Citation
Kinematic Analysis and Automatic Segmentation of the Menisci

Material Information

Title:
Kinematic Analysis and Automatic Segmentation of the Menisci
Creator:
Park, Brian H
Place of Publication:
[Gainesville, Fla.]
Florida
Publisher:
University of Florida
Publication Date:
Language:
english
Physical Description:
1 online resource (72 p.)

Thesis/Dissertation Information

Degree:
Doctorate ( Ph.D.)
Degree Grantor:
University of Florida
Degree Disciplines:
Mechanical Engineering
Mechanical and Aerospace Engineering
Committee Chair:
BANKS,SCOTT ARTHUR
Committee Co-Chair:
FREGLY,BENJAMIN J
Committee Members:
SARNTINORANONT,MALISA
POZZI,ANTONIO
Graduation Date:
12/17/2016

Subjects

Subjects / Keywords:
Dogs ( jstor )
Imaging ( jstor )
Kinematics ( jstor )
Knee joint ( jstor )
Knees ( jstor )
Magnetic resonance imaging ( jstor )
Menisci ( jstor )
Statistical models ( jstor )
Three dimensional modeling ( jstor )
Tibia ( jstor )
Mechanical and Aerospace Engineering -- Dissertations, Academic -- UF
automatic-segmentation -- dog -- kinematics -- knee -- menisus -- mri
Genre:
bibliography ( marcgt )
theses ( marcgt )
government publication (state, provincial, terriorial, dependent) ( marcgt )
born-digital ( sobekcm )
Electronic Thesis or Dissertation
Mechanical Engineering thesis, Ph.D.

Notes

Abstract:
The menisci of the knee are crescent-shaped structures that contribute to load distribution and stability of the tibiofemoral joint. The kinematics of menisci are of particular interest, but there remains a lack of robust evidence regarding meniscal translation in the healthy knee for comparative purposes. One of the greatest challenges in doing a study of the menisci is the lack of visibility and contrast between neighboring anatomies in the MRI. The purpose of this study is to set up a meniscal study framework by investigating the kinematics of the meniscus in dogs and developing an automatic segmentation algorithm to guide future studies. This information may be valuable because the dog has been extensively used as an experimental model to study meniscal treatments. In this study, menisci from 16 cadaveric dog knees were analyzed by 3T MR system with proton density sequence in the sagittal plane at four different knee positions. Prior to MR imaging, each knee was placed into a loading jig to mimic weight-bearing conditions. After imaging, the MRIs were manually segmented and the kinematics were observed. The manually segmented models were used to develop a statistical shape model, which was deformed to match the menisci in the MRI, automatically. The automatic segmentation scheme was validated for their shapes and sizes by comparing with manually segmented models. In results, the meniscal displacements were much larger than in humans, displacing 13.5mm and 13.7mm posteriorly on average for the lateral and medial menisci during flexion (p=0.90). The mean and standard deviations for Dice overlap index were 81.8% +/-3.0% and 81.9% +/-1.9% and absolute surface distance of 0.43mm +/-0.1mm and .48mm +/-0.1mm between manual and automatic segmentations. Through this study we have successfully provided the capabilities to automatically segment the menisci and the meniscal kinematic data for comparative purposes for future studies. ( en )
General Note:
In the series University of Florida Digital Collections.
General Note:
Includes vita.
Bibliography:
Includes bibliographical references.
Source of Description:
Description based on online resource; title from PDF title page.
Source of Description:
This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Thesis:
Thesis (Ph.D.)--University of Florida, 2016.
Local:
Adviser: BANKS,SCOTT ARTHUR.
Local:
Co-adviser: FREGLY,BENJAMIN J.
Statement of Responsibility:
by Brian H Park.

Record Information

Source Institution:
UFRGP
Rights Management:
Copyright Park, Brian H. Permission granted to the University of Florida to digitize, archive and distribute 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.
Classification:
LD1780 2016 ( lcc )

Downloads

This item has the following downloads:


Full Text

PAGE 1

KINEMATIC ANALYSIS AND AUTOMATIC SEGMENTATION OF THE MENISCI By BRIAN PARK A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR O F PHILOSOPHY UNIVERSITY OF FLORIDA 2016

PAGE 2

2016 Brian Park

PAGE 3

To my family

PAGE 4

4 ACKNOWLEDGMENTS I wish to express sincere appreciation to Dr. Scott Banks and Dr. Antonio Pozzi, my advisors, who gave me enormo us support and guidance throughout my quest I would like to thank Dr. Matthew Winter, Dr. Bunita Eichelberger Dr. Brian Saunders, Dr. Silvia Marches, and Alex Iorgulescu for their help during the data collection and MR imaging needed for the studies. I would also like to thank Dr. Peter Laz, Lowell Smoger, and Azhar Ali and their team from the University of Denver in there guidance and the program needed during the initial process of building the model for the SSM. Special thanks goes to my family and t o my wife, who has been with me from the beginning and constantly giving en couragement and support needed to finish my quest. Lastly, I would like to thank Jesus Christ, my Lord and my savior.

PAGE 5

5 TABLE OF CONTENTS page ACKNOWLEDGMENTS ................................ ................................ ................................ .............. 4 LIST O F TABLES ................................ ................................ ................................ ......................... 7 LIST OF FIGURES ................................ ................................ ................................ ....................... 8 LIST OF ABBREVIATIONS ................................ ................................ ................................ ........ 9 ABSTRACT ................................ ................................ ................................ ................................ 10 CHAPTER 1 INTRODUCTION ................................ ................................ ................................ ................. 12 Background ................................ ................................ ................................ .......................... 12 Menisci ................................ ................................ ................................ ........................... 12 Osteoarthritis ................................ ................................ ................................ ................ 13 Treatment of Meniscal Tears ................................ ................................ ..................... 13 Biomechanical Studies ................................ ................................ ................................ 15 Magnetic Resonance Imaging ................................ ................................ ................... 15 Medical Image Segmentation ................................ ................................ .................... 16 Statis tical Shape Model (SSM) ................................ ................................ .................. 18 Previous Meniscal Segmentation Work ................................ ................................ ... 19 Motivations for the Study ................................ ................................ ................................ ... 19 Research Overview ................................ ................................ ................................ ............ 20 Quantifying Meniscal Volume using 1.5T and 3T ................................ ................... 20 Quantifying Meniscal Kinema tics of Dogs ................................ ............................... 20 3 D Semi Automatic Segmentation of Menisci using Statistical Shape Model .. 21 2 QUANTIFYING MENSICAL VOLUME USING 1.5T AND 3.0T MRI .......................... 23 Summary ................................ ................................ ................................ .............................. 23 Background ................................ ................................ ................................ .......................... 24 Methods ................................ ................................ ................................ ................................ 26 Experimental Animals ................................ ................................ ................................ 26 Experimental Design ................................ ................................ ................................ ... 26 MRI Protocols ................................ ................................ ................................ ............... 26 Meniscal Volume Measurements ................................ ................................ .............. 27 Statistical Analysis ................................ ................................ ................................ ....... 28 Results ................................ ................................ ................................ ................................ .. 28 Discussion ................................ ................................ ................................ ............................ 30 Conclusion ................................ ................................ ................................ ............................ 32 3 QUANTIFIYING MENISCAL KINEMATICS OF DO GS ................................ ................ 37

PAGE 6

6 Summary ................................ ................................ ................................ .............................. 37 Introduction ................................ ................................ ................................ .......................... 37 Materials and Methods ................................ ................................ ................................ ....... 39 Experimental Animals ................................ ................................ ................................ 39 Experimental Design ................................ ................................ ................................ ... 39 MRI Protocols ................................ ................................ ................................ ............... 40 Segmentation ................................ ................................ ................................ ............... 40 Kinematic Analysis ................................ ................................ ................................ ....... 40 Statistical Analysis ................................ ................................ ................................ ....... 41 Results ................................ ................................ ................................ ................................ .. 4 1 Discussion ................................ ................................ ................................ ............................ 42 4 3 D SEMI AUTOMATIC SEGMENTATION OF THE MENISC US USING STATISTICAL SHAPE MODELS ................................ ................................ ..................... 50 Summary ................................ ................................ ................................ .............................. 50 Introduction ................................ ................................ ................................ .......................... 50 Materials and Methods ................................ ................................ ................................ ....... 52 MR Image Datasets ................................ ................................ ................................ ..... 52 Menisci Segmentation ................................ ................................ ................................ 52 Region of interest selection ................................ ................................ ................. 53 MR image preprocessing ................................ ................................ ..................... 53 Rigid body initialization ................................ ................................ ........................ 54 SSM fitting ................................ ................................ ................................ .............. 54 Manual segmentation ................................ ................................ ........................... 55 Validation ................................ ................................ ................................ ....................... 55 Results ................................ ................................ ................................ ................................ .. 57 Discussion ................................ ................................ ................................ ............................ 57 5 CONCLUSION ................................ ................................ ................................ ..................... 61 LIST O F REFERENCES ................................ ................................ ................................ ........... 63 BIOGRAPHICAL SKETCH ................................ ................................ ................................ ....... 72

PAGE 7

7 LIST OF TABLES Table page 2 1 Average medial and lateral meniscal volume measurements, 1 standard deviation (SD) and intersubject coefficient of variation (CV) results. ..................... 35 2 2 Agreement between vol ume measurement from segmented 1.5T and 3T MR 3D models, gold standard and 1.5T MR 3D models, and gold standard and 3T MR 3D models for both medial and lateral meniscus. ................................ ........ 36 3 1 Lateral and med ial meniscus movements relative to 30 flexion (full extension). Anterior/Posterior direction is represented by positive/negative values. ................................ ................................ ................................ .............................. 49 4 1 Validation results of the automatic segmentatio n algorithm for the medial and lateral meniscus models. ................................ ................................ ............................... 60

PAGE 8

8 LIST OF FIGURES Figure page 1 1 Meniscal anatomy (Anderson MW 2002) ................................ ................................ ... 22 2 1 MR images of the menisci and the resulting 3D meniscal models.. ....................... 33 2 2 Average mean (1 standard deviation) medial and lateral meniscal volum e measurements ................................ ................................ ................................ ................ 34 3 1 A specimen attached to the knee loading jig made with MR c ompatible non ferrous materials ................................ ................................ ................................ ............. 46 3 2 Images of the segmented menisci are superimposed on a left tibia model, with the anat omical coordinate systems shown ................................ ........................ 47 3 3 Sagittal views of the segmented tibia/fibula and menisci wit h fitted cylinder s superimposed ................................ ................................ ................................ ................. 48 3 4 Average medial and lateral meniscal t ranslations for four knee poses ................. 49 4 1 Flow diagram of the automatic segmentation scheme, steps I through IV are taken to develop a meniscus model automatically ................................ ................... 59 4 2 The correlation and Bland Altman analyses performed for the medi al and lateral meniscus volum es ................................ ................................ .............................. 60

PAGE 9

9 LIST OF ABBREVIATIONS 1.5T 1.5 Tesla 3T 3.0 Tesla ASM Active shape model CT Computer Tomography CV C oefficient of variation Dice Dice overlap coefficients FA Flip Angle FOV Field of view HD Hausdorff distanc e IACUC Institutional Animal Care and Use Committee ICC Intraclass correlation coefficient ICP Iterative closest point MASD M ean absolute surface distance s MR Magnetic Resonance (imaging) OA Osteoarthritis PCA Principal component analysis PD Proton density PDM Point distributed model RMS Root mean square ROI R egion of interest SD Standard deviation SSM Statically shape model TE Echo time TR Repeat time

PAGE 10

10 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 KINEMATIC ANALYSIS AND AUTOM ATIC SEGMENTATION OF THE MENISCI By Brian Park December 2016 Chair: Scott A Banks Major: Mechanical Engineering The menisci of the knee are crescent shaped structures that c ontribute to load distribution and stability of the tibiofemoral joint The kinematics of menisc i are of particular interest, but there remains a lack of robust evidence regarding meniscal translation in the h ealthy knee for comparative purposes One of the greatest challenges in doing a study of the menisci is the lack of visibility and contrast between neighboring anatomies in the MRI. The purpose of this study is to set up a meniscal study framework by inve stigating the kinematics of the meniscus in dogs and developing an automa tic segmentation algorithm to guide future studies. T his information may be valuable because t he dog has been extensively used as an experimental mod el to study meniscal treatments. In this study, menisci from 16 cadaveric dog knees were analyzed by 3T MR system with proton density sequence in the sagittal plane at four different knee positions. Prior to MR imaging each knee was placed into a loading jig to mimic weight bearing condi tions. After imaging, the MRIs were manually segmented and the kinematics were observed. The manually segmented models were used to develop a statistical shape model, which was deformed to match the menisci in the MRI, automatically. The automatic segmenta tion scheme was

PAGE 11

11 validated for their shapes and sizes by comparing with manually segmented models. In results, the meniscal displacements were much larger than in humans, displacing 13.5mm and 13.7mm posteriorly on average for the lateral and medial menisci during flexion (p=0.90). The mean and standard deviations for Dice overlap index were 81.8% 3.0% and 81.9% 1.9% and absolute surface distance of 0.43mm 0.1mm and .48mm 0.1mm between manual and automatic segmentations. Through this study we have succes sfully provided the capabilities to automatically segment the menisci and the meniscal kinematic data for comparative purposes for future studies

PAGE 12

12 CHAPTER 1 INTRODUCTION Background Menisci The menisci of the knee are crescent shaped structures that con tribute to load distribution, lubrication, proprioception and stability of the tibiofemoral joint (Kurosawa et al., 1980, Mow et al., 2005, Mcdermott et al. 2008). They are located between the femur and tibia. Because of their unique geometry, the menisci improve congruity and provide greater articular contact area between the femoral condyles and tibial plateaus. The main functions of the menisci are to protect the articular cartilage by minimizing contact stress by maximizing contact area within the knee. The knee is comprised of two menisci, the medial and lateral. The medial meniscus covers about 50% and the lateral covers about 70% of their compartme nts on the tibial plateau (Fox, 2007). During full extension, 50% of the body weight is transmitted throu gh the menisci and increases to 85% du ring 90 flexion (Fox, 2007; Ahmed, A.M. et al. 1983). At the macroscopic level, the meniscus can be divided into three parts, the anterior horn, posterior horn, and the body, with two different zones, the white zone and red zone a s shown in Figure 1 1 (Anderson, 2002). At the microscopic level, the meniscus is more than 90% type 1 collagen arranged in bundles that c ourse circumferentially (Lee 2000). The circumferential collagen bundles deform with axial loading, whi ch evenly distributes the load acros s the joint surface (Anderson 2002). Total or partial loss of meniscus due to injury or meniscectomy results in increased force observed by the articular cartilage. Degeneration of the meniscus can lead to progression of osteoarthr itis in the knee joint (Burr et al, 1984).

PAGE 13

13 Osteoarthritis Osteoarthritis (OA) is the most common form of knee arthritis and is the leading cause of chronic disability in the United St ates (CDC 2001). OA is a disease that slowly causes joint deg eneration through mechanical and biologic mechanisms OA is the leading indication for joint replacement surgery in the United States, and 905,000 knee replacements were performed in 2009 at a cost of $42.3 billion ( Murphy et al., 2000 ). It is reported tha t 9.3 million U.S. adults had symptomatic knee OA in 2005 (Lawrence et al., 2008). A significant loss of the meniscal tissue due to injury or degeneration leads to oste oarthritis in the knee (Burr et al, 1984). Exact causes for meniscal degeneration are still unknown and no cause and effect relationship between meniscal injury and development of OA has been established. There are a few different options for treating meniscal injury, including total or partial meniscectomy meniscal repair, meniscal transp lant and regeneration. T reatment of Meniscal T ears Total and partial meniscectomy is a procedure that consists of total or partial resection of the torn part of the meniscus However, because a portion of meniscus is cut out and tissue is reduced, the remaining tissues are forced to absorb more load with lower mechanical integrity. There are extensive studies, with numerical and experimental models, reporting that total meniscectomy can accelerate development of OA ( Englund and Lohmander 2004). Also, p artial meniscetomy does not eliminate the ri sk of OA (Englund, et al., 2003). It has been reported that 16% to 34% of meniscal removal increases the contact forces on the articul ar cartilage by 350% (Seedhom et al., 1979). However, partial meniscectomy is still the most common form of treatment

PAGE 14

14 for meniscal injury or meniscal tear and there are about 700,000 meniscectomies done annually in the U.S alone at an estimat ed cost of $4 billion (Cullen et al., 2006). Meniscal repair is a meniscus preserving techn ique that treats the torn meniscus typically by suture repair. Young patients with recent injuries are good candi dates for meniscal repairs Generally, there are 3 different types of fixation for meniscal repairs, idered a valuable treatment option for most pat ients with meniscal tears (Hoff elner et al. 2011). One critical drawback with meniscal repair is that the meniscus can only heal itself in the red (vascular) zone, which is the outer part of the meniscus and i s only comprised of one third of the meniscal body (Tudor et al. 2014). When menisci cannot be preserved, meniscal transplant is a treatment option. This technique typically uses one of four different types of allografts: cryopreserved, deep frozen (froze n fresh), fresh, and lyophilized (dried frozen) (Lubowitz et al. 2007). Meniscal transplant has been shown to partially replace the function of the normal meniscus, with concomitant improvements in pain and knee function (Felix et al., 2003). Some of the limitations of this procedure include tissue availability, risk of immune reaction, risk of disease transmission, and graft sizing (Lubowitz et al. 2007). Also, long term protection from joint degeneration is not yet proved and further research is needed ( Wirth et al., 2002, Felix et al., 2003). Due to the growing interest in meniscal tissue preservation techniques and the limitations of meniscal repair and allogr aft supply for transplants, some researchers have focused their interests on tissue regeneratio n and artificial tissue engineering. The potential possibilities using this treatment are for improving biological bonding for repairable meniscal tears and partial or total tissue regeneration to restore the removed

PAGE 15

15 tissues after partia l or total menisce ctomy (Scotti 2013). There has been extensive study of tissue regeneration/engineering, and a detailed review recently h as been written by Scotti et al U nfortunately there are apparently no published studies proving any of these treatment options are ef fective in preventing OA development. More investigations are needed to determine the most effective treatment options. Biomechanical S tudies Kinematics Many biomechanical studies have been done to better understand the kinematics and joint contact mec hanics of the healthy, injured, and surgically treated knees. Studies of the meniscal kinematics allow us to understand its behavior during different actives. Learning about meniscal kinematics of a healthy patient will give the gold standard and ideal c omparison and goal when treating patients with meniscal injuries. Also, learning about different types of injuries and how they affect the meniscal and knee kinematics will help during initial diagnosis process and the success of the treatments post opera tively. Contact mechanics There have been studies done using experimental model using in vivo and in vitro human patients and using in vivo and in vitro animal models. These models help us to understand the anatomy and its biology in the macro and micro level. Experimental models also give us the capabilities to measure stress, strain, volume, material and mechanical properties numerically. For those properties and measurements that are readily available and obtainable experimentally, numerical models have been developed for estimation and prediction capabilities. Magnetic Resonance Imaging Magnetic Resonance Imaging (MRI) system provides high quality images of the internal structures in two or three dimensions. This is a non invasive technique that is

PAGE 16

16 comparable with Computer Tomography (CT) x ray scans but without ionizing radiations that presents risk of cancer. The MRI scanner creates strong external magnetics field, 1 to 10 Tesla, around the patient that aligns the hydrogen atoms in the same dir ection. The images of the internal structures are created by collection the radio signals from the nuclei as it realigns into proper positions after the pulse of radio waves from the scanner knocks the particles out of alignment. MRI may be used in situa tions where organs or soft tissues are being studied. In orthopedics, MRI are used to examine bones, joints, and soft tissues such as cartilage, muscles and menisci in the knee for diagnosing injuries, OA or other joint degenerated diseases. There has be en extensive studies done on different MRI scanning protocols to determine the combinations needed when imaging differ ent soft tissues and bones in the body Medical Image Segmentation Segmentation is a process that creates a 3D model from stacks of mul tiple 2D images. In the medical field, segmentation is used to turn stacks of 2D images from devices such as MRI and CT scanners into a 3D model of internal organs, blood vessels, bones, and other soft tissues like cartilage and menisci. There has been a n extensive research done on both manual and automatic medical image segmentations algorithms ( Fu and Mui, 1981; Mclnerney and Terzopoulos, 1996; Pham DL et al. 2000 ; Mitchell et al., 2001; Fripp et al. 2010; Zhang et al. 2013; Paproki et al., 2014 ; Yin e t al, 2010 ). Segmentations assist in numerous biomedical image applications such as quantifications of tissue volume ( Stone et al., 1994; Bower et al., 2007 ), study of the anatomical structures (Wirth et al., 2010 Thompson et al. 1991 ), and kinematics (Ve di, et al., 1999; Yao et al., 2008 ). It is also used to better visualize the human anatomy

PAGE 17

17 during diagnosis and for research. Conventionally, segmentations were not need during medical image analysis for assisting with diagnosis and treatment planning, b ut with advances in technology segmentations are used to obtain more clear information from medical images to obtain accurate diagnosis and treatments. There are many different methods of image segmentations such as using deformable models, region based, intensity based, model based and graph cut etc. Currently most popular automatic segmentation algorithms are model based segmentations like Active shape model (ASM) and active appearance model ( AAM ), which uses statistical shape models (SSM) (Cootes et a l. 1995, Fripp et al., 2010, Paparoki et al., 2014) At the 2010 MICCAI conference, a competition for knee A Grand gmentation method ranked the highest in terms of accuracy (Seim et al., 2010; Vincent et al., 2010) ASM was first proposed by Cootes and Taylor and the term was adopted to make the distinction between their method and Active Contour Model (snakes ) (Coot es et al., 1995) The basic idea of ASM is to fit a model in an image by deformation. The key is that the model can only deform in the ways found in the training sets, giving it a shape constraints to the deformable model. ASM uses SSM where Cootes and T aylor called Point Distributed Model (PDM), which is a model represented by point cloud. The shape constraints are applied by training the base model by aligning with the same or similar models that has been deformed and calculating the variance of the po ints and restricting the model to only move the way it was trained. One of the major limitation for ASM is that it does not consider the image intensity information, so it is difficult to obtain

PAGE 18

18 accurate segmentations for image with less prominent edges, h owever it is relatively fast and simple. Statistical Shape Model (SSM ) SSM is a shape described by a trained mean shape plus the shape variation of each orthogonal mode, times the shape parameter, that control the modes of the variation: (1 1) ( 1 2) where a shape vector, is represented by a set of n points on its surfac e with every th point representing the same corresponding position for all shapes. Shape mean, and the covariance matrix, C, can be obtain ed after aligning the shape using Procrustes analysis method (ref) with m number of training data sets. ( 1 3) ( 1 4) After alignment, princip al component analysis (PCA) is used to analyze the shape variations. The corresponding eigenvectors, and eig envalues, are calculated by transforming C to a diagonal matrix using eigen decompositing method, ( 1 5) The variance can be explained by the eigenvectors and the corresponding eigenvalue and most of the variations can usual ly be explained by a smaller number of modes, t, where t <
PAGE 19

19 Since the variance of over the training set can be explained by, any new shapes can be ge nerated by varying the parameters within the suitable limits of, ( 1 7) Previous Meniscal Segmentation Work There are many work s on knee segmentations whether it is manual, semi automatic, or fully automatic, but only few on the m eniscal segmentation. The work of Fripp et al. and Yin et al. seem to be most famous Fripp uses 3D ASM to segment the bones and cartilages, then extract menisci by fitting medial and lateral shape models combined with Gaussian mixture m odel (GMM) which is an estimation of the intensity distribution and the probability of it being soft tissue in the ROI from the MR images. They reported accurate bones and bone cartilage interface (BCI) segmentation but not with the meniscal segmentation (Fripp et al. 2009). Yin et al. is based on LOGISMOS (layered optimal graph image segmentation of multiple objects and surfaces) framework for automatic segmentation of the knee bone and cartilage surfaces from volumetric MR image and the meniscus segment ation was done using pattern recognition technique based on the segmented cartilages (Yin et al. 2010). Also, most currently Paproki A et al. segmented the menisci using combination of ASM and SSM without constraints to cover the variability not shown in t he training sets (Paproki A et al. 2014). All these works provided promising results but there are still rooms for improvements. Motivations for t he S tudy There are still great interests in studying the menisci of the knee in the world of orthopedics. T here are extensive studies done in diagnosis and treatment techniques, which will allow increase in patient satisfactions after surgical treatments and surgical

PAGE 20

20 success rates. Our motivation for this study developed from the fact that for most of these st u dies, medical image segmentations were done t o generate 3D reconstructed model s from stacks of 2D images from MRI scanner and these segmentations were done m anually There have been studies done to automate the segmentations process by previous researche rs with promising results but none (commercial and/or open source) were publically available for uses The objectives of this study are to obtain kinematic da ta of the menisci and develop an accurate and robust automatic segmentation algorithm using stati stical shape models. Research Overview Quantifying M e niscal V olume u sing 1.5T and 3T The objective of the first a im is twofold: to determine if accurate measures of meniscal volume could be obtained using clinical veterinary MR facilities, and to compare results using 1.5T and 3.0T systems with similar common clinical protocols using a single proton density (PD) sequence. Stifles were imaged under a compressive load and in four different joint positions to simulate in vivo meniscal conditions. MR based men iscal volumes were compared with gold standard water displacement tests. Dogs are used because it has been extensively used as an experimental model to study meniscal treatments such as meniscectomy, meniscal repair, transplantation and regeneration Quan tifying Meniscal Kinematics of D ogs The o bjective of the second aim is t o provide a three dimensional (3D) description of meniscal kinematics of intact, loaded cadaveric stifle s at full flexion and extension, along with tib i al internal and external rotatio n applied at 45 flexion. Four

PAGE 21

21 poses were analyzed for each knee The meniscal kinematics were analyzed using 3D models reconstructed from 3T MRI taken at both sagittal and coronal views 3 D Semi Automatic Segmentation of Menisci u sing Statistical Shape M odel The objective of th is study was to develop and validate an automatic 3D segmentation scheme for canine menisci from MR images which would be made publically available for use in any future studies or further development and refinement. The accuracy of the automatic segmentation scheme was validated using two distance based measurements : mean absolute surface distance (MASD) and Hausdorff distance (HD) and three volume based measurements : sensitivity, specificity, and Di ce overlap coefficient (Dice)

PAGE 22

22 Figure 1 1 M eniscal anatomy (Anderson MW 2002)

PAGE 23

23 CHAPTER 2 QUANTIFYING MENSICAL VOLUME USING 1.5T AND 3.0T MRI Summary Background: The dog has been used extensively as an experimental model to study meniscal treatment s such as meniscectomy, meniscal repair, transplantation and regeneration Accurate quantification of meniscal size volume and morphology is a crucial step for developing models of the meniscus The goal of this investigation w as to determine if accurate measures of meniscal volume could be obtained using clinical veterinary MR facilities, and to compare results using 1.5T and 3.0T systems with similar protocols using a single proton density (PD) sequence. Results: MR based volume measurements for menisci from dogs have higher absolute errors than those reported for human meniscal volume measurements, with average volume errors as high as 1 6.3 % for medial and 1 9.3 % for lateral menisc i when compared to water displacement measurements (gold standard). Howeve r, MR Based volume measurements were reproducible, with inter and intra observer reproducibility root mean square coefficients of variation of 11% and 6% for medial and 14% and 4% for lateral meniscus. N o significant volume differences were found between M R based measurements for different joint poses, for either medial or lateral menisci (p>0.05). Conclusions: Repeatable measurements of meniscal volumes using MR may provide a useful capability for assessment of postoperative results following meniscal tre atments such as meniscectomy and meniscal regeneration. Despite the studies limitations, we have shown that common clinical MR scanners and signal sequences with reasonable accuracy and good repeatability.

PAGE 24

24 Bac kground The menisci of the knee are crescent shaped structures that contribute to load distribution, lubrication, and stability of the tibiofemoral joint ( Kurosawa et al. 1980 ; McDermott et al. ; 2008, Mov et al., 2005 ) Because of their unique geometry, t he menisci improve congruity and provide greater articular contact area between the femoral and tibial condyles. The importance of the structure function relationship of the knee meniscus has been demonstrated in experimental studies, guiding surgeons towa rds a conservative approach in treating meniscal injuries ( Fithian et al., 1990 ) Meniscal injury, repair, replacement and re generation are topics of great interest to surgeons in both human and veterinary medical fields Total meniscectomy is a treatment option that is rarely performed because of the direct relationship between loss of meniscal tissue and the development of osteoarthritis ( Burr et al. 1982; Roos et a., 1998 ; Englund et a., 2004; Berjon et al., 1991 ) Meniscal repair is preferred when pos sible, but partial meniscectomy is still indicated when the meniscal tear cannot be sutured ( Tudo et al. ; 2014, Thieman et al., 2010 ) In an attempt to restore meniscal function when repair is not possible, cell and tissue regenerative strategies have been investigated, mostly in experimental studies ( Scotti et al., 2013 ) The ability to quantify meniscal volumes noninvasively permits acute study of surgical procedures (e.g. partial meniscectomy) and longitudinal stud ies of meniscal regeneration following n ovel treatments such as intra arti cular mesenchymal stem cells ( Vangsness et al., 2014 ) Accurate quantification of meniscal size and morphology is a crucial step for developing models of the meniscus. Many finite element models have been developed to stud y the biomechanical contact behavior of the knee joint These models provide numerical solutions of stress distributions within tissues that cannot be readily

PAGE 25

25 measured in experimental models (Guo et al. 2014; Pena et al., 2005 ) P revious studies have accur ately measure d the meniscal volumes for humans (in vitro and in vivo) with different imaging protocols and different MR magnet strengths ( Blcker et a l., 2013; Bowers et al., 2007; Stone et al., 1994; Wirth et al., 2010; Siorpaes et al., 2012) Similar ass essments have not been reported for application in dogs and it cannot be assumed that the same results are possible in the se much smaller joint s nor broadly applicable since 3T MRI capability is uncommon in veterinary practice. The dog has been extensi vely used as an experimental model to study meniscal treatments such as meniscectomy, meniscal repair, transplantation and regeneration ( Bennett et a l., 1991; Cook et al., 2001; King et al., 1936; Pozzi et al. 2010 ) The goal of this investigation was twof old: to determine if accurate measures of meniscal volume could be obtained using clinical veterinary MR facilities, and to compare results using 1.5T and 3.0T systems with similar common clinical protocols using a single proton density (PD) sequence. Stif les were imaged under a compressive load and in four different joint positions to simulate in vivo meniscal conditions. MR based meniscal volumes were compared with gold standard water displacement tests. Since dog menisci are about 25% the volume of avera ge human menisci, we assumed MR based volume measurement of dog menisci would be less accurate than what has been reported for human meniscal volume studies. The significant reduction in size means each meniscus fills fewer 3D voxels and consequently may s uffer from greater partial volume averaging errors.

PAGE 26

26 Methods Experimental Animals IACUC# 20130788 was approved for this study involving 16 p airs of menisci (n=32 ), from 16 intact stifles from 16 medium breed dogs (range 20 35 Kg) euthanized for reasons un related to this study. The dogs were skeletally mature and without apparent signs of stifle disease confirmed during anatomical dissection Each intact stifle was prepared to have consistent lengths of femur and tibia 12 cm above and below the joint, and a transepicondylar k wire was drilled through the femur and tibia under fluoroscopic guidance to permit mechanical loading. Experimental Design Prior to MR imaging, e ach stifle was placed into a nonmetallic loading jig that provided a fixed flexion angle and a net 88N load across the joint. The load was selected to simulate a 30Kg dog in standing position. The stifles were placed at four different positions to simulate the full range of stifle fl exion during daily activities [Kim et al., 2015 ]: 35 flexio n (145 flexion us in g human angle conventions), 150 flexion (30 in humans ), and 135 flexion (45 in humans ) with tibial internal and tibial ex ternal rotation. The stifles were imaged serially in each of these positions without delay between tests. Tibia l internal and external rotation was obtained by applying a constant torsion to the proximal tibia with an elastic band. Specimens were kept hydrated during the entire MR imaging process by wra pping them in wet towels once thawed. MRI Protocols Intact s ti fles were thawed prior to imaging in two freeze/thaw cycles, with half scanned at 1.5T in the first thaw cycle, and half scanned at 3.0T in the first thaw cycle. A quadrature knee coil was used for 1.5T MR imaging (Toshiba Titan, Toshiba, Duluth,

PAGE 27

27 Georgia, U.S.A) and a 15 channel transmit/receive coil was used for 3.0T MR imaging ( Siemens Verio, Siemens, Malvern, Pennsylvania, USA ). Based upon preliminary exams performed to obtain good quality images of the menisci, the 3.0 T MR scans were per formed using a P D sequence (TR/TE/FA, 4300/18/9 0; FOV, 120mm; matrix, 512x512, slice thickness/gap, 2mm/0mm ), and the 3.0T MR scans were performed using a similar PD sequence ( TR/TE/FA, 4210/25/ 9 0; FOV, 120mm; matrix, 512x512, slice thickness/gap, 2mm/0mm ) All images we re stored on a picture archive and communication system (PACS) in digital imaging and communications in medicine (DICOM) format. Meniscal Volume M easurements After MR imaging the meniscal specimens were dissected from the joint and isolated from the attac hed cranial and caudal ligaments. The volume of the menisci was measured using a water volume test, which was considered the gold standard for comparison with the MR based volume measuremen ts [Bowers et al., 2007, Stone et al., 1994 ]. The menisci were plac ed in a 10 ml graduated cylinder filled with 9 ml of saline. The elevated saline was removed back to 9 ml with a micropipette and recorded as water volume of the specimen. Measurements for each specimen were repeated three time s by two observers. The avera ge of six water volume measurements three measurements each from two observers, was used as a reference standard for meniscal volumes. To calculate the meniscal volume using manual segment ation of MR images, an open source software (ITK SNAP www.itksnap.org ) ( Yushkevich et al., 2006 ) was used to reconstruct 3D surface representations of the medial and lateral menisci (Figure 2 1) A sub group of 14 stifles at a single position were segmented again by a single

PAGE 28

28 obse rver (at least 8 months later) to assess intra observer reproducibility. An additional subgroup of 5 stifles at a single position were segmented by 3 different observers to assess inter observer reproducibility. Model post processing was done using the alp ha shape algorithm ( MATLAB R2014b, MathWorks, Natick, Massachusetts, U.S.A), segmented contours were meshed with triangles to create 3D models, and volume s were calculated (Geomagic Studio Figure 2 1 ). Statistical A nalysis All measurements for significan ce were calculated using a one way ANOVA with alpha of 0.5 in this study. The meniscal volumes measurements from raw MR voxel and post processed 3D MR models from both 1.5MR and 3.0MR were compared for significance. Volume measurements from 3D model and th e gold standard were also compared The root mean square (RMS) standard deviation (SD) and coefficient of variation (CV) were calculated across 14 stifles for intra observer segmentation reproducibility involving 28 degrees of freedom, and 5 stifles for i nter observer segmentation reproducibility in volving 10 degrees of freedom (Gluer et al., 1 995 ) All statistical calculations were performed using Microsoft Excel. Results Raw voxel volumes from both 1.5T and 3.0T measurements were consistently less than v olumes from post processed 3D MR models (Fig. 2). For 1.5T images, raw voxel volumes (Fig. 1C) averaged 351 mm 3 70 mm 3 (p =0.006) and 520 mm 3 107 mm 3 (p < 0.001) while post processed 3D models (Fig. 1D) averaged 390 81 mm 3 and 596 128 mm 3 for medi al and lateral mensici, respectively. For 3T images, raw voxel volumes averaged 343 mm 3 65 mm 3 (p =0.004) and 479 mm 3 100 mm 3 (p < 0.001) while post processed 3D MR models averaged 403 74 mm 3 and 572 121 mm 3 for

PAGE 29

29 medial and lateral menisci, respecti vely (Table 2 1). Post processed 3D MR model volumes were also consistently smaller than water displacement volumes (Figure 2 2). Compared to water displacement volumes, 3T MR 3D model volume s from intact stifles were 1 4.4 % smaller (p < 0.001) for medial m enisc i and 1 9.3 % smaller (p < 0.001) for lateral menisc i. 3D models from 1.5T MR were 16.3% smaller (p < 0.001) and 16.1% smaller (p < 0.001), compared to medial and lateral menisci water displacement volumes, respectively The smallest volume differences were 12 .2 % for medial meniscus under tibial internal rotation for 3T MR 3D models, and 14 .4% for lateral meniscus at tibial external rotation for 1.5T MR 3D models (Table 2 2 ). However, n o significant meniscal volume differences were found between 1.5T and 3T MR 3D model measurements (p>0.34) and between different poses from both MR groups (p>0.6). There was also high average correlation, r = 0.8 for medial and r = 0.9 for lateral meniscus model volumes, between MR measurements and the water displacement vo lumes (Table 2 2). The intersubject volume CV for medial and lateral menisc al model volumes at 1.5T averaged 21 % for both, and for 3T scans averaged 18 % and 21 % respectively (Table 2 1). The SD for inter observer reproducibility for medial and lateral meniscus models were 39 mm 3 and 79 mm 3 and the RMS CV were 11% and 14%, respectively. The RMS CV ratio between intersubject and inter observer was 1.9 for medial and 1.5 for lateral meniscus. The SD and RMS CV for intra observer reproducibility were 21 m m 3 and 6% for medial and 20 mm 3 and 4% for the lateral meniscus models. The RMS CV ratio between intersubject and intra observer was 3.5 and 5.3 for medial and lateral meniscus, respectively.

PAGE 30

30 Discussion Accurate measurements of dog meniscal volumes using M R may allow to assess postoperative results followi ng meniscal treatments in dogs ( Cook et al., 2001 ; Arnoczky 1999; Arnoczky and Warren 1983 ) The objective of this study was to compare the volumes of 3D meniscal models segmented from 1.5T and 3T MR ima ges to determine if an accurate measurement of meniscal volumes can be obtained using common clinical veterinary MR capabi lities We selected a stress MR protocol to evaluate intra articular structures u nder joint loading conditions ( Tremolada et al. 2013 ) Since the ultimate goal of regenerating meniscal tissue is to restore the damaged or injured tissue to its normal function, the stifles were compressed during imaging to mimic weight bearing conditions. As expected, MR based volume measurements for dog menisci have higher average errors than have been reported for human meniscal volume measurements, with highest absolute mean volume errors of 16.3% and 19.3% for the medial and lateral menisci, respectively, when compared to unloaded water displacement volumes. A post processing was done to the 3D model because of the consistent underestimation and unknown error from partial volume averaging. The sole purpose of the post processing was to obtain outside boundaries of the meniscus model similar to a con vex hall but holding on to the interior boundaries that make up the anterior and posterior horns, henceforth developing a more smooth crescent shape meniscus model instead of block like structure (Figure 2 1:C, D). There are several factors potential ly c ontributing to consistent underestimation of meniscal volumes using MR imaging. The f irst, and the largest contribution is that the MR measured volumes are from stifles that are under compressive load during imaging while the water volume tests were not un der any compressive load. Kessler et al

PAGE 31

31 reported human meniscal volume changes after running as high as 10% for the medial meniscus and 8.2% for the lateral meniscus ( Kessler et al., 2006 ) Second, dog menisci are about one fourth the volume of human men isci, so equivalent volume measurement accuracy require s MR scans with smaller voxel volumes. Third the scanners available for this study coul d not conveniently provide 1mm slice spaci ng as reported by Bowers et al, who reported average meniscal volume er rors of 4.6% and 7.9% for both medial and lateral meniscus. Th ese factors lead to larger partial volume effects within each voxel, coarser spatial discretization of the reconstructed menisci, and likely explain consistent volume underestimat ion o f dog meni sci using common veterinary clinical scanners. Lateral meniscus is larger in volume and in surface area so it was likely to have been more influenced by the partial volume effect resulting in a slightly larger error than the medial meniscus. No statistica l results were discovered to support this theory, but it agrees with previous studies (Bowers et al., 2007; Stone et al., 1994 ; Siorpaes et al., 2012 ) Finally, th e MR signal sequences used for this study were typical clinical routines, not specialized se quences developed for maximum contrast and spatial resolution between cartilage, meniscus, and bone. Despite a consistent bias in meniscal volumes using either 1.5T or 3T MR compared to water volume measurements, the results were repeatable for all of the MR based measurements, including interim freeze/thaw cycles, different joint poses, and different field strengths with relativity high correlation coefficients (r>0.8) between the MR measurements. The RMS CV for both inter observer and intra observer segm entation reproducibility were lower than CV calculated for intersubject variability. Average intersubject CV was 1.9 times higher for medial and 1.5 time higher for the

PAGE 32

32 lateral meniscus than average CV for inter observer, and 3.5 times higher for medial an d 5.25 times higher for lateral meniscus than CV for intra observer reproducibility, which is comparable to meniscus volume results for inter observer reported by Siorpaes et al. This study has several important limitations. The largest limitation is that the MR based volume measurements were under compressive loads while the water volume tests were not. The consistent underestimation of volume measurement agrees with the fact that menisci are compressible and no obvious trend was detected from measurements from 35 flexion (first pose) to tibial external rotation (last pose). Ideally, the stifles would have been imaged with both MR systems in the same session immediately followed by harvesting and water displacement tests. This was not done, and the stifles underwent multiple freeze thaw cycles pending availability of the MR systems. Efforts were made always to keep the specimens hydrated, but the effect of thermal cycling on meniscal volumes is unknown. A different MR image sequence was used for this study than was used in previous reports, e.g. Bowers et al, making it difficult to directly compare the data or to identify specific causes for differing results. Conclusion In conclusion, despite these studies limitations, we have shown that common clinical M R scanners and signal sequences can be used to quantify volumes with reasonable accuracy and good repeatability.

PAGE 33

33 Figure 2 1 MR images of the menisci and the resulting 3D meniscal model s MR image of an intact stifle at 35 flexion ( A ) an intact stifle at 150 flexion ( B ) an image of 3D meniscus model of the raw voxel (C), and an image of 3D meniscal models from segmented images using the alpha shape algorithm plus smoothing (D).

PAGE 34

34 Figure 2 2 Average m ean (1 standard deviation) me dial and lateral meniscal volume measurements The thinner and more transparent bar graphs are the 1.5T/3T MR 3D model volume measurements and the thicker solid bar graphs on top are raw MR voxel volume measurements.

PAGE 35

3 5 Table 2 1 Average medial and late ral meniscal volume measurements, 1 standard deviation (SD) and intersubject coefficient of variation (CV) results Mean (mm 3 ) SD (mm 3 ) CV (%) Medial 1.5 T 35FLX 407 89 22 150FLX 381 76 20 EXT 392 70 18 INT 378 94 25 Overall 390 81 21 3 T 35FLX 418 87 21 150FLX 403 69 17 EXT 389 74 19 INT 402 64 16 Overall 403 74 18 H20 Overall 447 96 21 Lateral 1.5 T 35FLX 608 118 19 150FLX 588 112 19 EXT 610 141 23 INT 579 141 24 Overall 596 128 21 3 T 35FLX 572 128 22 150 FLX 595 117 20 EXT 557 199 20 INT 565 130 23 Overall 572 121 21 H20 Overall 713 149 21

PAGE 36

36 Table 2 2 Agreement between volume measurement from segmented 1.5T and 3T MR 3D models, gold standard and 1.5T MR 3D models, and gold standard and 3T MR 3D models for both medial and lateral meniscus diff (mm 3 ) error (%) r Medial 1.5T/3T 35FLX 70 15 0.7 150FLX 40 10 0.8 EXT 20 7 0.9 INT 41 11 0.9 Overall 41 11 0.8 H20/1.5T 35FLX 68 15 0.8 150FLX 82 17 0.8 EXT 71 14 0.8 INT 8 9 19 0.8 Overall 77 16 0.9 H20/3T 35FLX 77 17 0.6 150FLX 66 13 0.8 EXT 74 15 0.9 INT 62 12 0.9 Overall 70 14 0.8 Lateral 1.5T/3T 35FLX 65 13 0.8 150FLX 36 6 0.9 EXT 67 12 0.9 INT 59 11 0.9 Overall 57 10 0.9 H20/1.5T 35FLX 10 6 14 0.9 150FLX 125 17 0.9 EXT 103 14 0.9 INT 133 19 0.8 Overall 117 16 0.9 H20/3T 35FLX 141 20 0.9 150FLX 118 16 0.9 EXT 156 21 0.9 INT 148 21 0.9 Overall 141 19 0.9 diff: absolute difference, error: percent error, r: Pearson co rrelation coefficient.

PAGE 37

37 CHAPTER 3 QUANTIFIYING MENISCAL KINEMAT I C S OF DOGS Summary The dog has been used extensively as an experimental model to study meniscal treatments such as meniscectomy, meniscal repair, transplantation and regeneration. However, there is very little information on meniscal kinematics in the dog. This study used MR imaging to quantify in vitro meniscal kinematics in loaded dog knees in four distinct poses: extension, flexion, internal and external rotation. A new method was used to track the meniscal poses along the convex and posteriorly tilted tibial plateau. Meniscal displacements were large, displacing 13.5 mm and 13.7 mm posteriorly on average for the lateral and medial menisci during flexion (p = 0.90). The medial anterior ho rn and lateral posterior horns were the most mobile structures, showing average translations of 15.9 mm and 15.1 mm, respectively. Canine menisci are highly mobile and exhibit movements that correlate closely with the relative tibiofemoral positions. Intr oduction The menisci of the knee are crescent shaped structures that contribute to load distribution, lubrication, and stabi lity of the tibiofemoral joint ( Kurosawa et al. 1980 ; McDermott et al. ; 2008, Mov et al., 2005 ) Because of their unique geometry, t he menisci improve congruity and provide greater articular contact area between the femoral and tibial condyles. The kinematics of menisc i are of particular interest, and have been demons trated in exper imental studies on human knees ( Lee et al., 2000; Bran tigan et al., 1941 ; Bylski Austrow et al., 1994; Thompson et al., 1991; Boxheimer et al., 2004; Kawahara et al., 1999; Shefelbine et al., 2006 ; Epler et al., 2005 ; Rankin et al., 2006; Vedi et al., 1999; Yao et al., 2008 ; Scholes et al., 2015 ) One of the unique

PAGE 38

38 characteristics of the menisci is their complex system of attachments to tibia and femur (Gupte et al., 2007; Proffen et al., 2012; Amadi et al., 2008; Han et a l ., 2012 ) The function of the meniscal ligaments is multifold. These ligaments prevent extrusion of the menisci during load but also guide these structures during knee motion. The posterior horn of the medial meniscus is firmly attached to the tibia because of its attachments to the medial collateral ligament and the tibial plateau. The post erior horn of the lateral meniscus has a looser attachment to the tibia, but is firmly attached to the femur. Many researchers have investigated both human and animal meniscal kinematics in the past, but there remains a lack of robust evidence regarding me niscal translation in the healthy knee for comparative purposes ( Scholes et al., 2015 ) T his information may be valuable for comparative purposes because t he dog has been extensively used as an experimental model to study meniscal treatments such as menisc ectomy, meniscal repair, tr ansplantation and regeneration ( Bennett et al., 1991 ; cook et al., 2001; King et al., 1936; Pozzi et al., 2010 ) However, the authors were unable to locate studies report ing meniscal kinematics in dog s at the time of study initia tion The present study was performed to provide a three dimensional (3D) description of meniscal kinematics of intact, loaded cadaveric stifle s at full flexion and extension, along with tib i al internal and external rotation applied at 45 flexion. Four p oses were analyzed for each knee The meniscal kinematics were analyzed using 3D models reconstructed from 3T MRI taken at both sagittal and coronal views

PAGE 39

39 Materials and Methods Experimental Animals This study was approved by the Univer sity of Florida Inst itutional Animal Care and Use Committee ( # 20130788 ). Sixteen intact hind limbs (n=32 menisci) were collected from 16 dogs (range body weight 20 35 Kg) that were euthanized for reasons unrelated to this study. The dogs were skeletally mature and without ap parent signs of stifle disease confirmed during anatomical dissection. The skin and regional musculatur e were dissected and removed from the limbs. Both femur and tibia were sectioned 12 cm above and below the joint to prepare e ach intact stifle with con sistent lengths of femur and tibia A transepicondylar k wire was drilled through the femur and tibia under fluoroscopic guidance to allow mechanical loading (Figure 3 1) Experimental Design Prior to MR imaging, each stifle was placed into a nonmetallic loading jig that provided a fixed flexion angle and a net 88N load across the joint (Figure 3 1) The load was selected to simulate a 30Kg dog in a standing position. The stifles were placed at four different positions to simulate the full range of stifle flexion during daily activities [2 4 ]: 145 flexion (35 flexion using veterinary angle conventions), 30 flexion (150 flexion using veterinary angle conventions), and 45 flexion (135 flexion using veterinary angle conventions) with tibial internal and t ibial external rotation. The stifles were imaged serially in each of these positions without delay between tests. Tibial internal and external rotation was obtained by applying a constant torsion to the proximal tibia with an elastic band. Specimens were k ept hydrated during the entire MR imaging process by wrapping them in wet towels once thawed.

PAGE 40

40 MRI Protocols Intact stifles were thawed prior to imaging and scanned with the 3.0T MR system using a PD sequence (TR/TE/FA, 4210/25/90; FOV, 120mm; matrix, 512x 512, slice thickness/gap, 2mm/0mm). All images were stored on a picture archive and communication system (PACS) in digital imaging and communications in medicine (DICOM) format. Segmentation All 3D reconstructions were done manually by a single observer u sing open source software (ITK SNAP, www.itksnap.org ) ( Yushkevich et al., 2006 ) Both menisci, tibia, fibula and a cylindrical hole on the tibia drilled for mechanical loading were reconstructed for every knee (n = 6 4). Model post processing was done using the alpha shape algorithm (MATLAB R2014b, MathWorks, Natick, Massachusetts, U.S.A), segmented contours were meshed with triangles to create smooth 3D models much more similar to the real men iscus (Geomagic Studio, F igure 3 2 ). All mode l s from right limbs were mirrored as left limbs to permit direct comparison s Kinematic Analysis First, an anatomical coordinate system for the tibia was set by performing a rigid body registration with a representative tibia model a ligned in the desired anatomical coordinate system [ Joes et al., 2014 ]. The segmented meniscal models were then aligned in the standardized tibial reference frame by registering the transverse carbon fiber bar and fibular anatomic features (which are invar iant geometric features). The process was repeated for models from all flexion angles (145FLX, 30FLX, INT and EXT at 45FLX) Second, we took advantage of the fact the tibial condyles are very well approximated by cylinders to construct a size independent measurement of meniscal

PAGE 41

41 translation: meniscal translations were determined using angular coordinates. C ylinder s w ere fit to the inferior surface s of the medial and lateral menisci for all stifle poses. The central ax e s of the cylinders were aligned with t he medial/lateral axis of the tibia (global z axis, Figure 3 3 A,B ). How well the cylinders fit the distal meniscal surfaces was quantified as the root mean squared ( RMS ) distances between the two surfaces Third, each meniscus w as represented by 10 three dimensional points (5 points on the anterior horn, 5 points on the posterior horn determined as sagittal slice area centroids evenly spaced in the mediolateral direction) in each joint pose (Figure 3 3,C) Fo u rth, the location of the ten meniscal points were expressed as angles measured clockwise from the most posterior point on the cylinder (Figure 3 3 ,D ). Overall meniscal translations were calculated by observing the translation of the centroid of the medial and lateral meniscus at different poses. The se translations were expressed as angles relative to the reference cylinder, and also as linear displacements for an average cylinder radius (displacement = radius*angle). Statistica l A nalysis All measurements for significance were calculated using a one w ay ANOVA with alpha of 0.05. The displacement measurements for the entire meniscus centroid from full extension and full flexion were compared. The translations of anterior and posterior horn points were also compared All statistical calculations were pe rformed using Microsoft Excel. Results All knees showed normal soft tissue, bone, ligament and meni s cus an a tomy in the MRI which also was confirmed after di s section. RMS error s for fitting cylindrical surface s to the inferior meniscal surface s were less t han 0.6 mm for the medial and 1 .0

PAGE 42

42 mm for lateral meniscus All meni s cal measurement s were calculated by taking the average of the 10 points representing both anterior and posterior horns and are shown in both arc angles (in degrees) and in arc length (in millimeters, Figure 3 4). M eni s cal movement s for different pose s are expressed relative to their positions in full extension (30 flexion, Table 3 1) B oth menisc i displayed similar patterns of movement, with medial moving slightly more th an the lateral in the posterior direction during flexion but showing no significant differences (p=0.89 ). Out of 16 specimens, 7 had greater medial than lateral posterior translations, 8 had greater lateral than medial posterior translations, and one specimen had medial and lateral translations within 0.01 mm. During flexion, the lateral posterior horn was more mobile than the anterior horn (p = 0.019) while the medial anterior horn tended to be more mobile than the posterior horn (not significant, p = 0.371, Table 3 1 Figure 3 4). When tracking meniscal centroids, both medial and lateral menisci moved posteriorly with flexion by similar amounts (p = 0.90), while tibial external rotation induced anterior translation of the lateral meniscus and posterior translation of the medial meniscus. Tibial internal rotation induced translations opposite to those observed with external rotation. Discussion The objective of this study was to describe the meniscal kinematics of the intact dog knee at different flexion angles using a novel method to track the meniscal poses along the convex and posteriorly sloped tibial plateau. Ex vivo knee specimens were loaded to simulate weight bearing forces and positioned over their normal envelope of motion: full flexion, full extension, and i nternal and external rotation at 45 flex ion The human tibial plateau is relatively flat, making it convenient and intuitive to describe

PAGE 43

43 meniscal motions as linear displacements on the plane ap proximating the tibial surface (Lee et al., 2000; Bra ntigan et al., 1941; Bylski Austrow et al., 1994; Thompson et al., 1991; Boxheimer et al., 2004 ; Kawahara et al., 1999; Shefelbine et al., 2006; Epler et al., 2005; Rankin et al., 2006; Vedi et al., 1999 ; Yao et al., 2008 ; Scholes et al., 2015) The canine tibial p lateau is highly convex and more posteriorly sloped than in humans, and canine menisc i move significantly in both the anterior/posterior and superior/inferior direction s ( Matsuda et al., 1998 ; Hashemi et al., 2008 ; Pozzi, et al., 2006 ; Wilke et al ., 2002; Sabanci et al., 2014 ; Hildreth III et al. 2006 ) The measurement method presented in this study allowed us to account for the different surface geometry of the lateral and medial tibial plateau and to provide measures that were unaffected by differences in joint size. A similar approach has been used in human studies tracking articular contact locations along the surface of the femoral condyle s during movement ( Logan et al., 2004 ; Freeman et al., 2005 ) S tudies of human knees agree that both menisci transl ate posterior ly as the knee goes from extension to flexion However, there is little consensus on the magnitude s of meniscal displacements in vitro or in vivo In vitro studies uniformly report greater lateral than medial menisc al displacements ( Brantigan et al., 1941 ; Bylski Austrow et al., 1994; Thompson et al., 1991 ) but in vivo studies have reported both greater medial (Boxheimer et al., 2004; Kawahara et al., 1999; Shefelbine et al., 2006; Epler et al., 2005 ) and greater lateral ( Rankin et al ., 2006; Vedi et al., 1999; Y ao et al., 2008) translations. Our results are similar, finding roughly equivalent numbers of knees with greater medial than lateral, and knees with greater lateral than medial displacements.

PAGE 44

44 The most significant finding in our study i s the large magnitude of meniscal translation s during flexion, averaging more than 13 mm in both menisci. These are larger absolute translations than have been reported for human meniscal translations (11. 6 mm for medial anterior horn ( Lee et al., 2000 ), 8 .3 mm (Y ao et al., 2008 ) and 5.6 mm 9.5 mm ( Scholes et al., 2015 ) for the lateral meniscus), and represent much greater relative translations in the small canine knee. The large magnitude of meniscal translation measured in our study may depend on the ro ll back of the femur, which may be more pronounced in dogs than in people because of the geometry of the tibial plateau. The dog has more convex ( Pozzi et al., 2006 ; Wilke et al., 2002 ; Sabanci et al., 2014 ; Hildreth III et al., 2006 ; Ocal et al., 2012 ) an d posteriorly sloped ( Gupte et al., 2007 ; Wilke et al., 2002 ) tibial condylar articular surfaces than humans, and very different radii of curvature between lateral and me dial compartments on the tibia ( Sabanci et al., 2014 ) and femur ( Ocal et al., 2012 ) P osterior translation of the femoral condyles with flexion is influenced by the shapes of the femoral and tibial condyles ( Iwaki et al., 2000 ; Freeman et al., 2005 ) The more concave medial tibial condyle in the human knee may be responsible for the minimal translation reported in weight bearing and non weight bearing kinematics ( Johal et al., 2005 ) In contrast, the convex shapes of the canine tibial condyles may lead to a marked backward rolling of the femoral condyle and consequently large meniscal transl ations. Our methodology allowed describing the motion of each meniscal horn, indicating that meniscal kinematics cannot be reported as whole meniscus translation, but rather as a complex change in shape of this fibrocartilage structure. As an example, when using meniscal centroids for the calculation of motion during flexion, both medial and

PAGE 45

45 lateral menisci moved posteriorly by similar magnitude. However using the ten tracking points instead of the centroid, the lateral posterior horn was found more mobile than the anterior horn while the medial anterior horn tended to be more mobile than the posterior horn during flexion. These different motions patterns of the anterior and posterior horns may result from the coupled motion between axial rotation and flexio n, also known as screw home mechanism, described in both human and canine knees in vivo (Kim et al., 2005; Hallen et al., 1966 ) As expected from the canine anatomy of the meniscal attachments, the lateral posterior horn was more mobile than the anterior horn while the medial anterior horn tended to be more mobile than the posterior horn. In dogs there is a strong menisco femoral ligament, which is analogous to the human anterior menisco femoral ligament and a thin meniscotibial connection ( Gupte et al., 2 007 ) This firm attachment to the femoral condyle is likely responsible for greater mobility of the posterior horn during motion of the lateral condyle. In contrast, the posterior horn of the medial meniscus is strongly attached to the tibial plateau throu gh a large posterior ligament and a strong attachment to the medial collateral ligament, which limit its mobility. Our data confirm that the low incidence of tears of the posterior horn of the lateral meniscus may be explained by its greater mobility than the posterior horn of the medial meniscus ( McCready et al., 2016 ) The primary limitation of our study is that it was performed in vitro. Practically, this was the only method that allowed precise positioning and loading of the joints to obtain images rele vant to ambulatory conditions. Second, because very few studies of canine meniscal mechanics have been reported it is difficult to perform a direct

PAGE 46

46 comparison of our results to previous work. Third, we only observed four knee postures with a single joint c ompressive load, so we cannot comment on the entire gamut of possible meniscal motions in dog knees. In conclusion, this study presents normal weight bearing meniscal kinematics in dog knees. We report a new technique for quantifying meniscal translation that allowed us to account for the different surface geometry of the lateral and medial tibial plateau, and to provide measurements that were unaffected by differences in joint size. This new measurement technique may be helpful in quantifying meniscal tr anslations for all species with high tibial plateau angles, similar to dogs, to obtain better representation of the meniscus moving in multiple directions. Our data indicates much greater meniscal motions in dogs than in humans, with both menisci moving eq ually during flexion. Dog knees are commonly treated clinically for meniscal pathologies and often are used as translational study models for human knee treatments, so we are hopeful our data will provide useful observations of normal meniscal mechanics in dogs. Figure 3 1 A specimen attached to the knee loading jig made with MR compatible non ferrous materials. The jig allowed the joint to be positioned at different flexion angles. Elastic bands applied between the femur (A), tibia (B) and the base all owed application of joint compression and axial rotation loads. The knee positions can be change to deep flexion (C).

PAGE 47

47 Figure 3 2 Images of the segmented menisci are superimposed on a left tibia model, with the anatomical coordinate systems show n. A T he menisci at 30 flexion are shown in o range while at 145 flexion are shown in blue. B The menisci at full t ibial internal rotation are shown in orange and full tibial external rotation sho wn in blue, both captured at 45 flexion.

PAGE 48

48 Figure 3 3 Sa gittal views of the segmented tibia/fibula and menisci with fitted cylinders superimposed. A Lateral image. B Medial image. C Ten points used to represent the meniscus for all measurements. D Image of the medial knee showing the angle measurement f rom the most posterior point of the cylinder (P) to the selected meniscal point (M).

PAGE 49

49 Figure 3 4 Average medial and later al meniscal translation s for four knee poses Each point represents the average position of that meniscal cross section for all knees. The vertical axes represent sagittal plane translation while the horizontal axes show mediolateral displacements. Shaded areas represent 1 standard deviation. Table 3 1. Lateral and medial meniscus movements relative to 30 flexion (full extensio n). Anterior/Posterior direction is represented by positive/negative values. Joint Posture Segment Lateral displacement (mm) Medial displacement (mm) 145 Flexion Anterior horn 12.9* (2. 7 ) 15.9 (6.4) Posterior horn 15. 1 (2.5 ) 14.1 (4.8 ) Meniscus c entroid 13.5 (2.7) 13.7 (4.8) Tibial External Rotation Anterior horn 0.6 ( 2.2 ) 3. 6 ( 4.2 ) Posterior horn 0. 3 ( 2.1 ) 4. 6 ( 4.0 ) Tibial internal Rotation Anterior horn 3.3 ( 2.9 ) 2. 6 ( 3.2 ) Posterior horn 2. 7 (2.8 ) 2.9 ( 3.5 ) p<0.05, less anterior ho rn than posterior horn displacement.

PAGE 50

50 CHAPTER 4 3 D SEMI AUTOMATIC SEGMENTATION OF THE MENISCUS USING STATISTICAL SHAPE MODELS Summary We present a segmentation scheme that automatically and accurately segments the canine menisci from proton density (PD) magnetic resonance (MR) images. Our segmentation scheme involves deforming a three dimensional (3D) statistical shape model (SSM) to fit the meniscus in the MR image using a nonlinear least squares optimization. The accuracy of this approach was experimen tally validated using a leave one out approach comparing automatically extracted meniscal models with models manually segmented by a trained human expe r t Experimental results showed mean absolute surface distance s (MASD) of 0.43 mm and 0.49 mm, and obtain ed average Dice overlap coefficients (Dice) of 81.1% and 81.9% for medial and lateral menisci, respectively. High correlations between automatic and manual segmentation were observed ( Pearson coefficient r was 0.94 and 0.95 and intraclass correlation coef ficient, ICC was 0.93 and 0.95 ) for both medial and lateral menisc i Introduction The menisci of the knee are crescent shaped collagenous structures contributing to load distribution, lubrication and stabi lity of the tibiofemoral joint ( Kurosawa et al. 19 80 ; McDermott et al., 2008; Mov et al., 2005 ) The menisci improve congruity and protect the articular cartilage by maximizing contact area and minimizing contact stress between the femoral and tibial condyles. Quantitative analy ses of the medial and lat eral meniscus from three dimensional (3D) magnetic resonance (MR) imaging offer opportunities to study meniscal structure and function, and may lead to improved measurements for injury diagnoses and treatment outcome assessments

PAGE 51

51 3D MR segmentation allow s enhanc ed measurement precision for investigating the 3D morphological and kinematic characteristics of the menisci compared to planar identification of isolated meniscal features ( Bowers et al. 2007 ; Swanson et al., 2010 ; Thompson et al., 1991 ; Vedi et a l., 1999 ) However, m anual segmentation is a time and expertise intensive process, which still has inter observer differences ( Siorpaes et al., 2012 ) Automation of 3D MR segmentation and analysis will provide greater reproducibility and facilitate new m easues Several semi automatic and fully automatic methods for the 3D segmentation of the menisci for humans have been reported using variety of different approaches ( Fripp et al., 2010 ; Paproki et al., 2014; Swanson et al., 2010; Yin et al., 2010 ; Z hang et al. 2013 ) However, the author s w ere not able to locate or obtain any of these programs (open source or commercial) for use at the time th is study was undertaken The objective of th is study was to develop and validate an automatic 3D segmentation sche me for canine menisci from MR images which would be made publically available for use in any future studies or further development and refinement Dogs were used because of clinical interest, ready availability of cadaveric materials, and extensive use as an experimental model to study meniscal tr eatments ( Bennett et al., 1991 ; Cook et al., 2001; King et al., 1936; Pozzi et al., 2010 ) The accuracy of the automatic segmentation scheme was validated using two distance based measurements : mean absolute surfac e distance (MASD) and Hausdorff distance (HD) ( Gerig et al., 2001 ) and three volume based measurements : sensitivity, specificity, and Dice overlap coefficient (Dice) ( Dice et al., 1945 )

PAGE 52

52 Materials and Methods The performance of the new method was tested u sing a leave one out format by comparing 3D meniscal models automatically extracted from in vitro MR images with 3D meniscal models manually segmented by a trained human expe r t. This study was approved by the University of Florida Institutional Animal Care and Use Committee (# 20130788). MR Image D atasets Sixteen pairs of menisci, from sixteen intact stifles from sixteen medium breed dogs (range 20 35 Kg) were imaged with 3.0T MR system (Siemens Verior, Siemens, Malvern, Pennsylvania, USA) in the sagittal plane using a proton density ( PD ) sequence (TR/TE/FA, 4210/25/90; FOV, 120mm; matrix, 512x512, slice thickness/gap, 2mm/0mm). The dogs were skeletally mature and without apparent signs of knee disease confirmed during anatomical dissection. A 15 channel tr ansmit/received knee coil was used for all knees. Menisci Segmentation The automatic segmentation method is based on a 3D active shape model (ASM), and involves deforming a 3D statistical shape model (SSM) of the medial or lateral meniscus to match the MR image s ( Cootes et al., 1995 ) A 3D SSM mathematically describes the surface shape variability of the training sets. These models are deform ed by varying a set of shape modes to modulate meniscal shapes within the range of those included in the training set T wo SSMs, one for the medial and one for the lateral meniscus, were trained based on the manual ly segment ed 3D MR image datasets. SSMs were developed using the method described in Baldwin et al and Fitzpatrick et al ( Baldwin et al., 2010, Fitzpatrick e t al., 2011 ) These SSMs were

PAGE 53

53 deformed in the MR image domain to match the estimated meniscal surface points using a nearest neighbor method and comparing one dimensional intensity profiles surrounding the prospective 3D model ( Friedman et al., 1977 ) Th e segmentation scheme involved four steps: (1) region of interest selection (ROI), (2) image preprocessing, (3) rigid body initialization, (4) SSM fitting (Figure 4 1). The method was implemented in a custom program ( MATLAB R2016a, The MathWorks, Natick, M A and Visualization Toolkit ( Schroeder et al., 1997 ) Region of i nterest selection In this first stage, ROI selection is done using the 3D MR volume. The mean SSM meniscal model is voxelized into a binary 3D image and superimposed on the MR volume displa y. The use rotates, translates and scales the mean SSM meniscal model into rough correspondence with the meniscal region in the 3D MR image, insur ing it covers a majority of the meniscus in the 3D MR image. O n average, ROI sizes were 114x210 pixels by 6 sl ices for medial meniscus, and 115 x 193 pixels by 7 slices for the lateral meniscus Each meniscus for each specimen was segment ed separately. MR image preprocessing After the ROI is selected, the MR images are converted into binary images with the threshho ld level equal to an approximat e mean meniscal grayscale value in the MR images. Although t he threshold level can be change d for volumes using different MR sequences a threshold level of 0.12 i s used for image preprocessing all menisc al MR images in this study Canny edge detection i s then applied to identify the surface edge of the meniscus in the ROI ( Canny et al., 1986 )

PAGE 54

54 Rigid body initialization The initialization stage roughly aligns the mean SSM meniscal model with the binarized MR image using volume moments to define centroids and principal axes. The r igid body initialization is accomplished as follows. 1. Perform a PCA on the estima ted surface edge locations determined from Canny edge detection performed in MR image preprocessing step to obtain the ce ntroid and first and second principal axes 2. Scale the mean meniscal SSM to match the spatial resolution of the MR images. 3. Calculate the centroid voxel location for the scaled mean SSM. 4. Perform a PCA on the scaled mean SSM to obtain first and second princip al ax e s. 5. Translate and align the scaled mean SSM position by match ing the ir centroids in the ROI. 6. Align the scaled mean SSM orientation by matching the first and second princip al ax e s in the ROI. 7. Perform rigid body registration using an iterative closest p oint (ICP) algorithm ( Zhang et al., 1994 ) to match the scaled mean SSM outer surfaces with the image edges SSM fitting The SSM is then deformed and fitted to the meniscus in the ROI in two steps. The SSM is deformed using the first 10 modes, with maximum variation 3 standard deviation, which presents 97.5% and 97.6% of the total variation for the medial and the lateral SSM. A n onlinear least square solver ( Matlab lsqnonlin ) is used to minimize the cost function ( Coleman et al., 1996 ) In step one the cost is the Euclidean distance between the aligned mean SSM surface and edge detected points in the ROI. The rigid body initialization nearest neighbor ICP, and model deformation steps a re repeated until one of the convergence criteria is met. All properties for lsqnonlin are set to default values, ex c e p t minimum

PAGE 55

55 change in variables for finite difference gradients (diffminchange) i s set to 0.12, maximum function evaluation s i s set to 1800, function tolerance i s set to 0.005, and the step tolerance i s set to 0. 005. In step two the cost function includes the Euclidean distance and voxel distance between the mean SSM and meniscus in the ROI. A 1D image gradient profile at each vertex normal to the surface of SSM model, is extracted. Each profile i s sampled 10 pi xels along a 3D vector and smoothed with moving average filtering (MATLAB) The distance between the maximum image gradient location and the mean SSM surface i s used as the voxel distance for minimization. The lsqnonlin solver i s used with the same setting s as step one, exce pt diffminchange i s set to 0.15. The rigid body initialization nearest neighbor ICP, and mode shape variation are again performed in sequence until convergence. T he final model is smoothed (Geomagic studio) to remove noise from the defo rmed surface. Manual segmentation All manual segmentation w as performed by a single observer using open source software (ITK SNAP, www.itksnap.org ) ( Yushkevich et al., 2006 ) Model post processing was done using the a lpha shape algorithm (MATLAB R2016a), and segmented contours were meshed with triangles to create smooth 3D models similar to the real meniscus (Geomagic Studio, Figure 4 2). Validation Validation of the automatic segmentation algorithm was done by using two distance based measurement MASD and HD ( Gerig et al., 2001 ) and three volume based measurements of sensitivity, specificity, and Dice ( Dice et al., 1945 )

PAGE 56

56 ( 4 1 ) ( 4 2) ( 4 3) ( 4 4) ( 4 5) where is true positive, is t rue negative, is false positi ve, and is false negative, and A and M are the automatic and manual segmentation results, respectively. The MASD and HD measure the average of the forward and backward Euclidean distance (d(x,y)), and the maximum of the HD(x,y) between automatic an d manual segmentations in millimeter s, respectively The sensitivity, specificity, and Dice measure the percentage of correctly detect ed meniscus, correctly reject ed adjacent anatom y and the spatial overlap in volume respectively T he t hree volume based measurements were calculated for comparative purposes because they were used for validation for meniscal segmentation accuracy in previous studies ( Fripp et al., 2010 ; Pa proki et al., 2014; Zhang et al. 2013 ) The validation calculation s were performed usi ng MATLAB, Geomagic S tudio, and Mesh Lab ( Cignoni et al., 2008 ) The Pearson product moment correlation coefficient (pearson) ( Lee et al., 1988 ) the intraclass correlation coefficient (ICC) ( Koch et al., 1982 ) regression analysis (r 2 ) and Bland Altman ( B land et al., 1986 ) analysis were calculated to investigate the correlation between the automatic and the manual segmentation data. All statistical analyses were performed using Windows Excel.

PAGE 57

57 Results There was good spatial overlap between the manual and au tomatic segmentation for both medial and lateral menisc i (Table 4 1). MASD was less than 0.5 mm HD was less than 1.9 mm and Dice was just under 82% for both menisci. There were strong correlations between the manual and automatic meniscal parameters w it h Pearson coefficient s of 0.94 and 0.95, r 2 values of 0.881 and 0 .898 and ICC values of 0.93 and 0.95 for medial and lateral menisc i, respectively Comparison using Bland Altman plots showed a relatively even distribution of the differences between automa tic and manual segmentation with a slight bias towards under segmentation, 3.54% and 1.80% for medial and lateral meniscus models respectively (Figure 4 2). Discussion The objective of this study was to develop and validate an automatic segmentation sche me for the menisci in the canine knee. Leave one out experiments showed good spatial agreement between the manual and automatic segmentations for the medial and lateral meniscus, with overall mean surface deviations of less than 0.5mm, and volume overlaps greater than 81%. Our scheme slightly out performed the automatic segmentation algorithm reported by Fripp J et al., (mean Dice of 75% and 77% for medial and lateral meniscus) and is similar to the results reported by both Zhang et al. 2013 ] (mean Dice of 81.9%) and Paproki et al. (mean Dice of 78.3% and 83.9%). Strong correlation was observed (r = 0.94 and r = 0.95) between automatic and manual segmentation for both medial and lateral menisc i The primary advantages of our method are: (1) it requires very little user manipulation to initialize the registration/estimation pipeline, (2) it provides accurate individually segmented model s of the medial and lateral menisc i that resemble the

PAGE 58

58 actual shape of the meniscus regardless of the quality of the MR images and ( 3 ) the method was implement ed and validated using a clinically relevant MR sequence that commonly is used in v eterinary medic al practice. No special scanners or signal sequences were required to obtain high quality results. There are some limitation s with this study. First, the method was only evaluated using a small number of specimens (knees) Statistical shape models can only capture variations that are included in the training set, so there is great advantage to having much larger training sets f or methodologic robustness and segmentation. Second, t he method was only validated using MR images with PD sequences F urther validation is required to assess accuracy on MR image s obtained with other signal sequences. Third, this method is no t fully automated. E ven though it takes less than a minute to do, the current implementation requires the user to select a ROI, isolating the meniscus from adjacent anatom y in the MR images. However, we believe that ROI selection is a minor limitation bec ause similar to p r evious work, the ROI could be eliminated by segmenting adjacent anatom y first ( bones and articular cartilage ) in a hierarchical segmentation scheme ( Fripp et al., 2010 ; Yin et al., 2010 ) Lastly, the method was only evaluated and validat ed for healthy knees. In future work, we plan to increase the number of subject s in the training set and evaluate the segmentation/modeling accuracy in pathological knees.

PAGE 59

59 Figure 4 1 Flow diagram of the automatic segmentation scheme, steps I through I V are taken to develop a meniscus model automatically. I (ROI): The user identif ies a region of interest with a mouse on the computer screen by drawing a box and manually overlaying the voxelized mean SSM images on time to further isolate the meniscus from the rest of the anatomy. This step is the only manual process in the automatic segmentation scheme. II, MRI preprocessing: This step is done automatically and converts the grayscale images from ROI to binary images. III, Rigid body initialization: Automa tically initialized the mean SSM to the meniscus in ROI. IV, SSM fitting: Automatically deforms the surfaces and fits the mean SSM to the meniscus in the ROI using nonlinear least square solver and the step III is performed at each step.

PAGE 60

60 Figure 4 2. The correlation and Bland Altman analyses performed for the medial and lateral meniscus volumes. The scatter plots on the left column present automatic segmentation results against manual segmentation results. The Bland Altman analysis on the right column presents the relative differences between automatic and manual segmentation. Table 4 1. Validation results of the automatic segmentation algorithm for the medial and lateral meniscus models. The mean and 1 standard deviations for volume based measurement are represented by percentage, while distance base d measurements are represented in millimeters. Meniscus MASD (mm) HD (mm) Sensitivity (%) Specificity (%) Dice (%) Mean SD Mean SD Mean SD Mean SD Mean SD Medial 0.43 0.10 1.84 0.47 82.8 3.85 98.3 0.53 81.1 3.03 Lateral 0.49 0.1 1.83 0.37 84.2 2.47 97.0 2.20 81.9 1.94

PAGE 61

61 CHAPTER 5 CONCLUSION Meniscal injury, repai r, replacement and regeneration remain topics of great interest to surgeons in both human and veterinary medical fields. Through the studies described with in this dissertation, we have provided a meniscal study framework by developing and validating an automatic image segmentation algorithm, and investigating the kinematics of the canine meniscus. The dog has been e xtensively used as an experimental model to study meniscal treatments and our preliminary studies of the meniscal kinematics show mechanics of direct relevance to canine clinical care and the use of dog knees as translational models for human medicine. New measurement technique s proposed in our study may be helpful in quantifying meniscal translations for all species with high tibial plateau angles, similar to dogs, to obtain better representation s of meniscal mechanics Future work include s application/ext ension of the automatic segmentation scheme to investigation of meniscal kinematics using a statistical shape plus motion model along with improv ing the robustness, accuracy and ease of use of our automatic segmentation scheme. In vitro studies are a good representation and in many case s done prior to in vivo studies but ultimately in vivo studies are required to obtain information that is representative of liv ing subjects. Several specific improvement s to our automatic segmentation scheme include (1) c onverting to a fully automatic segmentation, (2) increas ing the number of experimental subjects in the SSM, (3) evaluating and validating the automatic scheme for other MR images created by different MR sequences and (4) extending the model to shape plus motion ( Smoger et al. 2015 ) Additional projects will include investigating different statistical procedures

PAGE 62

62 like independent component analysis, kernel principal component analysis, and other nonlinear analys e s when building a SSM and adding more inform ation to the data to SSM like meniscal kinematics, image intensity, and using combined medial and lateral menisci. This dissertation provides the beginning of a line of research work on meniscal segmentation and mechanics. Future work will always be necess ary to improve the quality of the automatic segmentation and analysis tool s Despite the limitations of the current work our automatic segmentation scheme proved to be accurate in generating 3D meniscus models from MR images. Th is automatic segmentation m ethod efficiently provides accurate menisc al morphology and reduce s potential variability between investigators. O ur method will be posted open source so these new techniques and tools are available for use by researcher s in our Orthopedic Biomechanics Lab oratory at the University of Florida, and to other researcher s around the wo rl d.

PAGE 63

63 LIST OF REFERENCES Ahmed AM, Burke DL. 1983. In vitro of measurement of static pressure distribution in synovial joints Part I: Tibial surface of the knee. Journal of biomec hanical engineering. Aug 1;105(3):216 25. Anderson MW. 2002. MR imaging of the meniscus. Radiologic Clinics of North America. Sep 30;40(5):1081 94. Amadi HO, Gupte CM, Lie DT, McDermott ID, Amis AA, Bull AM. 2008. A biomechanical study of the meniscofemora l ligaments and their contribution to contact pressure reduction in the knee. Knee Surgery, Sports Traumatology, Arthroscopy. Nov 1;16(11):1004 8 Arnocz ky SP. 1999. Building a Meniscus: Biologic Considerations. Clinical orthopaedics and related research. 3 67 S244 S253. Arnoczky SP, Warren RF. 1983. The microvasculature of the meniscus and its response to injury: an experimental study in the dog. American Jour nal of Sports Medicine. 11,131 141. Baldwin MA, Langenderfer JE, Rullkoetter PJ, Laz PJ. 2010. Deve lopment of subject specific and statistical shape models of the knee using an efficient segmentation and mesh morphing approach. Computer methods and program s in biomedicine. Mar 31;97(3):232 40 Bennett D, May C. 1991. Meniscal damage associated with cruci ate disease in the dog. Journal of small animal practice. 32 (3), 111 117 Berjon JJ, Munuera L, Calvo M. 1991. Degenerative lesions in the articular cartilage after meniscectomy: preliminary experimental study in dogs. The Journal of trauma 31 (3), 342 350. Bland JM, Altman D. 1986. Statistical methods for assessing agreement between two methods of clinic al measurement. The lancet. Feb 8;327(8476):307 10. Blcker K, Guermazi A, Wirth W, Benichou O, Kwoh CK, Hunter DJ, Englund M, Resch H, Eckstein F, OAI inve stigators. 2013. Tibial coverage, meniscus position, size and damage in knees discordant for joint space narrowing data from the Osteoarthritis Initiative. Ost eoarthritis and Cartilage. 21(3):419 27. Bowers ME, Tung GA, Fleming BC, Crisco JJ, Rey J. 2007. Quantification of meniscal volume by segmentation of 3T magnetic resonance images. Journal of biomechanics.Dec 31;40(12):2811 5. Boxheimer L, Lutz AM, Treiber K, Goepfert K, Cro ok DW, Marincek B, Weishaupt D. 2004. MR imaging of the knee: position related changes of the menisci in asymptomatic volunteers. Invest Radiol 39:254 263

PAGE 64

64 Brantigan OC, Voshell AF. 1941 The mechanics of the ligaments and menisci of the knee joint. J Bone Joint Surg Am 23:44 66 Burr DB, Radin EL 1982. Meniscal function and the importa nce of meniscal regeneration in preventing late medial compartment osteoarthrosis. Clinical orthopaedics and related research. 171 121 126 Bylski Austrow DI, Ciarelli MJ, Kayner DC, Matthews LS, Goldstein SA. 1994. Displacements of the menisci under joint load: an in vitro study in human knees. J Biomech 27:421 431. Canny J. 1986. A Computational Approach to Edge Detection, IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. PAMI 8, No. 6, pp. 679 698. Cignoni P, Callieri M, Corsini M, Del lepiane M, Ganovelli F, Ranzuglia G. 2008. MeshLab: an open source mesh processing tool. In: Sixth Eurographics Italian Chapter Conference, pp. 129 136. Coleman TF and Y Li. 1996. An Interior, Trust Region Approach for Nonlinear Minimization Subject to Bou nds. SIAM Journal on Optimization Vol. 6, pp. 418 445. Cook JL, Tomlinson JL, Arnoczky SP, et al. 2001. Kinetic study of the replacement of porcine small intestinal submucosa grafts and the regeneration of meniscal like tissue in large avascular meniscal defects in dogs. Tissue engineering. 7 (3), 321 334. Cootes TF, Taylor CJ, Cooper DH, Graham J. 1995. Active shape models their training and application. Computer visio n and image understanding. Jan 31;61(1):38 59. Cullen KA, Hall MJ, Golosinskiy A. 2006. D ivision of Health Care Statistics. Ambulatory su rgery in the United States. Dice LR. 1945. Measures of the amount of ecologic association between species. Ecology. Jul 1;26(3):297 302. Englund M, Lohmander LS 2004. Risk factors for symptomatic knee osteoar thritis fifteen to twenty two years after meniscectomy. Arthritis & Rheumatism. 50 (9), 2811 2819. Englund M, Roos EM, Lohmander L S. 2003. Impact of type of meniscal tear on radiographic and symptomatic knee osteoarthritis: A sixteen year followup of menisc ectomy with matched control s. Arthritis & Rheumatism. Aug 1;48(8):2178 87 Eple r M, Sitler M, Moyer R. 2005 Kinematics of healthy and meniscal repaired knees. Research in Sports Medicine 13 (2), 91 109.

PAGE 65

65 Felix NA, Paulos LE. 2003. Current status of menisca l transplantation. The knee. Mar 31;10(1):13 7 Fithian DC, Kelly MA, Mow VC. 1990. Material properties and structure function relationships in the menisci. Clinical orthopaedics and related research. 252 19 31 Fitzpatrick CK, Baldwin MA, Laz PJ, FitzPatr ick DP, Lerner AL, Rullkoetter PJ. 2011 Development of a statistical shape model of the patellofemoral joint for investigating relationships between shape and function Journal of biomechanics. Sep 2;44(13):2446 52. Freeman MA, Pinskerova V. 2005. The move ment of the normal tibio femoral joint Journal of biomechanics. Feb 28;38(2):197 208. Friedman JH, Bentely J, Finkel RA. 1977. An Algorithm for Finding Best Matches in Logarithmic Expected Time, ACM Transactions on Mathematical Software 3, 209. Fripp J, C rozier S, Warfield SK, Ourselin S. 2010. Automatic segmentation and quantitative analysis of the articular cartilages from magnetic resonance images of the knee. IEEE transa ctions on medical imaging. Jan;29(1):55 64. Fox MG. 2007. MR imaging of the meniscu s: review, current trends, and clinical implications. Magnetic resonance imaging clinics of North America. Feb 28;15(1):103 23. Fu KS, Mui JK. 1981. A survey on image segmenta tion. Pattern recognition. Jan 1;13(1):3 16 Gerig G, Jomier M, Chakos M. Valmet 2001. A new validation tool for assessing and impr oving 3D object segmentation. International Conference on Medical Image Computing and Compu ter Assisted Intervention. Oct 14 (pp. 516 523). Springer Berlin Heidelberg. Gluer CC, Blake G, Lu Y Blunt BA, Jer gas M, Genant HK. 1995. Accurate assessment of precision errors: how to measure the reproducibility of bone densitometry techniques. Osteoporos Int. 4:262 270. Gupte CM, Bull AM, Murray R, Amis AA. 2007. Comparative anatomy of the meniscofemoral ligament i n humans and some domestic mammals. Anatomia histologia, embryologia. Feb 1;36(1):47 52. Guo H, Shah M, Spilker RL. 2014. A finite element implementation for biphasic contact of hydrated porous media under finite deformation and sliding. Proceedings of th e Institution of Mechanical Engineers, Part H: Jo urnal of Engineering in Medicine 0954411914522782.

PAGE 66

66 Hildreth III BE, Marcellin Little DJ, Roe SC, Harrysson OL. 2006. In vitro evaluation of five canine tibial plateau leveling methods. American journ al of v eterinary research. Apr;67(4):693 700. Hallen LG, Lindahl O. 1966. joint. A cta Orthopaedica Scandinavica. Jan 1;37(1):97 106. Han SH, Kim DI, Choi SG, Lee JH, Kim YS. 2012. The posterior meniscofemoral ligament: morph ologic study and anatomic classif ication. Clinical Anatomy. Jul 1;25(5):634 40. Hashemi J, Chandrashekar N, Gill B, Beynnon BD, Slauterbeck JR, Schutt RC, Mansouri H, Dabezies E. 2008. The geometry of the tibial plateau and its influence on the biomechanic s of the tibiofemoral joint. The Journa l of Bone & Joint Surgery. Dec 1;90(12):2724 34. Hoffelner T, Resch H, Forstner R, Michael M, Minnich B, Tauber M. 2011. Arthroscopic all inside meniscal repair Does the meniscus heal?. Skeletal radiology. Feb 1;40(2) :181 7 Iwaki H, Pinskerova V, Freeman MA. Tibiofemoral movement 1: the shapes and relative movements of the femur and tibia in the unloaded cadaver knee. Johal P, Williams A, Wragg P, Hunt D, Gedroyc W. 2005. Tibio femoral movement in the living knee. A study of weight bearing and non weight bearing knee MRI Journal of biomechanics. Feb 28;38(2):269 76. Jones SC, Kim SE, Banks SA, Conrad BP, Abbasi AZ, Tremolada G, Lewis DD, Pozzi A. 2014. Accuracy of noninvasive, single plane fluoroscopic analysis for measurement of three dimensional femorotibial joint poses in dogs. American journ al of veterinary research. May;75(5):477 85. Kawahara Y, Uetani M Fuchi K, Eguchi H, Hayashi K. 1999. MR assessment of movement and morpholog ic change in the menisci during knee flexion. Acta Radiol 40:610 614 Kessler MA, Glaser C Tittel S, Reiser M, Imhoff AB. 2006. Volume Changes in the Menisci and Articular Cartilage of Runners An In Vivo Investigation Based on 3 D Magnetic Resonance Imagin g. The American J ournal of Sports Medicine. 34 (5), 832 836. Kim SE, Jones SC, Lewis DD, Banks SA, Conrad BP, Tremolada G, Abbasi AZ, Coggeshall JD, Pozzi A. 2005. In vivo three dimensional knee kinematics during daily activities in dogs. Journa l of Orthopaedic Research. Nov 1;33(11):1603 10. King D. 1936. The healing of semilunar cartilages. The Journal of Bone & Joint Surgery. 18 (2), 333 342.

PAGE 67

67 Koch GG. 1982. Intraclas s correlation coefficient. Encyclopedia of statistical sciences. 4:212 7. Kurosawa H, Fukubayashi T, Nakajima H. 1980. Load bearing mode of the knee joint: physical behavior of the knee joint with or without menisci. Clinical orthopaedics and related resea rch 149 283 290. Laakso MP, Juottonen K, Partanen K, Vainio P, Soininen H. 1997. MRI volumetry of the hippocampus: the effect of slice thickness on volume formation. Magnetic resonance imaging 15 (2), 263 265. Lawrence RC, Felson DT, Helmick CG, Arnold L M, Choi H, Deyo RA, Gabriel S, Hirsch R, Hochberg MC, Hunder GG, Jordan JM. 2008. Estimates of the prevalence of arthritis and other rheumatic conditions in the United States: Part I I. Arthritis & Rheumatism. Jan 1;58(1):26 35 Lee JM, Fu FH 2000. The men iscus: basic science and clinical applications. Oper Tech Orthop 10:162 168. Lee Rodgers J, Nicewander WA. 1988. Thirteen ways to look at the correlation coefficient. The American Statistician. Feb 1;42(1):59 66. Logan M, Dunstan E, Robinson J, Williams A, Gedroyc W, Freeman M. 2004. Tibiofemoral kinematics of the anterior cruciate ligament (ACL) deficient multiple resonance imaging. The Americ an journal of sports medicine. Apr 1;32( 3):720 6. Lubowitz JH, Verdonk PC, Reid III JB, Verdonk R. 2007. Meniscus allograft transplantation: a current concepts review. Knee Surgery, Sports Traumatology, Arthroscopy. May 1;15(5):476 92 Matsuda S, Miura H, Nagamine R, Urabe K, Ikenoue T, Okazaki K, Iwamoto Y. 1998. Posterior tibial slope in the normal and varus knee. The Americ an journal of knee surgery. Dec;12(3):165 8. McCready DJ, Ness MG. 2016. Systematic review of the prevalence, risk factors, diagnosis and management of meniscal injury in do gs: Part 2. Journal of Small Animal Practice. Apr 1;57(4):194 204. McDermott ID, Masouros SD, Amis AA 2008. Biomechanics of the menisci of the knee. Current Orthopaedics 22 (3), 193 201. McGraw KO, Wong SP. 1996. Forming inferences about some intraclass c orrelation coefficients. Psychological methods 1 (1), 30. McInerney T, Terzopoulos D. 1996. Deformable models in medical image analysis: a surve y. Medical image analysis. Jun 30;1(2):91 108

PAGE 68

68 Mitchell SC, Lelieveldt BP, Van Der Geest RJ, Bosch HG, Reiver JH Sonka M. 2001. Multistage hybrid active appearance model matching: segmentation of left and right ventricles in cardiac MR images. IEEE Transa ctions on medical imaging. May;20(5):415 23 Mow VC, Huiskes R, Eds. 2005. Basic orthopaedic biomechanics & mech ano biology Lippincott Williams & Wilkins. Murphy L, Helmick CG. 2012. The impact of osteoarthritis in the United States: a population health perspective. AJN The Am erican Journal of Nursing. Mar 1;112(3):S13 9 Ocal MK, Sevil Kilimci F, Yildirim IG. 2012 Geometry of the femoral condyles in dogs. Veterinar y research communications. Mar 1;36(1):1 6. Paproki A, Engstrom C, Chandra SS, Neubert A, Fripp J, Crozier S. 2014. Automated segmentation and analysis of normal and osteoarthritic knee menisci from magn etic resonance images data from the Osteoarthritis Initiative. Ost eoarthritis and Cartilage. Sep 30;22(9):1259 70. Pena E, Calvo B, Mar tinez MA, Palanca D, Doblar M. 2005. Finite element analysis of the effect of meniscal tears and meniscectomies on human knee biomechanic s. Clinical Biomechanics. 20(5):498 507. Pham DL, Xu C, Prince JL. 2000. Current methods in medical image segmentation 1. Annual review of biomedical engineering. Aug;2(1):315 37 Pozzi A, Kowaleski MP, Apelt D, Meadows C, Andrews CM, John son KA. 2006. Effect of medial meniscal release on tibial translation after tibial plateau leveling oste otomy. Veterinary surgery. Jul 1;35(5):486 94. Pozzi A, Tonks CA, & Ling HY. 2010. Femorotibial contact mechanics and meniscal strain after serial menis cectomy. Veterinary Surgery. 39 (4), 482 488. Proffen BL, McElfresh M, Fleming BC, Murray MM. 2012. A comparative anatomical study of the human knee and six animal species. The Knee. Aug 31;19(4):493 9. Rankin M, Noyes FR, Barber Westin SD, Hushek SG, Seow A. 2006. Human meniscus assessment under weightbearing conditions. Am J Sports Med 34:98 107 Roos H, Laurn M, Adalberth T, Ro os EM, Jonsson K, Lohmander LS. 1998. Knee osteoa rthritis after meniscectomy: prevalence of radiographic changes after twenty one years, compared with matched controls Arthritis & Rheumatism. 41(4):687 93.

PAGE 69

69 Sabanci SS, Ocal MK. 2014. Lateral and medial tibial plateau angles in normal dogs. Veterinary and Comparative Orthopaedic s and Traumatology (VCOT). 27(2):135 40. Schroeder W Martin K, Lorensen B. 1997. An Object Oriented Approach t o 3D Graphics. Prentice Hall: Scholes C, Houghton ER, Lee M, Lustig S. 2015. Meniscal translation during knee flexion: what do we really know?. Knee Surgery, Sports Traumatology, Arthroscopy. Jan 1 ;23(1):32 40. Scorzin JE, Kaaden S, Quesada CM, Mller CA, Fimmers R, Urbach H, Schramm J. 2008. Volume determination of amygdala and hippocampus at 1.5 and 3.0 T MRI in temporal lobe epilepsy. Epilepsy research 82 (1), 29 37. Scotti C, Hirschmann MT, An ti nolfi P, Martin I, Peretti GM. 2013. Meniscus repair and regeneration: review on current methods and research p otential. Eur Cell Mater. 26:150 70. Seedhom BB, Hargreaves DJ. 1979. Transmission of the load in the knee joint with special reference to the ro le of the menisci part II: experimental results, discussion and conclusions Engineering in Medicine. Oct 1;8(4):220 8 Shefelbine SJ, Ma CB, Lee KY, Schrumpf MA, Patel P, Safra n MR, Slavinsky JP, Majumdar S. 2006. MRI analysis of in vivo meniscal and tibi ofemoral kinematics in ACL deficient and normal knees. J Orthop Res 24:1208 1217 Siorpaes K, Wenger A, Bloecker K, Wirth W, Hudelmaier M, Eckstein F. 2012. Interobserver reproducibility of quantitative meniscus analysis using coronal multiplanar DESS and I WTSE MR imaging. Magne tic resonance in medicine. May 1;67(5):1419 26. Smoger LM, Fitzpatrick CK, Clary CW, Cyr AJ, Maletsky LP, Rullkoetter PJ, Laz PJ. 2015. Statistical modeling to characterize relationships between knee anatomy and kinematics. Journa l of Orthopaedic Research. Nov 1;33(11):1620 30. Stone KR, Stoller DW, Irving SG, Elmquist C, Gildengorin G. 1994. 3D MRI volume sizing of knee meniscus cartilage. Arthroscopy: The Journal of Arthro scopic & Related Surgery. 10(6):641 4. Swanson MS, Prescott JW Best TM, Powell K, Jackson RD, Haq F, Gurcan MN. 2010. Semi automated segmentation to assess the lateral meniscus in normal and osteoarthritic knees. Ost eoarthritis and Cartilage. Mar 31;18(3):344 53. Thieman K, Pozzi A Ling C, Horodyski MB, Lewis DD. 2010. Comparison of contact mechanics of three meniscal repair technique and partial meniscectomy in cadaveric dog stifles. Vet Surg 39(3): 355 362.

PAGE 70

70 Thompso n WO, Thaete FL, Fu FH, Dye SF.1991. Tibial meniscal dynamics using three dimensional reconstruction of magnetic resonance images. Am J Sports Med 19:210 216. Tremolada G, Winter MD, Kim SE, Spreng D, Pozzi A. 2014. Validation of stress magnetic resonance imaging of the canine stifle joint with and without an intact crani al cruciate ligament. American jour nal of veterinary research. Jan;75(1):41 7. Tu dor F, McDermott ID, & Myers P. 2014. Meniscal repair: a review of current practice. Orthopaedics and Trauma. 28 (2), 88 96. Vangsness CT, Farr J, Boyd J, Dellaero DT, Mills C R, LeRoux Williams M. 2014. Adult human mesenchymal stem cells delivered via intra articular injection to the knee following partial medial meniscecto my. J Bone Joint Surg Am. 96(2):90 8. Vedi V, Williams A, Tennant SJ, Spouse E, Hunt DM, Gedroyc WM. 1999 Meniscal movement. An in vivo study using dynamic MRI. J Bone Joint Surg Br 81:37 41. Wilke VL, Conzemius MG, Besancon MF, Evans RB, Ritter M. 2002. Comparison of tibial plateau angle between clinically normal Greyhounds and Labrador Retrievers with and w ithout rupture of the cranial cruciate ligament. Journal of the American Veter inary Medical Association. Nov 1;221(10):1426 9 Wirth W, Frobell RB, Souza RB, Li X, Wyman BT, Graverand L, Hellio MP, Link TM, Majumdar S, Eckstein F. 2010. A three dimensional quantitative method to measure meniscus shape, position, and signal intensity using MR images: A pilot study and preliminary results in knee osteoarthritis. Magnet ic resonance in medicine. 63(5):1162 71. Yao J, Lancianese SL, Hovinga KR, Lee J Lerner AL. 2008. Magnetic resonance image analysis of meniscal translation and tibio menisco femoral contact in deep knee flexion. J Orthop Res 26:673 684 Yin Y, Zhang X, Williams R, Wu X, Anderson DD, Sonka M. 2010. LOGISMOS layered optimal graph image segmentation of multiple objects and surfaces: cartilage segmentation in the knee joint. IEEE transa ctions on medical imaging. Dec;29(12):2023 37. Yushkevich PA, Piven J, Hazlett HC, Smith RG, Ho S, Gee JC, Gerig G. 2006. User guided 3D active contour segm entation of anatomical structures: significantly improved efficiency and reliability. Neuroimage 31 (3), 1116 1128. Zhang K, Lu W, Marziliano P. 2013. The unified extreme learning machines and discriminative random fields for automatic knee cartilage and m eniscus segmentation from multi contrast MR images. Machin e vision and applications. Oct 1;24(7):1459 72.

PAGE 71

71 Zhang Z. 1994. Iterative point matching for registration of free form curves and surfaces. Internation al journal of computer vision. Oct 1;13(2):119 5 2.

PAGE 72

72 BIOGRAPHICAL SKETCH in m echanical e ngineering from University of North Carolina at Charlotte, NC, USA, in 2008 He joined the University of Florida in the fall of 2009, received his Master of Science (MSc) degree in mechanical engineering in 2011, and started to pursue for the Doctor of Philosophy (PhD) thereafter.