0.001 0.000000e+000 0.002 0.000000e+000 0.003 0.000000e+000 0.004 0.000000e+000 0.005 0.000000e+000 0.006 0.000000e+000 0.007 0.000000e+000 0.008 0.000000e+000 0.009 0.000000e+000 0.010 1.667037e-007 0.011 6.702871e-005 0.012 3.215110e-003 0.013 8.050436e-002 0.014 9.994548e-001 0.015 7.633517e+000 0.016 3.520179e+001 0.017 1.272994e+002 0.018 3.409927e+002 0.019 6.563028e+002 0.020 1.237942e+003 0.021 1.916401e+003 0.022 2.939904e+003 0.023 3.951859e+003 0.024 5.278006e+003 0.025 6.599543e+003 0.026 7.828999e+003 0.027 9.105412e+003 0.028 1.042640e+004 0.029 1.126040e+004 0.030 1.215664e+004 0.031 1.276260e+004 0.032 1.341548e+004 0.033 1.367764e+004 0.034 1.392818e+004 0.035 1.411292e+004 0.036 1.408450e+004 0.037 1.397117e+004 0.038 1.375997e+004 0.039 1.348599e+004

PAGE 178

163 0.040 1.316100e+004 0.041 1.280902e+004 0.042 1.236704e+004 0.043 1.163411e+004 0.044 1.090961e+004 0.045 1.046537e+004 0.046 9.917670e+003 0.047 9.092364e+003 0.048 8.248741e+003 0.049 7.703012e+003 0.050 7.118414e+003 0.051 6.561164e+003 0.052 5.800405e+003 0.053 5.088577e+003 0.054 4.350744e+003 0.055 3.597729e+003 0.056 2.868704e+003 0.057 1.945594e+003 0.058 1.198114e+003 0.059 5.977838e+002 0.060 1.499110e+002 0.061 1.102908e+001 0.062 0.000000e+000 0.063 0.000000e+000 C Tally energy spectrum in cell 8 F4:P 7 E4 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.009 0.010 0.011 0.012 0.013 0.014 0.015 0.016 0.017 0.018 0.019 0.020 0.021 0.022 0.023 0.024 0.025 0.026 0.027 0.028 0.029 0.030 0.031 0.032 0.033 0.034 0.035 0.036 0.037 0.038 0.039 0.040 0.041 0.042 0.043 0.044 0.045 0.046 0.047 0.048 0.049 0.050 0.051 0.052 0.053 0.054 0.055 0.056 0.057 0.058 0.059 0.060 0.061 0.062 0.063 0.064 0.065 0.066 0.067 0.068 0.069 0.070 0.071 0.072 0.073 0.074 0.075 0.076 0.077 0.078 0.079 0.080 0.081 0.082 0.083 C Material Cards m1 6000 -0.000124 7000 -0.755267 8000 -0.231781 18000 -0.012827 $ dry air m2 13000 1 $ aluminum m3 1000 -0.080548 6000 -0.599816 8000 -0.319636 $ acrylic m4 82000 1 $ Pb nps 100000000 Description of Image Receptor Input File This input file contains essentially the sa me elements as the table top input file. Some materials have been added and they can be seen by the graphica l representation in

PAGE 179

164 Figure D-2. The only significant di fference in the input files is the tally used. In this case an F4 tally is used. This tally provides a tr ack length estimate of the flux in cell seven (the image receptor). Figure D-1. A geometrical depi ction of the cells defined in the table top input file. A cross section along the z-axis is shown.

PAGE 180

165 Figure D-2. A geometrical depic tion of the cells defined in th e image receptor input file. A cross section along the z-axis is shown.

PAGE 181

166 APPENDIX E RESULTS OF THE POSTHOC STATISTICAL TESTS Table E-1. Post-hoc INPSC mean comparisons for variati ons with current-time product. Acquisition mAs p-value Base mAs Compared mAs Tamhanes T2 Dunnetts T3 Games-Howell 0.4 0.80* 1.00* 1.20* 2.00* 3.20* .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 0.8 0.80* 1.00* 1.20* 2.00* 3.20* .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 1.0 0.80* 1.00* 1.20* 2.00* 3.20* .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 1.2 0.80* 1.00* 1.20* 2.00* 3.20* .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 2.0 0.80* 1.00* 1.20* 2.00* 3.20* .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 3.2 0.80* 1.00* 1.20* 2.00* 3.20* .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 *Indicates a significant mean differ ence for all three tests at the = 0.05 level.

PAGE 182

167 Table E-2. Post-hoc INPSC mean comparisons for variati ons with peak-tube potential. Acquisition kVp p-value Base kVp Compared kVp Tamhanes T2 Dunnetts T3 Games-Howell 50 60* 70* 80* .000 .001 .003 .000 .001 .002 .000 .000 .002 60 50* 70 80 .000 1.000 .998 .000 .999 .994 .000 .980 .958 70 50* 60 80 .001 1.000 1.000 .001 .999 1.000 .000 .980 .999 80 50* 60 70 .003 .998 1.000 .002 .994 1.000 .002 .958 .999 *Indicates a significant mean differ ence for all three tests at the = 0.05 level. Table E-3. Post-hoc INPSC mean comparisons for variat ions with pro cessing option. Processing Option p-value Base Compared Tamhanes T2 Dunnetts T3 Games-Howell System Diagnosis/Flat Field Chest PA* Hand AP* .000 .000 .000 .000 .000 .000 Chest/Chest PA Flat Field* Hand AP* .000 .000 .000 .000 .000 .000 Upper Extremity/Hand AP Flat Field* Chest PA* .000 .000 .000 .000 .000 .000 *Indicates a significant mean differ ence for all three tests at the = 0.05 level.

PAGE 183

168 Table E-4. Post-hoc scan IMTFC mean comparisons for variations with current-time product. Acquisition mAs p-value Base mAs Compared mAs Tamhanes T2 Dunnetts T3 Games-Howell 0.4 0.80 1.00 1.20 2.00 3.20 .785 .419 .557 .239 .432 .606 .252 .342 .155 .253 .460 .189 .259 .114 .190 0.8 0.40 1.00 1.20 2.00 3.20 .785 .996 1.000 .876 .995 .606 .960 1.000 .735 .949 .460 .853 .982 .570 .840 1.0 0.40 0.80 1.20 2.00 3.20 .419 .996 .999 .999 1.000 .252 .960 .993 .986 1.000 .189 .853 .933 .910 1.000 1.2 0.40 0.80 1.00 2.00 3.20 .557 1.000 .999 .918 .998 .342 1.000 .993 .757 .984 .259 .982 .933 .604 .900 2.0 0.40 0.80 1.00 1.20 3.20 .239 .876 .999 .918 .998 .155 .735 .986 .757 .155 .114 .570 .910 .604 .878 3.2 0.40 0.80 1.00 1.20 2.00 .432 .995 1.000 .998 .998 .735 .986 .757 .971 .253 .190 .840 1.000 .900 .878 *Indicates a significant mean differ ence for all three tests at the = 0.05 level.

PAGE 184

169 Table E-5. Post-hoc subscan IMTFC mean comparisons for variations with current-time product. Acquisition mAs p-value Base mAs Compared mAs Tamhanes T2 Dunnetts T3 Games-Howell 0.4 0.80 1.00 1.20* 2.00* 3.20* .345 .141 .077 .048 .049 .194 .094 .047 .030 .029 .146 .069 .035 .022 .022 0.8 0.40 1.00 1.20 2.00* 3.20* .345 .976 .163 .053 .010 .194 .866 .104 .034 .007 .146 .726 .076 .025 .005 1.0 0.40 0.80 1.20 2.00 3.20 .141 .976 .996 .796 .619 .094 .866 .970 .630 .440 .069 .726 .867 .478 .328 1.2 0.40* 0.80 1.00 2.00 3.20 .077 .163 .996 .999 .967 .047 .104 .970 .989 .888 .035 .076 .867 .917 .734 2.0 0.40* 0.80* 1.00 1.20 3.20 .048 .053 .796 .999 1.000 .030 .034 .630 .989 1.000 .022 .025 .478 .917 .999 3.2 0.40* 0.80* 1.00 1.20 2.00 .049 .010 .619 .967 1.000 .029 .007 .440 .888 1.000 .022 .005 .328 .734 .999 *Indicates a significant mean differ ence for all three tests at the = 0.05 level.

PAGE 185

170 Table E-6. Post-hoc scan IMTFC mean comparisons for variations with peak-tube potential. Acquisition kVp p-value Base kVp Compared kVp Tukeys HSD Bonferroni Test LSD 50 60* 70* 80* .000 .000 .000 .000 .000 .000 .000 .000 .000 60 50* 70* 80* .000 .000 .000 .000 .000 .000 .000 .000 .000 70 50* 60* 80* .000 .000 .001 .000 .000 .001 .000 .000 .000 80 50* 60* 70* .000 .000 .001 .000 .000 .001 .000 .000 .000 *Indicates a significant mean differ ence for all three tests at the = 0.05 level. Table E-7. Post-hoc subscan IMTFC mean comparisons for va riations with peak-tube potential. Acquisition kVp p-value Base kVp Compared kVp Tukeys HSD Bonferroni Test LSD 50 60* 70* 80* .000 .000 .000 .000 .000 .000 .000 .000 .000 60 50* 70* 80* .000 .000 .000 .000 .000 .000 .000 .000 .000 70 50* 60* 80* .000 .000 .000 .000 .000 .000 .000 .000 .000 80 50* 60* 70* .000 .000 .000 .000 .000 .000 .000 .000 .000

PAGE 186

171 Table E-8. Post-hoc IMTFC mean comparisons for variati ons with processing option in the scan direction. Processing Option p-value Base Compared Tamhanes T2 Dunnetts T3 Games-Howell System Diagnosis/Full Range Chest PA* Hand AP* .014 .000 .010 .000 .009 .000 Chest/Chest PA Full Range* Hand AP* .014 .007 .010 .005 .009 .005 Upper Extremity/Hand AP Full Range* Chest PA* .000 .007 .000 .005 .000 .005 *Indicates a significant mean differ ence for all three tests at the = 0.05 level. Table E-9. Post-hoc IMTFC mean comparisons for variati ons with processing option in the subscan direction. Processing Option p-value Base Compared Tukeys HSD Bonferroni LSD System Diagnosis/ Full Range Chest PA* Hand AP* .000 .000 .000 .000 .000 .000 Chest/Chest PA Full Range* Hand AP* .000 .000 .000 .000 .000 .000 Upper Extremity/Hand AP Full Range* Chest PA* .000 .000 .000 .000 .000 .000 *Indicates a significant mean differ ence for all three tests at the = 0.05 level.

PAGE 187

172 Table E-10. Post -hoc scan IDQEC mean comparisons for vari ations with current-time product. Acquisition mAs p-value Base mAs Compared mAs Tukeys HSD Bonferroni Test LSD 0.4 0.80* 1.00* 1.20* 2.00* 3.20* .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 0.8 0.40* 1.00* 1.20* 2.00* 3.20* .000 .049 .000 .000 .000 .000 .073 .000 .000 .000 .000 .005 .000 .000 .000 1.0 0.40* 0.80* 1.20* 2.00* 3.20* .000 .049 .047 .000 .000 .000 .073 .069 .000 .000 .000 .005 .005 .000 .000 1.2 0.40* 0.80* 1.00* 2.00* 3.20* .000 .000 .047 .014 .000 .000 .000 .069 .019 .000 .000 .000 .005 .001 .000 2.0 0.40* 0.80* 1.00* 1.20* 3.20 .000 .000 .000 .014 .003 .000 .000 .000 .019 .004 .000 .000 .000 .001 .000 3.2 0.40* 0.80* 1.00* 1.20* 2.00 .000 .000 .000 .000 .003 .000 .000 .000 .000 .004 .000 .000 .000 .000 .000 *Indicates a significant mean differ ence for all three tests at the = 0.05 level.

PAGE 188

173 Table E-11. Post-hoc subscan IDQEC mean comparisons for variations with current-time product. Acquisition mAs p-value Base mAs Compared mAs Tukeys HSD Bonferroni Test LSD 0.4 0.80* 1.00* 1.20* 2.00* 3.20* .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 0.8 0.40* 1.00* 1.20* 2.00* 3.20* .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 1.0 0.40* 0.80* 1.20* 2.00* 3.20* .000 .000 .022 .000 .000 .000 .000 .031 .000 .000 .000 .000 .002 .000 .000 1.2 0.40* 0.80* 1.00* 2.00* 3.20* .000 .000 .022 .000 .000 .000 .000 .031 .000 .000 .000 .000 .002 .000 .000 2.0 0.40* 0.80* 1.00* 1.20* 3.20* .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 3.2 0.40* 0.80* 1.00* 1.20* 2.00* .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 *Indicates a significant mean differ ence for all three tests at the = 0.05 level.

PAGE 189

174 Table E-12. Post -hoc scan IDQEC mean comparisons for variations with peak-tube potential. Acquisition kVp p-value Base kVp Compared kVp Tukeys HSD Bonferroni Test LSD 50 60* 70* 80* .050 .000 .000 .068 .000 .000 .011 .000 .000 60 50* 70* 80* .050 .000 .000 .068 .000 .000 .011 .000 .000 70 50* 60* 80* .000 .000 .000 .000 .000 .000 .000 .000 .000 80 50* 60* 70* .000 .000 .000 .000 .000 .000 .000 .000 .000 *Indicates a significant mean differ ence for all three tests at the = 0.05 level. Table E-13. Post-hoc subscan IDQEC mean comparisons for variations with peak-tube potential. Acquisition kVp p-value Base kVp Compared kVp Tukeys HSD Bonferroni Test LSD 50 60 70 80* .994 .121 .000 1.000 .179 .000 .801 .030 .000 60 50 70 80* .994 .186 .000 1.000 .296 .000 .801 .049 .000 70 50 60 80* .121 .186 .000 .179 .296 .000 .030 .049 .000 80 50* 60* 70* .000 .000 .000 .000 .000 .000 .000 .000 .000

PAGE 190

175 Table E-14. Post-hoc IDQEC mean comparisons for variatio ns with processing option in the scan direction. Processing Option p-value Base Compared Tamhanes T2 Dunnetts T3 Games-Howell System Diagnosis/ Full Range Chest PA* Hand AP .030 .992 .021 .986 .019 .956 Chest/Chest PA Full Range* Hand AP* .030 .008 .021 .006 .019 .005 Upper Extremity/Hand AP Full Range Chest PA* .992 .008 .986 .006 .956 .005 *Indicates a significant mean differ ence for all three tests at the = 0.05 level. Table E-15. Post-hoc IDQEC mean comparisons for variatio ns with processing option in the subscan direction. Processing Option p-value Base Compared Tamhanes T2 Dunnetts T3 Games-Howell System Diagnosis/ Full Range Chest PA* Hand AP .026 .380 .019 .298 .016 .259 Chest/Chest PA Full Range* Hand AP* .026 .011 .019 .009 .016 .008 Upper Extremity/Hand AP Full Range Chest PA* .380 .011 .298 .009 .259 .008 *Indicates a significant mean differ ence for all three tests at the = 0.05 level.

PAGE 191

176 Table E-16. Post-hoc CDSC mean comparisons for variatio ns with current-time product for the TO.10 phantom. Acquisition mAs p-value Base mAs Compared mAs Tukeys HSD Bonferroni Test LSD 0.4 0.80* 1.00* 1.20* 2.00* 3.20* .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 0.8 0.40* 1.00* 1.20* 2.00* 3.20* .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 1.0 0.40* 0.80* 1.20* 2.00* 3.20* .000 .000 .003 .000 .000 .000 .000 .003 .000 .000 .000 .000 .000 .000 .000 1.2 0.40* 0.80* 1.00* 2.00* 3.20* .000 .000 .003 .000 .000 .000 .000 .003 .000 .000 .000 .000 .000 .000 .000 2.0 0.40* 0.80* 1.00* 1.20* 3.20* .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 3.2 0.40* 0.80* 1.00* 1.20* 2.00* .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 *Indicates a significant mean differ ence for all three tests at the = 0.05 level.

PAGE 192

177 Table E-17. Post-hoc CDSC mean comparisons for variatio ns with current-time product for the UF Radiology phantom. Acquisition mAs p-value Base mAs Compared mAs Tukeys HSD Bonferroni Test LSD 0.8 0.40* 1.00* 1.20* 2.00* 3.20* .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 1.0 0.40* 0.80* 1.20* 2.00* 3.20* .000 .000 .001 .000 .000 .000 .000 .003 .000 .000 .000 .000 .001 .000 .000 1.2 0.40* 0.80* 1.00* 2.00* 3.20* .000 .000 .001 .000 .000 .000 .000 .003 .000 .000 .000 .000 .001 .000 .000 2.0 0.40* 0.80* 1.00* 1.20* 3.20* .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 *Indicates a significant mean differ ence for all three tests at the = 0.05 level.

PAGE 193

178 Table E-18. Post-hoc CDSC mean comparisons for variatio ns with processing option for the TO.10 phantom. Processing Option p-value Base Compared Tukeys HSD Bonferroni Test LSD System Diagnosis/ Full Range Chest PA* Hand AP* .004 .000 .004 .000 .000 .000 Chest/Chest PA Full Range* Hand AP* .004 .000 .004 .000 .000 .000 Upper Extremity/Hand AP Full Range* Chest PA* .000 .000 .000 .000 .000 .000 *Indicates a significant mean differ ence for all three tests at the = 0.05 level. Table E-19. Post-hoc CDSC mean comparisons for variatio ns with processing option for the UF Radiology phantom. Processing Option p-value Base Compared Tukeys HSD Bonferroni Test LSD System Diagnosis/ Full Range Chest PA* Hand AP .021 .000 .025 .000 .008 .000 Chest/Chest PA Full Range* Hand AP* .021 .000 .025 .000 .008 .000 Upper Extremity/Hand AP Full Range Chest PA* .000 .000 .000 .000 .000 .000 *Indicates a significant mean differ ence for all three tests at the = 0.05 level.

PAGE 194

179 LIST OF REFERENCES 1K. M. Hintenlang, J. L. Williams, and D. E. Hintenlang, "A survey of radiation dose associated with pediatric plain-film ch est X-ray examinations," Pediatr Radiol 32, 771-7 (2002). 2M. J. Flynn and E. Samei, "Experimental comparison of noise and resolution for 2k and 4k storage phosphor radiography systems," Med Phys 26, 1612-23 (1999). 3K. A. Johnson, Quantifying computed radiography (CR) and digital radiography (DR) image quality and patient dose for pediatric radiology Doctoral Dissertation, Department of Nuclear and Ra diological Engineering, Univ ersity of Florida, (2003). 4J. T. Dobbins, 3rd, D. L. Ergun, L. Rutz, D. A. Hinshaw, H. Blume, and D. C. Clark, "DQE(f) of four generations of computed radiography acquisition devices," Med Phys 22, 1581-93 (1995). 5K. A. Fetterly and N. J. Hangiandreou, "Ima ge quality evaluation of a desktop computed radiography system," Med Phys 27, 2669-79 (2000). 6R. Padgett and C. J. Kotre, "Assessment of the effects of pixel loss on image quality in direct digital radiogra phy," Phys Med Biol 49, 977-86 (2004). 7T. Ideguchi, Y. Higashida, K. Himuro, M. Ohki, S. Nakamura, A. Yoshida, R. Takagi, H. Hatano, R. Kuwahara, M. Toyonaga, I. Ta naka, and F. Toyofuku, "[Full-field digital mammography with amorphous silicon-based fl atpanel detector: physical imaging characteristics and signal detection]," Nippon Hoshasen Gijutsu Gakkai Zasshi 60, 399405 (2004). 8H. G. Chotas and C. E. Ravin, "Digital ch est radiography with a so lid-state flat-panel xray detector: contrast-detail evaluation with processed images printed on film hard copy," Radiology 218, 679-82 (2001). 9T. Katou, K. Ogawa, Y. Umezu, E. Kinoshita E. Shinkai, M. Ohki, F. Toyofuku, and Y. Higashida, "[Low contrast detectability of a new CR system with two-sided reading]," Nippon Hoshasen Gijutsu Gakkai Zasshi 59, 410-5 (2003). 10J. T. Bushberg, The essential physics of medical imaging (Lippincott Williams & Wilkins, Philadelphia, 2002). 11E. L. Siegel and R. M. Kolodner, Filmless radiology (Springer, New York, 1999).

PAGE 195

180 12S. Lehar, in An intuitive explanation of Fourier theory (Boston University, http://cnsalumni.bu.edu/~slehar/fourier/fourier.html#basic July, 2004) 13R. B. Blackman, J. W. Tukey, and Ka rreman Mathematics Research Collection., The measurement of power spectra, from the point of view of communications engineering (Dover Publications, New York, 1959). 14H. H. Barrett, and W. Swindell, Radiological imaging: the theory of image formation, detection, and processing (Academic Press, New York, 1981). 15E. O. Brigham, The fast Fourier transform and its applications (Prentice Hall, Englewood Cliffs, N.J., 1988). 16E. O. Brigham, The fast Fourier transform (Prentice-Hall, Englewood Cliffs, N.J.,, 1974). 17C. D. Bradford, W. W. Peppler and J. T. Dobbins, 3rd, "Performance characteristics of a Kodak computed radiography system," Med Phys 26, 27-37 (1999). 18K. A. Fetterly, N. J. Hangi andreou, B. A. Schueler, and E. R. Ritenour, "Measurement of the presampled two-dimensional modula tion transfer function of digital imaging systems," Med Phys 29, 913-21 (2002). 19I. A. Cunningham and A. Fenster, "A method for modulation transfer function determination from edge profiles with correction for finite-element differentiation," Med Phys 14, 533-7 (1987). 20E. Samei, M. J. Flynn, and D. A. Reima nn, "A method for measuring the presampled MTF of digital radiographic systems us ing an edge test device," Med Phys 25, 102-13 (1998). 21H. Fujita, D. Tsai, T. Itoh, K. Doi, J. Morishita, K. Ueda, and A. Ohtsuka, "A simple method for determining the modulation transf er function in digital radiography," IEEE Transactions on Medical Imaging Vol. 11, (1992). 22C. R. Crawford, "Reprojection using a parallel backprojector," Med Phys 13, 480-3 (1986). 23J. H. Launders, S. M. Kengyelics, and A. R. Cowen, "A comprehensive physical image quality evaluation of a selenium based digital x-ray imaging system for thorax radiography," Med Phys 25, 986-97 (1998). 24F. H. Attix, Introduction to radiological physics and radiation dosimetry (Wiley, New York, 1986). 25W. R. Hendee, G. S. Ritenour, and E. R. Ritenour, Medical imaging physics (Wiley, New York Chichester, 2002).

PAGE 196

181 26R. F. Wagner, and D. G. Brown, "Unified SNR analysis of medical imaging systems," Phys Med Biol 30, 489-515 (1985). 27M. Arreola, Objective assessment of image quality in conventional fluoroscopy and the low-contrast detectability of obj ects in carbon dioxide angiography Doctoral Dissertation, Department of Nuclear and Radiological En gineering, University of Florida, (1993). 28R. Aufrichtig, and P. Xue, "Dose efficiency and low-contrast detectability of an amorphous silicon x-ray detector for digital radiography," Phys Med Biol 45, 2653-69 (2000). 29Z. F. Lu, E. L. Nickoloff, J. C. So, and A. K. Dutta, "Comparison of computed radiography and film/screen combination using a contrast-detail phantom," J Appl Clin Med Phys 4, 91-8 (2003). 30K. W. Brooks, J. H. Trueblood, K. J. Kearfo tt, and D. T. Lawton, "Automated analysis of the American College of Radiology ma mmographic accreditation phantom images," Med Phys 24, 709-23 (1997). 31A. D. Castellano Smith, I. A. Castellano Smith, and D. R. Dance, "Objective assessment of phantom image quality in ma mmography: a feasibility study," Br J Radiol 71, 48-58 (1998). 32A. L. Kwan, L. J. Filipow, and L. H. Le, "Automatic quantitative low contrast analysis of digital chest phantom radiographs," Med Phys 30, 312-20 (2003). 33M. Cristy, "Mathematical phantoms representi ng children of various ages for use in estimates of internal dose," Oa k Ridge National Laboratory Report ORNL/NUREG/TM-367), Oak Ridge, TN, (1980) 34M. W. Bower, A physical anthropomorphic phantom of a one-year-old child with realtime dosimetry Doctoral Dissertation, Department of Nucl ear and Radiological Engineering, University of Florida, (1997). 35L. G. Bouchet, W. E. Bolch, D. A. Weber, H. L. Atkins, and J. W. Poston, Sr., "A revised dosimetric model of the a dult head and brain," J Nucl Med 37, 1226-36 (1996). 36D. J. Gladstone, X. Q. Lu, J. L. Humm, H. F. Bowman, and L. M. Chin, "A miniature MOSFET radiation dosimeter probe," Med Phys 21, 1721-8 (1994). 37M. J. Butson, A. Rozenfeld, J. N. Mathur, M. Carolan, T. P. Wong, and P. E. Metcalfe, "A new radiotherapy surface dose detector:the MOSFET," Med Phys 23, 655-8 (1996). 38P. Francescon, S. Cora, C. Cavedon, P. Scalchi, S. Reccanello, and F. Colombo, "Use of a new type of radiochromic film, a ne w parallel-plate micro-chamber, MOSFETs, and TLD 800 microcubes in the dosim etry of small beams," Med Phys 25, 503-11 (1998).

PAGE 197

182 39R. Ramani, S. Russell, and P. O'Brien, "Clinical dosimetry using MOSFETs," Int J Radiat Oncol Biol Phys 37, 959-64 (1997). 40M. Soubra, J. Cygler, and G. Mackay, "E valuation of a dual bias dual metal oxidesilicon semiconductor field effect transistor detector as radiation dosimeter," Med Phys 21, 567-72 (1994). 41B. V. Zeghbroeck, in Principles of Semiconductor Devices (University of Colorado at Boulder, http://ece-www.colorado.edu/~bart/book/book/title.htm May, 2002) 42M. W. Bower and D. E. Hintenlang, "The characterization of a commercial MOSFET dosimeter system for use in di agnostic x ray," Health Phys 75, 197-204 (1998). 43B. D. Pomije, C. H. Huh, M. A. Tressl er, D. E. Hintenlang, and W. E. Bolch, "Comparison of angular fr ee-in-air and tissue-equi valent phantom response measurements in p-MOSFET dosimeters," Health Phys 80, 497-505 (2001). 44International Commission on Radiological Protection., 1990 recommendations of the International Commission Radiological Prot ection: adopted by the Commission in November 1990 (Published for the Commission by Pergamon, Oxford; New York, 1991). 45E. Storm, "Calculated bremsstrahlung spectra from thick tungsten targets," Phys Rev A 5, 2328-2338 (1971). 46J. M. Boone and J. A. Seibert, "An accu rate method for comput er-generating tungsten anode x-ray spectra from 30 to 140 kV," Med Phys 24, 1661-70 (1997). 47R. Aufrichtig, "Comparison of low contrast detectability between a digital amorphous silicon and a screen-film based imaging sy stem for thoracic ra diography," Med Phys 26, 1349-58 (1999). 48X. J. Rong, C. C. Shaw, X. Liu, M. R. Lemacks, and S. K. Thompson, "Comparison of an amorphous silicon/cesium iodide flat-p anel digital chest ra diography system with screen/film and computed ra diography systems--a contra st-detail phantom study," Med Phys 28, 2328-35 (2001). 49A. R. Owen, O. F. Clarke, N. J. Coleman, D. M. Craven, S. McArdle, and G. A. Hay, Leeds x-ray test objects instruction manual (University of Leeds, Academic Unit of Medical Physics, Leeds, 1992). 50K. A. Fetterly and N. J. Hangiandreou, "Effects of x-ray spec tra on the DQE of a computed radiography system," Med Phys 28, 241-9 (2001). 51L. N. Rill, Improvement of the clinical use of computed radiography for mobile chest imaging : image quality and patient dose. Doctoral Dissertation, Department of Nuclear and Radiological Engineering, University of Florida, (2001).

PAGE 198

183 52F. D. Pazik, Organ doses received during co mmon newborn radiographic and fluoroscopic procedures Master's Thesis, Department of Nuclear and Radiological Engineering, University of Florida, (2003).

PAGE 199

184 BIOGRAPHICAL SKETCH Christopher David Pitcher was born in Cincinnati, Ohio, on April 4, 1971, to Albert Wayne and Susan Davis Pitcher. He attended hi gh school in Saint Petersburg, Florida, at Northeast High School. He received a bachelo rs degree, with high honors, in nuclear engineering from the College of Engineering of the University of Florida in 1994. As a Department of Energys Oak Ridge Institute of Science and Education fellow he received a masters degree in health physics from the Co llege of Engineering of the University of Florida in 1995. Following completion of his masters degree, he was commissioned as a First Lieutenant in the United States Army Medical Department. After achieving the rank of Captain, he was selected to pursue a doctoral degree through the Armys Long Term Health Education and Training Program. He received his Doctor of Philosophy degree in December of 2004 in medical physics from the College of Engineering of the University of Florida.