<%BANNER%>

Segmentation in 3D Virtual Spine Modeling for Assistance in Surgical Planning and Guidance


PAGE 1

SEGMENTATION IN 3D VIRTUAL SPINE MODELING FOR ASSISTANCE IN SURGICAL PLANNING AND GUIDANCE By MATTHEW JAMES WILLIAMS A THESIS PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLOR IDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF MASTER OF ENGINEERING UNIVERSITY OF FLORIDA 2005

PAGE 2

Copyright 2005 by Matthew James Williams

PAGE 3

ACKNOWLEDGMENTS There are many people that I am pleased to recognize who have had a very positive influence on my journey at the University of Florida. First, I would like to thank Dr. Frank Bova for his guidance and patience, and for making possible a very clinically interactive biomedical engineering experience. It is the kind of project I had hoped for when I began my studies at UF. Dr. Wesley Bolch and Dr. David Gilland both have my sincere appreciation for their commitment to come on board at such a late stage to help out. Literally, I could not have finished without them. Also, I would like to extend gratitude to Dr. Bernard Guiot, for introducing me to neurosurgery; Dr. Lionel Bouchet, for his tutelage and example of tireless dedication; Drs. Yunmei Chen and Bernard Mair, for lending their ears for mathematical reflection; and Drs. Christopher Batich and Anthony Brennan, for their continued support and words of motivation. In addition, I would like to acknowledge the financial contributions of the Department of Neurosurgery, the Department of Biomedical Engineering, and the Whitaker Foundation. I have encountered and have had the pleasure of knowing many friendly personalities in the lab, in the operating room, in the classroom, and in various other nooks and crannies, and each has helped to make this a memorable experience; and for that I give thanks. Finally, I would like to extend a heartfelt appreciation to Alan Wineman, Jennifer Kadlowec, and Jeanette Clute for their belief in my abilities; my fraternity brothers, for their companionship; and my exceptional friends Matt, JP, Vishal, Christina, Chris, Tom, and Kent for their encouragement and distracting adventures. And iii

PAGE 4

most of all, I am indebted to my brother, my parents, and the rest of my loving family as they are the foundation of my strength; and to Jennifer, the delightful blend of sunshine and spice that has helped me to believe in all of the wonderful things in life. iv

PAGE 5

TABLE OF CONTENTS page ACKNOWLEDGMENTS.................................................................................................iii LIST OF FIGURES..........................................................................................................vii LIST OF OBJECTS...........................................................................................................ix ABSTRACT.........................................................................................................................x CHAPTER 1 IMAGE-GUIDED SPINE SURGERY.........................................................................1 Preoperative Image Guidance.......................................................................................2 Intraoperative Imaging..................................................................................................5 Fluoroscopy...........................................................................................................5 Computed Tomography.........................................................................................6 Magnetic Resonance Imaging...............................................................................7 Ultrasound.............................................................................................................7 Computer-assisted Surgery Systems.............................................................................8 Virtual Surgical Model..........................................................................................8 Surgical Planning.................................................................................................10 Registration and Tracking...................................................................................11 Surgical Guidance...............................................................................................14 Surgical Application............................................................................................14 Systems................................................................................................................17 Segmentation Introduction..........................................................................................18 2 SEGMENTATION BACKGROUND........................................................................20 Image Processing........................................................................................................21 Mathematical Morphology.........................................................................................22 Binary Operations................................................................................................24 Grayscale Operations...........................................................................................28 Voxel-based Segmentation Methods..........................................................................29 Thresholding........................................................................................................30 Morphological Segmentation..............................................................................32 Edge-based Segmentation...........................................................................................34 Region-based Segmentation.......................................................................................39 v

PAGE 6

Shape-based Segmentation.........................................................................................40 Atlas Warping......................................................................................................40 Modeling..............................................................................................................41 Methods...............................................................................................................42 3 SEGMENTATION.....................................................................................................44 Problem Evaluation....................................................................................................45 Method Selection........................................................................................................47 Research Tools............................................................................................................48 Cursory Examination..................................................................................................50 4 REGION GROWING.................................................................................................53 Kernel Definition........................................................................................................53 Shape Correlation.......................................................................................................53 Hysteresis....................................................................................................................56 Algorithm Enhancement.............................................................................................59 Image Variations.........................................................................................................60 Prelabeling..................................................................................................................62 Instrumentation...........................................................................................................65 Results.........................................................................................................................65 5 CONCLUSIONS AND FUTURE WORK.................................................................69 LIST OF REFERENCES...................................................................................................73 BIOGRAPHICAL SKETCH.............................................................................................81 vi

PAGE 7

LIST OF FIGURES Figure page 1-1 Change in anatomical orientation occurring during transarticular screw placement...................................................................................................................4 1-2 Segmentation of a surgical model..............................................................................9 1-3 Surgical plan.............................................................................................................11 1-4 Registration of the surgical model to the patient......................................................12 1-5 Active surgical guidance..........................................................................................14 2-1 Intensity histograms and histogram equalization.....................................................21 2-2 Morphological structuring elements and neighborhood configurations..................24 2-3 Binary morphological dilation.................................................................................25 2-4 Compound morphological operations......................................................................27 2-5 Morphological distance transforms..........................................................................28 2-6 Threshold modeling..................................................................................................31 2-7 Binary image segmentation......................................................................................33 2-8 Singularity in contour evolution...............................................................................37 2-9 Level set application for contour evolution..............................................................38 3-1 Building a threshold model......................................................................................45 3-2 Manual segmentation and display............................................................................46 3-3 Surface characteristics of vertebrae..........................................................................47 3-4 Toolset GUI..............................................................................................................49 3-5 3D intensity gradient................................................................................................51 vii

PAGE 8

4-1 Manual kernel selection...........................................................................................54 4-2 Segmentation using distance maps...........................................................................55 4-3 High threshold values result in diminished kernels.................................................56 4-4 Region growing algorithm iteration.........................................................................57 4-5 Algorithm input processing sequence......................................................................58 4-6 Region growing........................................................................................................59 4-7 Open and close operators are used to alter the evolution of the growth algorithm..60 4-8 Unique properties of the input CT image may be isolated or combined to produce a more favorable segmentation...................................................................62 4-9 Segmentation results using variations on the input CT image.................................63 4-10 Instrumentation subtraction......................................................................................66 4-11 Segmentation failure: boundary leakage................................................................67 4-12 Segmentation failure: weak bone intensity profile..............................................68 5-1 Intensity/distance-to-boundary profiles....................................................................71 viii

PAGE 9

LIST OF OBJECTS Object page 4-1 High threshold movie (object1-highthreshold.mpg, 27.6 MB)................................56 4-2 Region growing movie (object2-success.mpg, 25.2 MB)........................................59 4-3 Boundary leakag e movie (object3-failureleakage.mpg, 25.8 MB)...........................67 4-4 Weak bone intensity profile movie (object4-failureweak.mpg, 28.6 MB)...............68 ix

PAGE 10

Abstract of Thesis Presented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Master of Engineering SEGMENTATION IN 3D VIRTUAL SPINE MODELING FOR ASSISTANCE IN SURGICAL PLANNING AND GUIDANCE By Matthew James Williams May 2005 Chair: Frank J. Bova Major Department: Biomedical Engineering Surgical procedures of the spine are technically demanding, requiring precise navigation to avoid critical vascular and nerve structures. Often, image guidance systems are employed to improve accuracy of surgical tool placement and increase the likelihood of a successful outcome. The most commonly used systems generate a patient-specific virtual 3D model. This model is used to create a surgical plan that, during the operation, safely guides the surgeon in placement of the surgical tools. Unfortunately, the current modeling technique limits the application of image-guided surgery. The typical surgical model is created using the thresholding technique performed on a pre-operatively acquired diagnostic CT image. The result is a single rigid body model that appears as a surface-rendered outline of the involved bones. During the procedure, this model is registered or aligned with the patient. The drawback to the rigid model is its inability to account for the mechanical flexibility of the spine. Since the patient is likely to change positions between image acquisition and placement for x

PAGE 11

surgery, errors are introduced into the guidance. In order for the surgical model to properly deform, the virtual bones must be segmented and individually registered to the vertebrae. The current method of segmentation is very time-consuming and not clinically feasible on a routine basis. The unfortunate consequence is that, given accuracy restrictions, image-guided procedures of the spine are generally limited to one or two vertebral levels. It is the purpose of this research to develop a method that provides a segmented model for use in image-guided surgery. It is desired that this method be easy to work with, provide swift results, and require minimal user intervention. The segmentation method developed here utilizes a simple region-growing scheme. It dilates seed regions that are uniquely assigned to the underlying bones. This allows the growing regions to retain their identity while expanding into the full bone profile. The method was tested on 31 high-resolution clinical scans. Over 80% of the resulting segmentations showed significant improvement over the standard manual methods. Success was gauged by the amount of time and effort required to achieve a segmentation using the method as compared to manual methods. The factors contributing to segmentation failures are attributed to poor bone resolution and inadequate starting information. This method showed marked success in the segmentation of the CT-derived models used in guided spine surgery. This method may be employed to significantly reduce segmentation time and facilitate surgical image-guided applications to multi-level spinal procedures. xi

PAGE 12

CHAPTER 1 IMAGE-GUIDED SPINE SURGERY Surgical procedures of the spine are technically demanding. They often involve placement of rigid instrumentation in close proximity to critical vascular and nerve structures and accordingly, demand a high degree of accuracy. In order to avoid complications, the neurosurgeon or orthopedic surgeon must accurately plan the trajectories for placement of this instrumentation. This requires visualization of the involved anatomy in three dimensions. Physical inspection allows the surgeon to make an initial assumption of specific anatomical orientation. However, imaging systems verify this assumption without unnecessary surgical exposure. In addition, they provide crucial information such as condition of vasculature or other hidden structures or extent of pathology such as fracture dimensions and tumor location. Image-guided surgery, at its simplest, is the use of an imaging modality to facilitate surgical intervention. Systems dedicated to image guidance, though, actually allow non-invasive evaluation and planning through a virtual surgical environment. In addition, during surgery, these systems direct the surgeons tools to the proper trajectories. The use of image guidance systems in surgery, through increased accuracy and precision, has reduced the risk of damage to critical areas and increased the ability of the surgeon to handle technically difficult surgeries. In addition to a reduction in the duration of the operation, these advancements have led to significant reductions in patient morbidity and mortality (Cleary, 1999). 1

PAGE 13

2 The most common image guidance systems utilize a model of the vertebrae generated from a diagnostic Computed Tomography (CT) image. It appears as a surface-rendered image of the bone profile. Unfortunately, this model is not able to deform, or account for any of the natural motions that may occur with the spine. Since motion of the patient is likely to occur after image acquisition, the model of the vertebrae constructed may not adequately reflect the orientation of the vertebrae presented during surgery. This limits the accuracy of the model and its utility in guiding surgical procedures. In order to create a spine model that can properly reorient or deform each of the vertebrae must be isolated. The process of identifying these individual vertebrae in the model is called segmentation. The current method of segmenting spinal models is a labor-intensive, time-consuming procedure, which is infeasible on a routine clinical basis. The purpose of this research is to create a more user-friendly method for segmenting the models intended for image-guided surgery. The goal is to provide results in a more suitable time frame while requiring minimal user intervention. Preoperative Image Guidance The standard of care is the use of a presurgical scan for image-guided surgery. Typically, plain radiographs are supplemented with an additional modality such as myelography, computed tomography, or magnetic resonance imaging (MRI). Each modality provides another layer of information that the surgeon can utilize. However, the use of both CT and MRI is discouraged because of cost constraints (Cleary, 1999). The preoperative scans that are at the surgeons disposal are generally used to mentally construct a 3D representation of the anatomy. This envisioned model is associated with the anatomy exposed during the operation. It allows the surgeon to proceed with more confidence without relying on any anatomical statistical convention. This is especially

PAGE 14

3 true in cases of severe pathologic deformation. Overall, the greater the knowledge of anatomical positioning, the less invasive a procedure must be to assess the correct orientation of the involved elements (Lavalle et al., 1996). The accurate determination of vertebral alignment is essential for a successful surgical outcome. Of the modalities available for presurgical imagery, CT and MRI are typically used for spinal applications. They provide multiple 2D slices that can be mentally stacked to form a 3D model. CT has excellent bone and soft tissue contrast, yet requires ionizing radiation (Cleary, 1999). MRI, on the other hand, provides soft tissue contrast without the ionizing radiation and bone flare usually associated with CT. They are both susceptible to artifacts, such as non-uniformity distortions in MRI and starburst patterns in CT. The primary drawback of preoperative imaging used for surgery is that it does not account for any anatomical alignment changes that occur after imaging. Any changes of positioning that occur detract from the accuracy of the representation. This change of positioning can be isolated from two sources. One is the movement of the patient between the presurgical scan and the operative positioning. The presurgical scan is taken in the supine position to minimize breathing artifacts. A process called reduction whereby the surgeon optimally orients the patient for the procedure determines the operative position. The second source of position change occurs during the operation. Throughout the procedure the vertebrae can experience significant surgeon-induced motion, especially in the case of spinal instability (Glossop & Hu, 1997). Patients with unstable spines may exhibit abrupt translations of spinal segments, and these movements are difficult to model and predict (Cleary, 1999). Accounting for these alignment changes is especially critical in cases where two vertebral bodies are to be fixed by the

PAGE 15

4 Table 1-1. Change of orientation between the vertebrae during a surgical procedure. Presurgical and postsurgical CT scans were compared, and the change of the positions of C1 relative to C2 was recorded. The translation and rotation between the two vertebrae in orthogonal planes are shown for three patients. Patient 1 Patient 2 Patient 3 AP Lat. SI AP Lat. SI AP Lat. SI Trans. (mm) 2.7 -6.1 3.7 -9.3 3.8 7.4 -6.7 0.7 10.0 Rot. (degree) 3.0 -5.4 -6.5 0.8 -14.0 9.7 -1.9 -18.0 2.4 Figure 1-1. Change in anatomical orientation occurring during transarticular screw placement. A profile of vertebral positioning is taken from the presurgical CT scan (left) and overlaid onto the postsurgical CT scan (right). A shift in position can be seen between C1 and C2. same screw (Figure 1-1). Table 1-1 quantifies the changes of position that occur after imaging by comparing preoperative and postoperative CT scans. Even though this particular example only involves two vertebral levels, it is apparent that the use of pre-operative imaging for image guidance has a limited accuracy. There are, however, two possible solutions for managing the problem of spinal motion after image acquisition. The first is intraoperative imaging and the second is the use of a guidance system that tracks each individual vertebral body.

PAGE 16

5 Intraoperative Imaging Imaging systems can be used to acquire scans during the operation and provide an instant indication of anatomical positioning and surgical tool location. Despite this advantage, however, there are certain considerations such as cost, intraoperative image quality, and impact on the operating room environment that must be considered before the investment in such a system. Fluoroscopy Fluoroscopy is the dominant intraprocedural imaging modality for spinal surgery. The unit is a common piece of equipment that easily fits around the table and provides minimum disruption to the operative field or operative technique. It is a low-cost, useful and familiar technology. It is also easily accessible and portable (Foley et al., 2001; Cleary, 1999). In the application of spinal surgery, fluoroscopy is commonly used to verify vertebral alignment and instrumentation and tool positioning. Fluoroscopy units provide high resolution, large field of view (FOV) scans in real time that have excellent bone to soft tissue contrast (Cleary, 1999). There are some drawbacks however, that limit the utility of x-ray fluoroscopy. The primary one is fluoroscopy only acquires a 2D image in one plane, and hence only gives one aspect of 3D positioning. It is impractical to gauge depth in the picture, as it is a display of overlapping tissues. In order to image in additional planes, the fluoroscope must be repositioned repeatedly (Foley et al., 2001). Also, fluoroscopic images have poor soft tissue discrimination, which makes the visualization of vascular and nerve structures difficult (Cleary, 1999). Finally, the capability of imaging in real time is at the expense of constant x-ray radiation exposure. In the case of spine surgery, where there may be long periods of imaging activity, there is

PAGE 17

6 significant radiation exposure to the patient and operating team (Foley et al., 2001; Rampersaud et al., 2000; Cleary, 1999). Computed Tomography Computed tomography is an imaging modality in which a three dimensional image is constructed from a series of plane cross-sectional radiographic images made along an axis. The scans are high resolution and have excellent bone to soft tissue contrast. Intraoperatively, CT provides an excellent view of bony structures and their relationships, and can accurately localize the tip of interventional instruments (Cleary, 1999). However, the imaging is not real time and subjects the patient and those in the scanning field to ionizing radiation. There are three main types of CT scanning systems that are used during surgery. The first is spiral CT. It is very fast with a large bore and excellent image quality, yet requires high capital and maintenance costs (Cleary, 1999). In addition, it is a large, fixed machine that has limited accessibility in a surgical environment. Mobile CT, on the other hand, is smaller, portable, and has a comparatively lower radiation dose. The disadvantages to this imaging system are slower acquisitions, decreased tube capacity, and lower image quality, which can lead to registration difficulties (Cleary, 1999). Mobile CT is costly as well. The last type of CT imaging systems uses a sweeping fluoroscope to generate images. This fluoro-CT has the advantage of quick reconstruction and display for 3D imaging as well as easy patient access and targeting. It offers a low cost, low patient dose alternative to spiral CT or mobile CT. However, fluoro-CT has comparatively poor tissue contrast which results in minimally acceptable bone reconstruction. Two examples of Fluoro-CT systems are the SIREMOBIL Iso-C3D (Siemens Medical Solutions USA, Inc., Iselin, NJ) and FluoroCAT (G.E. Healthcare

PAGE 18

7 Technologies, Waukesha, WI). Given these options, there is still one primary limiting factor in the use of intraoperative CT for spinal alignment determination. Scans from the typical surgical position are subject to breathing artifacts, and the resulting images are unsuitable for accurate image guidance. Magnetic Resonance Imaging Magnetic resonance imaging is a non-ionizing imaging modality that can be employed intraoperatively. The images are high-resolution with excellent soft tissue contrast. Safety considerations with the magnetic fringe field however, make it difficult to adapt to the surgical environment. The need for specialized tools and equipment that can operate in the intense magnetic field and radio frequency-rich environment make it a costly endeavor. This is in addition to the intrinsic cost of the device. Also, specific to guidance, there is insufficient tool tip viewing accuracy. MRI images have poor definition of bony structures, making edge interpretation difficult. The correct determination of bone boundaries is critical in spinal surgery. Finally, MRI has its own handful of potential artifacts that must be considered (Cleary, 1999). An example of an MR system optimized for surgical application is the Achieva I/T Interventional MR (Philips Medical Systems, Bothell, WA). Two other examples that allow intraoperative use are the Signa SP (G.E. Healthcare Technologies, Waukesha, WI) and the MAGNETOM Concerto (Siemens Medical Solutions USA, Inc., Iselin, NJ). Similar to computed tomography, this imaging modality is also susceptible to the breathing artifacts that plague scans performed in standard operative position. Ultrasound Ultrasound is an inexpensive, easily portable, real-time imaging modality that does not rely on ionizing radiation. However, the poor image quality, weak discrimination of

PAGE 19

8 certain critical spinal tissues, and reliance on operator skill, does not make it a candidate for precise surgical visualization. Also, the use of ultrasound for surgery often requires increased intrasurgical access, which should be avoided if at all possible (Cleary, 1999). This modality is mentioned because it has utility in being used for image to patient registration, a process necessary for computer-assisted surgery. Computer-assisted Surgery Systems The problem of realizing spinal motion after pre-operative image acquisition is a primary concern for accurate surgical intervention. In response, computer-assisted surgical guidance systems have been developed which actively track spinal movement and relay real-time positioning of the vertebrae to the surgeon. Unlike the intraoperative imaging systems, which typically operate over discrete periods of time with some degree of surgical field interruption, these systems relay positioning information throughout the critical portion of the surgical intervention. Image-guided surgical systems accomplish their task by overlaying a virtual surgical field with the patients operative field. The computer-generated virtual field is created preoperatively and includes a geometric model of the involved anatomy with visual indications of intended surgical pathways and targets of instrumentation. Once this virtual field is matched or registered intraoperatively to the physical surgical field, the surgeon can compare his surgical tool placement with the plan. The ability of the model to properly represent the spinal anatomy has a direct effect on the accuracy of the surgical guidance. Virtual Surgical Model The most preferred anatomical model used for spine surgery is one that best approximates the mechanical flexibility of the spine. The spine should be modeled as a series of rigid non-deformable bodies, vertebrae, connected by deformable tissue, muscle,

PAGE 20

9 Figure 1-2. Segmentation of a surgical model. A rigid model of the cervical vertebrae is created using the thresholding technique (left). An applied segmentation method produces a model that properly represents the anatomy (right). cartilage, and ligaments. The model used for guided surgery, however, only needs to include each individual vertebra as isolated bodies, so that they can slide, or translate and rotate relative to each other. Such a model best reflects the motion seen during the operation. The computer-based anatomical model is typically created from a presurgically acquired CT image. As indicated by Peters (2000), a high-resolution, three-dimensional image is best suited to clearly represent the patients anatomy. For example, a CT scan intended for use in spine surgery has a .6 to 1.2 mm slice thickness with a 0.2 0.3 mm in plane pixel size. These voxel dimensions (0.3 x 0.3 x 0.3 mm) equate to an approximate 18 cm field of view with 512 x 512 pixel images. Currently, thresholding is the scheme most commonly used to create the anatomic virtual model used in spine surgery. It is a convenient procedure that takes little of the clinicians time. Thresholding is a process that creates a boundary at disjoint intensity ranges such as between bone and soft tissue on CT images. It is easy to indicate the

PAGE 21

10 proper bone boundary using this method, however, it often leaves adjacent bones as one contiguous mass (Kikinis et al., 1998). The resulting single rigid body model, once registered to the patient, is ill suited to account for any movement of the spine that may occur during the surgical procedure, which potentially introduces errors into the surgical guidance. This challenges the efficacy of the guidance system and presents the need for a segmented model. The single rigid body model created through thresholding can be converted into a model that allows for flexibility. This process is called segmentation, or the division of the single spinal bone model into its constituent vertebrae (Figure 1-2). The resulting composite model, if properly matched to the patient, will allow the orientation of the vertebrae to be tracked intraoperatively. Currently, segmentation is a time-consuming process in which the clinician must use basic processing tools to manually contour the boundaries of each vertebra; there is to date no automatic, reliable, and robust method of doing so. Therefore, the development of real-time image processing techniques for model creation is of primary importance for image-guided surgery (Cleary, 1999). Surgical Planning Once an adequate virtual model is created, it can be used to aid treatment selection (Cleary, 1999). The surgeon will use the model to visualize inside structures, and establish optimal surgical pathways. Preparation involves the placement of virtual tools, which represent the actual instruments used in surgery. The surgeon can plan the incision, define resection margins, and determine the appropriate orientation for instrumentation (Welch et al., 1997). This includes identifying working corridors that provide adequate access while minimizing the risk of damage to fragile tissues (Cleary, 1999). An example of tool placement is shown in Figure 1-3. After the planning phase

PAGE 22

11 Figure 1-3. Surgical plan. Pedicle screw trajectory is planned using a virtual tool. Note the surgical model in the lower-right corner. of surgical treatment, the virtual model must be matched, or registered to the patient to allow for guidance. Registration and Tracking Registration is the mapping of coordinates between any two spaces. In a guided surgery application, it is the process that links the coordinate frame of the virtual computer field to the surgical field. Registration is done so the displayed computer model of the involved anatomy and overlaid virtual surgical tools accurately represents the physical operation. A popular system at the University of Florida accomplishes registration using a stereotactic, optical camera system. It tracks precise positioning of the anatomy and surgical tools in three-dimensional space using reference frames. These frames have attached light emitting diodes (LEDs) that allow the infrared (IR) camera to

PAGE 23

12 Figure 1-4. Registration of the surgical model to the patient. The reference frames attached to the probe and vertebra are used by the IR camera system to track stereotactic positioning. The surgeon uses a probe to touch various points on an exposed vertebra for surface registration (left). The corresponding points are reflected on the virtual model (right). track their position and pose. There are several steps taken to perform registration for this system. First, a reference frame is attached to an exposed spinal process and to each of the tools involved in the guided surgery. Since the thresholding process creates a single rigid-body model, only one vertebra can be registered. Therefore, for best utilization, the model is registered to the most significant or important vertebral level. Next, the computer-built virtual anatomic model must be registered or geometrically associated to the patients selected vertebra. Registration is required because the computer is unaware of the actual position of the anatomy relative to the attached reference frame. Registration for the tracked tools is much simpler because of predefined attachment sites for the reference arrays. The primary registration method for the spinal system is point matching. A set of points, usually landmarks, are identified on the virtual vertebrae and then matched to corresponding points on the patients vertebrae using a tracked probe. Once point matching is completed, and an estimate of the association between the surgical model and exposed vertebrae is achieved, an additional registration step, surface matching, may be

PAGE 24

13 employed. In order to surface match, the surgeon touches a multitude of points on the surface of the exposed vertebrae (Figure 1-4). These points are used to create a surface profile, which is then aligned to the surface of the virtual model. This step may improve the accuracy of the registration. Once registration is complete, whichever vertebra the reference frame is attached to is tracked dynamically to account for changes in the patients vertebral orientation. Registration assures that the virtual surgical tools, which correspond to the real tools, are properly overlaid upon the spine model as the operation is performed. Registration is usually limited to one vertebral level. This is due to both current modeling restrictions and the ability of the tracking system to easily recognize multiple rigid bodies. Therefore, to preserve accuracy, the use of surgical navigation is generally reserved for those procedures that only involve a couple of vertebral levels. An increase in availability of segmented models, however, will encourage registration and tracking of multiple vertebral levels. This will allow the benefits of surgical guidance to be applied to a broader scope of procedures. Alternate techniques exist for tracking and registration of the spine. One noteworthy tracking system uses electromagnetic (EM) markers to locate anatomy. These point localizers minimize signal blockage commonly associated with the cumbersome IR reference frames, but may be susceptible to EM interference and the presence of magnetic or ferrous objects. Registration, alternatively, can be accomplished with a localized imaging system. Specifically, ultrasound can be used to create a surface profile for use in the surface-matching method of registration. One novel development presented by Medtronic is the FluoroMerge software (Medtronic Surgical Navigation

PAGE 25

14 Figure 1-5. Active surgical guidance. A targeting screen (right) guides the surgeon to the proper trajectory of the surgical drill guide (left). Technologies, Louisville, CO). It can perform the registration of a preoperative CT image using only two fluoroscopic images. This not only provides a quick, hands-free registration method, but may also be useful for multiple rigid body registration. Surgical Guidance Once all of the steps of preparation (model creation, planning, registration) have been completed, the system is ready to provide the surgeon with the feedback necessary to perform a successful operation. A visual terminal in the operating room allows the surgeon to simultaneously view surgical tool placement relative to both the virtual model and the patient. Also included is a targeting screen that compares the trajectories of tools relative to the surgical plan (Figure 1-5). This feedback assures proper placement of the instrumentation during the surgical procedure. Surgical Application There are several benefits to the use of image guidance in surgery. First, it provides a multidimensional view of anatomic relationships in the operative field, including extent of bone and soft tissue resection (Welch et al., 1997). It is especially useful when traditional surgical landmarks are obscured or altered, as may be the case with pathology or bony fusion (Austin et al., 2002). Image guidance reduces the need for

PAGE 26

15 exposure and increases the confidence of the surgeon during the procedure (Welch et al., 1997). It also improves accuracy of the surgical tool placement, which leads to the reduced risk of structurally significant violations such as neurovascular injury. Ultimately, image guidance lessens the technical difficulty of spine surgery and allows for more safe and effective outcomes (Assaker et al., 2001; Ohmori et al., 2001; Youkilis et al., 2001; Welch et al., 1997). Currently, there are three general classifications of spinal surgery that take advantage of image-guidance. The first is decompression, or removal of anatomical pressure on the spinal cord. This category has the highest volume of cases. Stabilization is the category with the next highest volume of procedures. Surgical intervention is almost always required when instability occurs below the level of C2. The final group is deformity correction, which has the highest risk of undesirable outcomes (Cleary, 1999). The hard tissues manipulated in image-guided surgical procedures often include the vertebral body, facet joints, iliosacral joint, and intervertebral disc. The pathology addressed during the procedure varies from fracture to inflammation and may involve the spinal cord and other adjacent soft tissues (Cleary, 1999). The use of computer-assisted surgery has especially seen use in cervical procedures, given the need for a high level of accuracy. Vaccaro & Singh (2001) discusses various applications in cervical spine surgery. He states the primary use for computer-assisted surgery is placement of C2-C1 transarticular screws for atlantoaxial fusion in order to minimize the risk of injury to the vertebral artery. Other infrequent uses include transoral odontoid resections, cervical subaxial pedicle screw placement, and anterior cervical corpectomies.

PAGE 27

16 Spinal surgery often includes placement of instrumentation such as screws, rods, hooks and wires. Screws can be used alone to repair fracture or instability as in C1-C2 transarticular screw fixation, or as an anchor for further instrumentation as in the case of pedicle screw fixation. Pedicle screws are often used in the treatment of pathological conditions such as arthritic deformity (spondylosis), fractures (iliosacral or vertebral), and can be used to support bony fusion (arthrodesis) (Cleary, 1999). In certain levels of the spine, extremely high accuracy is needed for the placement of pedicle screws to avoid perforation of the pedicle wall (Rampersaud et al., 2001). In many cases, misplacement of the screw can result in vertebral artery injury (Weidner et al., 2000). Cervical procedures are even more technically difficult as the screw trajectory is in very close proximity to the spinal canal, vertebral artery, and spinal nerve root. Image-guided surgery has been shown to improve the placement of pedicle screws and reduce the risk of screw misplacement (Austin et al., 2002; Youkilis et al., 2001; Weidner et al., 2000; Amiot et al., 2000; Henderson et al., 1996). There are a few considerations in the use of image-guidance for spinal surgery. First, navigation systems eliminate the need for repetitive intraoperative fluoroscopy for tool placement, dramatically reducing radiation exposure (Foley et al., 2001; Welch et al., 1997). Conversely, the use of fluoroscopy or other intraoperative modality can be used to verify a CT-based image guidance system and avoid complications resulting from registration errors, modeling errors, shifting of the reference frame, or untracked intraoperative shifting of anatomy (Dickman, 2000). As far as how guided-surgery effects the overall duration of the surgery, there are arguments for no appreciable effect

PAGE 28

17 and increased time consumption (Assaker et al., 2001; Weidner et al., 2000; Henderson et al., 1996). Systems There are a number of image guidance systems that are used in spine surgery. Each of them relies on the data from an imaging modality and a tracking system to relay vertebral orientation to the surgeon. Medtronic offers three configurations that use IR tracking, yet utilize different imaging modalities (Medtronic Surgical Navigation Technologies, Louisville, CO). The first, the StealthStation, uses a traditional CT-derived model to perform navigation. The second system uses 3D fluoro-CT images by interfacing the StealthStation with a Siemens SIREMOBIL Iso-C3D, an isocentric, automated fluoroscopy system (Siemens Medical Solutions USA, Inc., Iselin, NJ). Fluoroscopic-CT images generally have excellent spatial accuracy, yet their poor image contrast results in a 3D image reconstruction with poor tissue resolution and differentiation (Foley et al., 2001). These qualities hinder the construction of suitable virtual models for use in surgical guidance. The third pertinent system from Medtronic is the FluoroNav virtual fluoroscopy system. It facilitates real-time navigation using C-arm fluoroscopy. This system employs a two-plane display, yet the drawback is similar to standard fluoroscopy in that 2D views do not give an accurate appreciation of 3D anatomy. G.E. Healthcare Technologies (Waukesha, WI) offers surgical guidance system configurations that use EM tracking, such as the OEC 9800 FluoroTrakTM, which couples a high-resolution fluoroscopic imaging system with surgical navigation. A software upgrade to the FluoroTrakTM Surgical Navigation allows the use of 3D Fluoroscopic-CT images.

PAGE 29

18 In addition to CT and Fluoro-CT, there are 3D image-based guidance systems that use MR technology. An example is the PoleStarTM Intraoperative MR Image Guidance System (Odin Medical Technologies, Inc., Newton, MA). It uses an infrared navigation system similar to CT-based systems; however, this system allows intraoperative image acquisition. The primary drawbacks to use of this system are very high cost and small size which prohibits spinal operations on the trunk. Considering the options for tracking intraoperative spinal motion during surgical procedures, the cost of intraoperative MR and CT, the difficulty in resolving breathing artifacts, the lack of dimensionality of fluoroscopy, and poor 3D image construction from fluoro-CT, it is most clinically appealing to perform surgical guidance using a navigation system that utilizes preoperatively acquired CT images. These images provide a suitable platform from which to create an accurate anatomical model, plan the surgery, and program the navigation. However, the primary obstacle is that multi-level spinal procedures require the registration of a segmented model to be accurate. Segmentation Introduction Segmentation, as it is considered in medical image application, is the division of an image into anatomically labeled sections. A clinician experienced in the modality can easily recognize and outline the pertinent anatomy, however, a similar computational recognition or identification scheme is challenging to develop. Algorithms that automatically segment images have the potential to significantly reduce involvement of the clinician, encouraging the beneficial application of segmented models. These models can not only be used for visualization, surgical planning, and surgical navigation, but as a medium to study structural information particular to the modality.

PAGE 30

19 In order for a segmentation algorithm to be successful, many factors must be considered. First, the time of the clinician is at a premium, so the algorithm must operate quickly (in accordance with current computational technology) and with minimum user input. It also must be flexible to adjustment; as such models will always be subject to the fine-tuning of a trained eye. Also, the algorithm must be accurate and robust despite the varied pathologies presented such as crushed vertebrae, fractures, scoliosis, and involvement of tumors. Artifacts are problematic to any imaging modality, and a segmentation algorithm must be tolerant to these instances. In the specific case of CT scans, instrumentation can cause such artifacts. Finally, the performance of the algorithm must be consistent and reliable, optimally eliminating the variability introduced through human error. These parameters, combined with a thorough review of methodology, will appropriately guide research of automatic segmentation algorithms. Image segmentation is acknowledged as the most difficult and prohibitive step in the modeling of anatomical data. Despite this impediment, segmentation of images has a clear benefit to spinal surgery. The proper registration of an extended-level surgical plan to a patient diminishes the inaccuracies introduced during intraoperative movement. It is the focus of this research to develop and test methods to automate the process of vertebral segmentation of 3D spine models constructed from CT image slices. This will encourage image-guided spine surgery involving multiple vertebral levels, ultimately improving patient outcomes.

PAGE 31

CHAPTER 2 SEGMENTATION BACKGROUND Segmentation is the division of a whole image into a subset of connected pixel regions that have some common property. Segmentation in medical image processing, however, includes the implied step of anatomical labeling. That is the association of the image information to the appropriate anatomy for analysis. Computer algorithms that accurately segment medical images can be challenging to develop. However, there are many potential benefits such as enhanced visualization and ability to perform complex surgical planning and navigation procedures. There are several approaches to the problem of medical image segmentation. Each of them relies on different information and can give somewhat different results, depending on the application. The first, voxel classification, uses globally defined characteristics to determine segmentation. For example, the use of an intensity threshold to identify bony anatomy qualifies as voxel classification. The next approach to segmentation involves the creation of a boundary concept that is used to mark the division between regions. This concept relies on the identification of discontinuities between regions. The contrast to this edge-based segmentation is region-based segmentation. This approach uses regional characteristics such as common intensity patterns to identify clusters. The final approach to image segmentation is the use of deformable atlases. The atlas is a general model that is associated to a particular anatomic structure. To perform segmentation, the general atlas is deformed to match an 20

PAGE 32

21 Figure 2-1. Intensity histograms and histogram equalization. An intensity histogram (upper right) is calculated from the sample image (upper left). The horizontal axis shows grayscale value and the vertical axis shows frequency. Histogram equalization improves the contrast over the most densely populated intensity ranges. A histogram equalization (lower right) is performed on the sample image (lower left). individual image. Often, these techniques for medical image segmentation can be integrated in an algorithm to produce a more favorable result. Image Processing Certain methods can be used to extract information from an image for use in a segmentation algorithm or to enhance an image to better suit a segmentation algorithm. Examples include histogram equalization, boundary detection, and filtering. An intensity histogram of an image is a frequency distribution of pixel intensities. The resultant graph indicates which intensity ranges have the most pixels (Figure 2-1). One use of the information provided in a histogram is to allow contrast normalization. This normalization process, called histogram equalization (Figure 2-1), maps pixels to

PAGE 33

22 new intensity values to approximate a flat histogram. This process creates no new intensity values. Boundary detection is a feature detection method used to identify boundary pixels in an image. These boundary pixels correspond to the region between different tissue types and are usually defined by a high intensity gradient (Roberts, 1965; Sobel, 1970; Prewitt, 1970). The gradient of a three-dimensional image with image intensity I is defined as: ),,(),,(),,(),,(zyxIzkzyxIyjzyxIxizyxI Various tracking systems can use the gradient information to make an edge trace. However, ambiguous or discontinuous edge data can produce errors. A drawback to the use of gradient information is that the image often must be smoothed because gradient calculations are susceptible to noise. Filters are used to enhance the qualities of an image in accordance with some desired characteristic. In one instance, multi-dimensional adaptive filters are used to resample the image data to reduce partial volume effects and noise. They also handle the low off-plane resolution of CT images (Westin et al., 2000). Another filter of note is specifically designed to allow for a more robust image segmentation for use in guided surgery. It proposes to enhance separation of joint spaces in a CT scan, while allowing the retention of important edge information (Westin et al., 1998). Mathematical Morphology Mathematical morphology is a geometrical approach to signal processing (Matheron, 1975; Serra, 1982). It performs many image processing tasks using object quantification and easily deals with attributes such as shape and size, connectivity, and

PAGE 34

23 contrast. Also, edge information is preserved during boundary manipulation. Common applications include noise reduction, texture analysis, and shape changing such as thickening, pruning, or skeletonization. These characteristics make mathematical morphology well suited to the task of image segmentation. Mathematical morphology uses set theory as a foundation for many of its functions. In accordance, its primary operations use binary structures which are defined with an object and complementary background: Object: TRUEapropertyaA)( Background: AaaAC A common technique for the creation of such a structure is the threshold mask. The threshold mask is a binary overlay that indicates which voxels in a three-dimensional image I are included within a prescribed intensity range MT 10,tt as shown: 10,,,,,1,,,ttgIgkjikjiTM Simple operations such as reflection and translation form the basis of more complex set functions, and can be applied to this binary image structure. Reflection of set A is indicated by A : AaawwA, In other words, A is the set of elements, w, such that w is formed by multiplying each of the coordinates of the elements, a, of set A by Note that the elements or voxels in an image are considered vectors, so in a three-dimensional image in Z3 space Another basic set operation is translation: 321,,aaaa AaxaccAx,

PAGE 35

24 111111111 Figure 2-2. Morphological structuring elements and neighborhood configurations. A flat 3x3 structuring element (FSE) (left) is equivalent to a pixel with its eight nearest neighbors (8NN configuration) (middle). On the right is a structuring element composed of a pixel and its four nearest neighbors (4NN). The equation says that the shifting of set A by vector x is the set of pixels c such that c is equal to a+x for all pixels that are a member of A. Binary Operations Two operations, dilation and erosion, are part of a core of image processing algorithms used in mathematical morphology. They produce a result by passing a structuring element over the image. This structuring element is analogous to the convolution kernel used in linear filter theory and it must have the same dimension as the image. Technically, in this application, the structuring element can be considered an image. For example, in an algorithm for a two-dimensional image, a flat structuring element (FSE) is commonly used. It has each element in the structuring array set to true or (Figure 2-2). Other types of structuring elements can be defined according to their neighborhood configuration. For example, equivalent to the 3x3 FSE is a structure consisting of a pixel and its adjacent horizontal and diagonal pixels. These adjacent pixels are known as the eight nearest neighbors (8NN). Another common structure is a central pixel with its adjacent orthogonal pixels or four nearest neighbors (4NN). These neighborhood configurations are shown in Figure 2-2. Nearest neighbors are defined by their connectivity patterns; the arrangements that determine if adjoining pixels are part of the same object. In the 4NN structure example

PAGE 36

25 = Figure 2-3. Binary morphological dilation. A binary image (left) dilated by a structure (middle) results in an expanded object (right) above, all of the four nearest neighbors are 4-connected to the central pixel. On the other hand, pixels are 8-connected if they are connected in the orthogonal or diagonal direction, as in the 8NN structure. Connectivity arrangements for three-dimensional structures include: 1. 6-connected: voxels are connected by their faces 2. 18-connected: voxels are connected by their faces or edges 3. 26-connected: voxels are connected by their faces, edges, or corners One of the simplest morphological functions to implement is dilation. It serves to expand the binary image structure and is analogous to convolution (Figure 2-3). In typical notation, A is the image set and B is the structuring element. The set theory formula for dilation is given by: })({AABzBAz That is, all coordinates z where the translation of the reflected structure B by z intersects with the binary image A. The implementation of dilation on a computer, also known as Minkowski addition, is in a slightly different form. It is the union of the sets where the

PAGE 37

26 binary image is translated according to each of the points in the dilating structure: BbbABA Erosion, another simple morphological image processing operation has the effect of shrinking a binary object. The set operation for erosion as well as Minkowski subtraction is given by: A ABxBx)(: BxxA The set equation simply states that the erosion of A by B is the set of points x such that B translated by x is contained in A. The Minkowski variant computes the erosion by taking the intersection of all of the sets from the result of A translated by each element of B. The two simple morphological operations dilation and erosion can be combined in series to form the compound morphological operations open and close. The open operation serves several functions such as smoothing of object contours, breaking of narrow isthmuses, eliminating thin protrusions, and acts as a filter to remove background noise: ABA( BB) The close operation, on the other hand, fills gaps in contours, fuses breaks, eliminates small holes, and acts as a filter to remove foreground noise: )(BABA B Opening and closing have the advantage of being idempotent, which means that repeated applications will not further change the signal. Examples of opening and closing are shown in Figure 2-4. There are a handful of other binary morphological operations that are useful for segmentation. One is boundary extraction. A boundary is extracted from a binary object

PAGE 38

27 Figure 2-4. Compound morphological operations. Comparison of open (middle) and close (right) morphological operations performed on a binary image (left) using a 3x3 Flat Structuring Element. by subtracting the result of an erosion operation performed on that object. The connectivity of this boundary is determined by the structure used for the erosion. Region filling is another useful operation. It utilizes the background or complement of a binary image as the area to be filled. A point X0 is selected in the background and repeatedly dilated until it achieves the desired fill result: ckkABXX1 This technique is especially successful for filling internal gaps. Finally, distance transforms can be integrated into mathematical morphology (Cuisenaire, 1999). Distance transforms create distance maps from binary image objects (Figure 2-5). Distance maps are images in which the intensity of a pixel p is an indication of proximity or nearest distance to the object O: OqqpdistpDM ),,(min)( Distance maps have a smooth surface and an even gradient making them desirable as shape representations. In addition, distance maps have an interesting property. The result of an object dilated by a spherical structure can be expressed as the threshold of a distance map of that same object. B is a spherical structure with radius d, created by the

PAGE 39

28 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 0 0 0 0 1 1 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 2.2 2 2 2.2 2.8 3.6 1.4 1 1 1.4 2.2 3.2 1 0 0 1 2 3 1 0 0 1 2 3 1.4 1 0 1 2 3 2.2 1.4 1 1.4 2.2 3.2 Figure 2-5. Morphological distance transforms. A Euclidean distance transform is used to form a distance map (right) from a binary image object (left). The pixel intensity, as indicated by the numerical value, relays the distance of the pixel to the object. selection of all points that are less than or equal to a certain distance d away from point (0,0,0): dbdistbBM))0,0,0(,( And an object X dilated by the structure B is equivalent to the threshold of the distance transform DT(x) of the object at distance d: dxDTxBX)( Grayscale Operations Binary morphological image processing methods can be extended to grayscale (Bangham & Marshall, 1998). Dilation and erosion become useful filtering functions. Dilation extends an object by using the maximum filter to remove low-valued regions. Accordingly, erosion contracts an object by using the minimum filter to remove high valued regions. These operations both have the effect of smoothing an image. Grayscale morphological dilation assigns to each pixel the maximum of the sum of the local region and the structuring element. In the special case that this structuring element contains all

PAGE 40

29 zeroes, this operation is equivalent to the maximum filter: kzjyixBkjiABA,,),,(max And the dual operation for erosion using the minimum operator: kzjyixBkjiABA,,),,(min These operations, as with their binary analogues, can also be combined to form the compound operations open and close. Opening acts as a high intensity point filter and closing acts as a low intensity point filter. Another useful grayscale operator is the morphological gradient. Given an image I, the morphological gradient is given by the difference between the respective dilation and erosion: )()()(BIBIIg This intensity gradient is very useful for boundary extraction. The advantage of using these morphological operations is that they provide useful image-processing features that are easy to implement, and integrate into a segmentation routine. Voxel-based Segmentation Methods The simplest of the segmentation methods utilizes global image information to assign the memberships of voxels into particular anatomical regions. Boundaries of regions are then implicitly determined from a complete label map. Given a set of anatomical structures contained in an image {1k}, a label map L(x) overlays the image I(x) and L(x)= i where i represents the anatomical structure at I(x). The assignment of voxels takes into account such factors as intensity value, neighboring pixel classification, and relative distance of neighboring pixels. However, if the chosen factors result in an overlap of neighboring anatomical structures, the global representation makes

PAGE 41

30 additional helpful information such as shape and geometric relationships difficult to incorporate. On the other hand, voxel methods are quickest in speed and do not require a cumbersome training model to achieve a result. Thresholding Thresholding is a simple, extensively used image processing technique that isolates a region of an image based solely on intensity criteria. It is a computationally inexpensive and fast low-level technique. In addition, the result may be used as an input to a higher-level segmentation model. It is defined as: ),,( kjiG 1 for TkjiI ),,( 0 for TkjiI ),,( G is the resultant three-dimensional binary image of a grayscale image I thresholded at intensity value T (Figure 2-6). Semithresholding is a similar technique which masks out the image background leaving gray level information present in the objects: ),,(kjiG),,(kjiITkjiI),,(TkjiI for 0 for ),,( An upper bound may also be incorporated into these thresholding forms. It serves to isolate a specific intensity band. Overall, this class of voxel methods is typically applied in cases where particular anatomy or tissue can be identified within a certain intensity range. The most apparent example is the segmentation of bone, where all included voxels are for the most part at a higher intensity than adjoining tissues. The selection of a threshold value can be done through manual inspection or probability estimation. To perform the estimation, a tissue model is created that predicts which voxels belong to which structures based on probability distributions of intensities. The probability of

PAGE 42

31 Figure 2-6. Threshold modeling. A binary image (upper right) is created by thresholding the sample image (upper left). The red line on the histogram of the sample image (bottom) indicates the threshold intensity value. Each of the pixels with an intensity value higher than the threshold value is highlighted in the binary image. intensity value x, estimated at each tissue class I is based on a set of training data )(ixP Histograms are commonly used to collect training data for this technique. Care must be taken in application of these voxel methods if intensity ranges are not disjoint. For example, in spinal CT images, it is not uncommon to have certain intensity values coinciding with both bone and cartilage or ligament. Also, thresholding or tissue class estimation is susceptible to imaging noise and artifacts that cause intensity overlap and unclear boundaries between tissues. A popular method for probability estimation is called Expected Maximization (Dempster et al., 1977). It is especially effective at handling incomplete data.

PAGE 43

32 Tissue class estimators can actually use criteria other than intensity to perform segmentation of tissues. One example, which makes a classification according to regular neighborhood intensity patterns, uses a technique called statistical clustering (Leahy et al., 1991). Texture patterns are typically measured though a co-occurrence of distance and intensity. Morphological Segmentation Methods in mathematical morphology can be used to segment an image (Dougherty, 1993). And although morphology is a language for shape representation and manipulation, its basic segmentation methods can be considered voxel-based. The first concept in morphological segmentation is seeded region growing. A seed is a point voxel or a group of voxels. The seed is expanded by checking to see if neighboring boundary voxels are within specified criteria. A common criterion is if the absolute value of the intensity difference of a seed and its neighboring voxel is beneath a threshold. A very simple version of region growing is found in the watershed transform (Vincent & Soille, 1991; Beucher & Meyer, 1992). The watershed transform is a segmentation method for grayscale images. It interprets the topology of an image and assigns watershed lines to boundaries between catchment basins. Catchment basins are the areas of local minima on the topographical map and are analogous to the depressions that would collect drops of water. The watershed lines are the crests between these basins. The common input to the watershed transform is a morphological gradient of an image so that watershed lines would correspond to the areas of strong edge evidence and divide the original image into homogonous regions. In application, however, simply taking local minima can result in oversegmentation, especially if there is a large noise contribution from false minima.

PAGE 44

33 Figure 2-7. Binary image segmentation. The distance transform can be used to create a topographical image (center) from a binary image (left). The binary image is a silhouette of three overlapping shapes. Utilizing the topographical image, the watershed transform can then be called to approximate a segmentation (right). The problem of oversegmentation can be handled in a couple of ways. The first is to apply a merging scheme, which combines adjacent areas according to some guideline. This guideline is usually based on statistical gray level properties. The second way to deal with an oversegmented image is to use marker selection. First, a new analogy should be introduced, and that is the gradual flooding of the image topography using a rising water table. Given this, marker selection is the decision of which local minima or seed regions will flood (Beucher & Meyer, 1992; Meyer & Beucher, 1990). The result is regions that are selected will flood the regions that are not, eliminating oversegmentation. Caution must be used in attempting to use smoothing operators or other such filters to reduce the false minima from noise contributions as they have the potential to remove areas of strong edge information. In modern application, the watershed transform has been used to segment out maxillofacial bone in CT (Bhm et al., 1999). The cited method also uses a tissue classification scheme to label the segmented regions. The watershed transform has applications on binary images (Figure 2-7). It can use a

PAGE 45

34 topographical image created by the distance transform to separate overlapping objects (Cuisenaire, 1999). The last morphological medical image segmentation method of note uses Recursive Erosion (RE) and Geodesic Influence (GI) (Kaneko et al., 2000). The recursive erosion is used generate candidate seeds while the geodesic reconstruction recovers the separated organs. Edge-based Segmentation Some segmentation problems can be solved using boundary localization. This concept involves the creation of an edge or surface model that is designed to converge on an object boundary. An accurate convergence will then describe the border of the segmented object. Given a three-dimensional image, closed surfaces are defined {S1Sk} with all points inside surface Si corresponding to an anatomical structure I. Also, all points that represent i are contained by Si. The information for the model is gathered by relating the edge representation to its associated image information. This includes image gradient, texture discontinuities, or any other useful measure that can be geometrically associated with the boundary. One indirect scheme for the segmentation of 3D objects involves the unification of multiple segmentations in two dimensions. An example is the segmentation of bone in CT images. Once the proper segmentation of one image slice is accomplished, it can be used to direct the result of adjacent slices. The combination of all of these individual slices results in a segmented three-dimensional image. Active contours are dynamic deformable models used for edge-based segmentation (Blake & Isard, 1998; Caselles et al., 1993; Malladi et al., 1995; Tek & Kimia, 1995).

PAGE 46

35 One such model is called the snake (Kass et al., 1987). It makes use of an energy function to guide the evolution of the boundary. The snake itself is a spline function parameterized by a set of node points. The cubic polynomial is a common choice for the spline function: xxxxducubuaux23)( yyyyducubuauy23)( The energy equation is then used to direct the movement of the nodes. The terms in the equation are a balance of forces that pull the spline toward the desired edge features. A proper segmentation is achieved through the minimization of this energy equation: ][intintconstraimageernaltotalEEEE This equation includes a term for internal energy, image energy and constraint energy. Internal energy is solely dependent on the shape of the spline. It includes parameters for stretching and flexing, which are optimized according to the geometric knowledge of the object to be segmented. The image energy is based on the image values along the path of the spline and can include attraction to strong gradients or regions of light or dark intensity. Constraint energy is intended to capture the higher level knowledge about the image and features. Example constraint energy terms are manually defined attraction or repulsion fields such as volcanoes and springs. A specific snake model designed for image segmentation was introduced by Kass et al. (1987): dqqCIdqqCCE))(()(')(2 It includes internal and external energy terms for the curve C. Internal energy is a regulating force that keeps the curve smooth by penalizing high curvature. This also

PAGE 47

36 makes the curve robust against noise. External energy is image energy and is designed to act as an attraction force to high gradients. The coefficients of these terms are empirically adjusted according to the specific application. One consideration in the optimization of snakes is the dynamic adjustment of nodes. If the number of nodes remains consistent, then snake expansion will result in decreased resolution and snake contraction will result in unnecessary computational expense. Algorithms that effectively adjust number and spacing of nodes will maximize accuracy and speed. Snakes have the advantage of being conveniently autonomous in their search for a minimal energy state. Also, since the integral operator is an inherent noise filter, they are insensitive to noise and other image ambiguities. These ambiguities include spatial aliasing and sampling artifacts that can cause boundaries to be indistinct and disconnected (McInernery & Terzopoulos, 1996). In contrast, snakes often overlook minute features in the process of minimizing the energy over the entire path of their contours. Initiation of the snake is critical to the success of the algorithm and accordingly there are a few points to consider. First, the snake must be initialized close to the expected boundary for good performance since they can get stuck in lower minima states. This is a concern since contours are generally difficult to initialize around the region of interest. For best results, the snake should be initiated depending on the uniformity of the intensity distribution, either inside or outside the boundary. Caution must be observed if the curve is initialized on the inside of the expected boundary in a place where the image information has little influence as the curvature penalty will cause the spline to shrink to a point. The addition of an outward pointing force addresses the shrinking problem

PAGE 48

37 Figure 2-8. Singularity in contour evolution. A self-intersecting curve creates a singularity which can be described as in the process of either splitting or merging. Grass burning is one method used to solve this ambiguity (right). It determines the state by marking the path (texture) traveled by the curve (blue). In this case, the contour is in the process of splitting. inherent in the internal energy term of the snake. This edge-based active contour is called a balloon (Cohen, 1991). The balloon function includes stability aids such as the normalizing of image force to restrict boundary movement and interpolation to avoid discretization errors. A particular concern with the evolution of contours occurs when image topography causes the contour to self-intersect. The intersection point can be in one of two states, splitting or merging. The nodes can be reparameterized to adjust, as in establishing two separate contours in the case of splitting (Figure 2-8). However, the state must be established. Several solutions exist to tackle this problem. One is to use a grass-burning assumption (Sethian, 1996a). This creates an entropy condition where the evolving front leaves a burnt or irreversibly marked path as it travels. This will determine what direction the contour was moving. Another solution is to use T-snakes, which use a triangulation of the embedded space to determine new node points (McInerney & Terzopoulos, 1997). The Level Set Method is yet another method for tracking the evolution of interfaces (Osher & Sethian, 1998). Level set methods offer highly robust and accurate methods for tracking interfaces moving under complex motions (Osher & Sethian, 1988). They handle topologically breaking and merging naturally as in the creation of channels in a surface contour.

PAGE 49

38 Figure 2-9. Level set application for contour evolution. The contour (red) is embedded as a zero level set of a higher dimensional surface (blue). This allows the smooth adaptation of the contour to topographical changes. Details about theory, implementation, and application of level set methods are in Sethians Level Set Methods (1996a). It relies on two central embeddings. One, the contour is embedded as the zero level set of a higher dimensional function (Figure 2-9). Two, the velocity of the contour is integrated into the higher level function. This is called the extension velocity, and is curvature dependent. A companion technique to Level Set Methods is Fast Marching Methods (Sethian, 1999, 1996b). They are numerically efficient methods for evolving a front traveling in one direction. They can be used to construct a distance map and provide appropriate extension velocities for level sets (Adalsteinsson & Sethian, 1999). There are several associated active contour methods of note that are used for image segmentation. The first uses a geodesic or minimal path formulation for active contours in the level set approach and applies it to segmentation of medical imagery data (Caselles et al., 1997). The second uses gradient flows to direct the curve evolution (Kichenassamy et al., 1995). The last integrates other geometric techniques for image segmentation (Malladi et al., 1995).

PAGE 50

39 Much of the current research in active contour models deals with generalizing the form of the contours and overcoming the convergence and stability problems encountered during the energy minimization process (Davatzikos & Prince, 1992). Region-based Segmentation In applications such as medical imaging, adequate edge information such as strong gradients may not always be present for an edge-based segmentation approach to be consistently accurate. In these cases region-based models may be employed. In the region-based segmentation approach, the boundaries of the deformable model are determined by statistics inside and outside a region or cluster of voxels. These statistics measure properties such as unique texture patterns, homogeneity of intensity, or some other pixel-based statistic. The goal in evolving the region is that variation of these properties is less inside a region then between regions. Similar to edge-based methods, region based methods are evolved through the minimization of an energy term. This global energy term, however, is defined for the entire area of the region rather than only the boundary. There are a variety of region-based segmentation methods that include Bayesian segmentation (Geman & Geman, 1984), piecewise constant energy (Mumford & Shah, 1989), region competition with balloons (Zhu & Yuille, 1996), and an energy-based watershed (Bleau & Leon, 2000; Nguyen et al., 2003). One technique uses the level set formulation to evolve the boundary of a region based on texture statistics (Yezzi et al., 1999). Region-based segmentation algorithms have several advantages. First, they have a greater capture range and are not as dependent on initialization as edge-based methods. Also, they are not reliant on high frequency information and are not as susceptible to

PAGE 51

40 noise (Chan & Vese, 1999). Care must be taken in application, though, as the determination of regional parameters can be computationally expensive. As a final note, efforts have been made to integrate region-based and edge-based modeling techniques (Chakraborty et al., 1996; Paragios & Deriche, 1999). The addition of a boundary regulation term is used to control smoothness in a region-based segmentation model. Also, geometric information, such as a shape prior, is much more easily incorporated into a boundary formulation. However, the tradeoff of boundary regularization comes at a cost of strength of region description. Shape-based Segmentation The final approach to medical image segmentation is to incorporate shape information in a deformable modeling algorithm. This is accomplished through the construction of an atlas. An atlas is a collection of prior information used to direct the deformation of the algorithm. It can be a set of points, finite elements, flow fields, or other similar parameter, but is most commonly a shape descriptor. Once defined, the atlas or standard template molds itself to the target image, imprinting an inherent label map. The data associated with the atlas can also be used to direct the mentioned edge or region-based segmentation strategies. Atlas Warping The classic warp consists of a representative image scan that has an assigned label map L*(x). Given a new image, some transformation T must be computed that deforms the representative scan I*(T(x)) to correspond with I(x). The same transformation applied to the label map L*(T(x)) will then describe the label map for the new image L(x). There are several examples of this template-driven segmentation (Pentland & Sclaroff, 1991; Sclaroff & Pentland, 1995; Wang & Staib, 1998b; Warfield et al., 1998).

PAGE 52

41 One specific method computes the deformation field in a way that allows a course to fine solution (Christensen et al., 1996; Christensen, 1999). Another method begins with a rigid transformation and iteratively progress in plasticity to find a solution (Miller et al., 1997). A technique especially useful in the atlas-based segmentation of composite boney structures incorporates the rigidity of the tissue class in the deformation model (Little et al., 1997). More examples of atlas warping include works by Cootes and Taylor (1992a), Jones and Poggio (1998), and Pichumani (1997). Modeling There are several factors to consider in the development of a shape-based model. First, for automatic interpretation, it is essential to have a model that not only describes the size, shape, location and orientation of the target object but that also permits expected variations in these characteristics. In order to properly account for this object variability seen in application, a statistical analysis must be performed (Dryden & Mardia, 1998; Neumann & Lorenz, 1998). Casting the fitting process of deformable modeling into a probabilistic framework allows incorporation of prior statistics as well as an inherent measure of uncertainty (McInerney & Terzopoulos, 1996). These statistical models have the advantage of being flexible to ambiguity and noise. Since the development of a model and its associated training data set can be data intensive and computationally laborious, techniques are used to minimize the number of statistically significant parameters. One technique in particular is Principle Component Analysis (PCA). This process identifies the primary modes of variation and reduces the dimensionality of the data set, optimizing the framework (Golland et al., 2000; Lorenz & Krahnstover, 1999). Multiple strategies for shape description have been developed. Feature detection is a common approach. A feature is a geometric property of the object to be segmented that

PAGE 53

42 can be easily identified by an algorithm. The algorithm performs a segmentation by matching the set of features that is extracted from the unlabeled image to the training set. Depending on the application, the association of these features should be defined with certain invariances such as translation, rotation, or homogeneity. One method of feature extraction uses medial axis or skeletonization. The geometric features produces are scale-invariant, which allow detection at different resolutions (Dougherty, 1993). A drawback to the use of skeletons is that original size and edge information are lost. An improvement on skeletonization uses fixed topology skeletons. They have the advantage of being robust to the noise and quantization errors that traditional skeletons are susceptible to (Golland et al., 1999). Mathematical morphology can be used to build a shape classification strategy as well (Dougherty, 1993). The binary object of interest is probed with an array of simple shape primitives, from which a statistically appropriate feature set can be collected. In another method, distance maps are used as a shape descriptor (Golland et al., 2000). Distance maps provide smooth and wide minima for a matching algorithm. Methods There are a variety of shape-based segmentation algorithms. One particular group uses Active Shape Models (ASM) (Cootes & Taylor, 1992a; Cootes et al., 1992b; Cootes et al., 1999b). The ASM is a parametric deformable model that is represented in the image as an n-point polygon. The algorithm then deforms the points of the polygon to match the landmark points in the target image object. The points are evolved according to a point density model (PDM), which is constructed using PCA. The PDM provides a statistical distribution of global shape variation. Once the model is deformed, an accompanying label atlas identifies the segmented object. The ASM is significant

PAGE 54

43 because it is fast and accurate and can describe complex shapes with minimal parameters. In application, models have been developed to account for both Gaussian and non-Gaussian data (Cootes et al., 1995; Cootes & Taylor, 1999a). In one instance, ASMs are incorporated into a Bayesian formulation for boundary location (Wang & Staib, 1999a). ASMs have also been extended to accommodate grayscale information as Active Appearance Models (Cootes & Taylor, 1998). Shape-based segmentation approaches have been integrated into edge and region-based algorithms to retain the benefits of those approaches. For example, contours are well suited to handle shape information, and such information can easily be incorporated as priors. One method uses statistical priors on the Fourier coefficients of the contour to represent shape (Staib & Duncan, 1992). A related method combines snakes with a shape-based Fourier parameterization method (Szekely et al., 1996). Chen et al. (2001), uses shape priors with a level-set evolved geometric active contour model. Leventon (2000) tests a method that incorporates distance-intensity profiles and shape contour with a probability-based segmentation approach. It uses a level set implementation of geodesic active contours for shape evolution. Also, PCA is implemented to reduce the dimensionality of the training set. A couple of region-based shape methods are also of note. Cremers et al. (2002) incorporate statistical shape knowledge into the Mumford-Shah functional to perform segmentation. In application of cervical spine segmentation, Pichumani (1997) uses a finite-element shape model integrated into a region-based approach.

PAGE 55

CHAPTER 3 SEGMENTATION Currently, to create a segmented model appropriate for guided surgery, the clinician follows a few basic steps. First, he reviews the volumetric CT scan and selects an intensity value that defines the edge of the bone. Isocontours are displayed on the terminal to assist in this task. Typically, voxels of bone have much higher intensity values than adjacent soft tissues, and the boundary value can be easily identified. The selected intensity is used to define a threshold mask. The mask is a binary overlay which identifies all voxels that are associated with bone (Figure 3-1). The highlighted voxels describe an object which can usually be seen with accompanied 3D visualization software. There are a few considerations in the creation of a threshold model. First, it is crucial that the edge of the bone is accurately defined, as many of the registration techniques use surface points to do the matching. A model created with a threshold value that is low will include non-bony soft tissue such as cartilage and ligaments. A model created with a threshold value that is high will diminish the volume of the bone, shrinking the edge and withering the interior. The process of using a threshold to describe an area of the image can also be utilized to isolate an object from high intensity artifacts or noise. Instead of a just a single lower intensity threshold, an upper intensity bound is used as well. The model created by thresholding rarely results in a structure that is segmented or appropriately sectioned to match the underlying anatomy. Therefore, the model must be 44

PAGE 56

45 Figure 3-1. Building a threshold model. On the left is a sagittal slice of a cervical CT image. On the right is an isocontour (red) scribed at the boundary of the bone, with its implied threshold mask area highlighted in blue. It can be seen that several of the bones are inappropriately grouped together. manually segmented. This is a tedious and time-consuming process whereby the clinician flips through each of the series of CT image slices and uses simple pixel highlighting tools to identify each bone. The result is a segmented model that can be used for guided surgery (Figure 3-2). However, even though this method produces an acceptable result, the time commitment involved in this manual process restricts the routine use of guided-surgery for multi-level applications. Therefore, the focus of this research is to develop methods to automate the segmenting process and facilitate these procedures. Problem Evaluation Several guidelines were established to direct this segmentation research. First, the time spent by the clinician to prepare a segmented model must be minimized. This includes increasing simplicity and minimizing user input. In contrast, however, there must be enough prior knowledge provided by the user to direct the algorithm to a correct solution. Another contributor to the involved time of the clinician is the speed or computational efficiency of the algorithm, but this optimization is secondary at this stage of the research. Second, there must be an accurate approximation of bone edge for the

PAGE 57

46 Figure 3-2. Manual segmentation and display. The image on the left is a 3-D view of a contiguous model formed using the thresholding method. The image on the right is the same model that has been manually segmented. The goal of this research is to reduce the involved time to produce segmentation. registration methods to be successful. Third, the segmentation process must be robust to the assorted image variations. These include pathologies and artifacts, such as those induced with instrumentation. The initial step in development of the segmentation algorithm is to assess a random sample of images used for image guided surgery. Currently, most of the image-guided surgical procedures are implemented on the cervical spine. This is because there is a very small tolerance available for tool placement to avoid critical vascular and nerve structures. A series of the high-resolution cervical CT scans used for these surgeries were evaluated. Upon analysis, it was apparent that the general shape and intensity distribution of the vertebrae were similar. Some patients had an overall higher bone density, with clear distinction of boundaries. Others had bone density that was close to that of soft tissue. In some of the most severe cases, the individual bones were difficult to recognize. Some of the pathologies observed include fractures and fusion, each which showed unique discontenuities and intensity patterns. Instrumentation, such as rod and screw assemblies or wires used to stabilize bones, produce unique artifact patterns in the

PAGE 58

47 Figure 3-3. Surface characteristics of vertebrae. The typical cervical vertebra has very complex topology, with areas of high curvature and channels. These properties make it a troublesome candidate for edge-based or shape-based algorithms. (Left) Superior view. (Right) Lateral view. (Grays Anatomy). CT scan and disrupt the normal intensity patterns. A careful review of image variations was helpful in selecting the approach to algorithm design for this research. Method Selection Several avenues exist for the development of a segmentation algorithm. One common approach is to use boundary localization techniques. These involve the creation of a boundary concept that will deform until it reaches a minimum energy state at an appropriate edge. They have the capacity to handle small gaps in information, but also have the potential to arrive at false minima. These are erroneous minimal energy states that locate an incorrect boundary. A primary difficulty in the use of boundary localization techniques is that the typical cervical vertebra has very complex topological characteristics including areas of high curvature and channels. This makes it a difficult surface to evolve (Figure 3-3). In addition, the peculiar shape would exacerbate the initialization problems that have the potential to plague these methods. Another avenue for image segmentation is to use region-based methods. These methods have the advantage of not being as dependent on initialization as edge-based methods and are less affected by the gaps in information or high frequency errors that may cause boundary methods to reach a false minima. A drawback, however, is that

PAGE 59

48 region-based parameters are often computationally expensive. In addition, region-based methods may not be best suited to handle the complex boundaries and topology presented with a typical vertebra. One advantage to a region-based approach, however, is that shape information can be more easily incorporated. An atlas, if properly developed, could capture variance across a population and be used not only to direct segmentation, but also to identify abnormality or pathology. The incorporation of shape information, however, can be costly. First, reliable features must be identified. And then, a comprehensive database must be put together and maintained. Given the number of pathologic conditions to account for, this could be a very time consuming process. The third avenue to develop a segmentation algorithm is to use voxel-based techniques, which are independent of any boundary or region components. This significantly reduces complexity, yet has the drawback of not being able to easily account for geometric information. Thresholding is a proven, if inefficient, voxel-based modeling technique, and makes a good approximation of the bone surface. Therefore, it provides a convenient platform from which to launch other voxel-based modeling techniques. Given that this is the simplest approach, it shall be investigated first. To facilitate this work, morphological operators will be used, as they are computationally efficient, work in 2D and 3D and preserve edge information. Research Tools A programming language and accompanying toolset were selected to facilitate the segmentation research. IDL (Research Systems Inc., Boulder, CO) was chosen because of its strong image processing toolset, assortment of visualization methods, a GUI interface, and ability to process large arrays. First, a GUI was created to visualize volumetric CT images. Through it, any orthogonal image slice could be viewed and

PAGE 60

49 Figure 3-4. Toolset GUI. Multiple functions are easily arranged to facilitate segmentation. Shown is an axial CT slice with an overlaid indexed array (yellow highlight), indicating the segmentation boundary and corresponding anatomical association. manipulated. Plus, a function for overlaying intensity contours to aid in the selection of threshold values was included. Since a 2D view does not give an accurate account of 3D connectivity, several 3D visualization schemes were also setup. Once a viewing system was constructed, multiple groups of 2D and 3D image processing functions were added to the menu structure, such as the morphological functions open, close, and gradient. The next step in toolset development was the addition of functions to the program environment that mimicked the manual segmentation process. Before this could be implemented, however, a data representation of the segmentation model had to be

PAGE 61

50 chosen. We decided on an indexed overlay. This is an integer data array equal in size to the image array with an index number assigned to each voxel. This index identifies that voxel as being associated to a specific bone. Once created, this array could be blended with the original picture for visualization (Figure 3-4). This data representation for segmentation is certainly inefficient, yet is easy to associate with any voxel-based manipulation. Next, several options were added to the GUI to allow the creation of the index array based on threshold techniques. Several tools were created to assign or change index assignments. They include 2D single pixel and polygon and 3D sphere and rectangular prism assignment options. In addition, continuity functions in 2D and 3D were also added. They can be used to paint interior spaces or to fill holes. Throughout the research, functionality of the interface was expanded accordingly. Cursory Examination In order to test the toolset interface, a few simple, preliminary examinations were made. The primary failure of segmentation by threshol ding is that the pro cess does not properly isolate each of the bones, leaving one c ontiguous clump of bone. These initial experiments, then, attempt to use a few simple techniques to accomplish improved separation. There are two approaches to test. One is to cut the threads that leave the bones connected. The other is to identify the boundary and subtract it from the contiguous model, leaving separated regions. The morphological open operator is a simple function that can be used to effect a break of thin connections on a binary object. Hypothetically, this could be used to separate bones on a threshold model. Several tests were performed using the open function on sample images. Symmetric structuring elements were applied in an order of increasing width to see what size would effectively break the connections. The result is

PAGE 62

51 Figure 3-5. 3D intensity gradient. The morphological gradient is constructed from a CT volume. On the left is an axial slice of a cervical vertebra. On the right is an equivalent slice from the gradient map. Notice the interior areas of strong edge-information. that the size of structure required was thicker than some of the natural boney structures, and the application of which significantly altered the surface of the vertebra. The intensity gradient is a very commonly used indicator of edge strength. Several 3D gradient intensity maps were calculated from sample CT scans and analyzed. A majority of the boundaries were highlighted in these trials; however, there were multiple gaps or areas of low gradient. Therefore, without further modification, the gradient map could not be directly used to make a boundary subtraction and foster segmentation. However, an interesting property of the vertebrae is that there is a strong gradient on the interior of the bone (Figure 3-5). One observation is that, given these circumstances, an edge-based algorithm solely dependent on gradient information would have difficulty localizing the correct boundary. The interior areas of strong gradient would tempt the curve during evolution. As a result of these preliminary experiments, a rectangular region of interest (ROI) tool was setup to allow the application of image processing algorithms on a defined area. It was theorized that this would allow unique image faults to be addressed without

PAGE 63

52 affecting the whole image. For example, it could be used to address specific gaps in the gradient map. After a few trials, it was apparent that ROI definition in accordance with a specific region in 3D was a time-consuming process subject to burdensome iterations and is something that should be avoided.

PAGE 64

CHAPTER 4 REGION GROWING Kernel Definition The fundamental problem in using a threshold model created at the edge defining value for bone is that it often connects adjacent anatomy as one contiguous model. However, it has been observed that a regression of the isocontour to a higher threshold value results in a disjoint model that can be uniquely associated with the underlying anatomy. It is hypothesized that these regions can be used to create a successful segmentation. A method has been developed to create and label this kernel-based threshold model using IDL. First, an isocontour is selected that allows the creation of disjoint regions. This is an iterative process that involves using the visualization software to check for contiguous regions. Once an intensity level is selected, a binary threshold is created in the label array. The proper IDs are assigned through point selection and continuity labeling tools (Figure 4-1). Since the goal is to minimize the intervention of the clinician, not every unassigned region is labeled. It is assumed that the user will use knowledge of the segmentation method to label the regions that will have the most profound effect on the results. Shape Correlation It is observed in the creation of the disjoint threshold model that the shapes of the kernel regions bear a resemblance to their associated boundaries. It is reasoned that if 53

PAGE 65

54 Figure 4-1. Manual kernel selection. A threshold model created at the bone-defining intensity value of a sagittal slice of a cervical spine CT image (left). The threshold value is raised, resulting in a model with multiple discontinuous bony elements (middle). Each area is labeled to correspond to the underlying anatomy (right). enough boundary information is preserved in the shape of the kernels, then they can be used for segmentation. One simple voxel-based method to test this theory uses both the labeled region kernel model and the contiguous boundary threshold model. They are compared, and each of the highlighted voxels in the boundary model are assigned according to the closest labeled kernel region. The assumption is that there is a high likelihood that an unassigned voxel has the same label as the closest pre-labeled region. This can be accomplished through a system of constrained morphological dilations. Specifically, the kernels of adjacent anatomies are dilated equally until they touch. The voxel IDs assigned in this process cannot be overwritten, so the boundary created is fixed. The process of repeated dilation continues until all of the voxels in the boundary threshold model are assigned. Expanding on this idea, distance maps may be used to approximate this process. A distance map is an image whereby each of the voxels is assigned an intensity value corresponding to the Euclidean distance to an object. This value, as it turns out, is approximately equal to the number of unit dilations necessary to reach that voxel from the object. Further, if a distance map is created for two sets of objects, the

PAGE 66

55 Figure 4-2. Segmentation using distance maps. Unique, discontinuous kernel regions are created from a cervical CT scan (upper left). Then, two distance maps are created. One represents the kernels associated with C2 (yellow) and the other represents the kernels associated with C3 (Cyan). The difference of the resulting maps (upper right) approximates a boundary between C2 and C3 (lower left). A segmentation results when a boundary threshold mask is applied (lower right). difference of the two maps scribes a zero line exactly equidistant between the object sets. Figure 4-2 illustrates this process enacted upon two sets of adjacent kernels. This segmentation method is a good way to quickly approximate a boundary. However, there is one failing that made further pursuit inadvisable. The amount of shape information contained in the labeled kernels rapidly deteriorates as the threshold level is

PAGE 67

56 Figure 4-3. High threshold values result in diminished kernels. The threshold contour (red) is increased from the boundary defining value (left) to one that allows proper label associations. The resulting high threshold and label (right) shows kernels that retain very little boundary shape information. Object 4-1High threshold movie (object1-highthreshold.mpg, 27.6 MB) increased (Figure 4-3). Therefore, circumstances that warrant a high threshold to achieve isolated kernel regions produced the least accurate boundaries. Hysteresis The previous study showed that the shapes of the segmentation kernels are not infallibly correlated to boundaries. More information is needed to direct the growth of these kernels to arrive at the appropriate segmented boundary. In the kernel labeling process, an intensity threshold is used to withdraw the isocontour to the interior high intensity areas of the bone. This threshold information, if appropriately utilized, can assist the kernels to revert back to the bone edge. An algorithm was developed using a combination of morphologic dilation and thresholding functions to progress labeled kernel regions to a fully segmented model. The inputs to this algorithm are the CT image, the user-determined bone boundary threshold intensity value, and the kernel region dataset. Each iteration of the region-growing algorithm involves three basic actions (Figure 4-4): 1. The step to a lower threshold and resulting isocontour. 2. The constrained orthogonal dilation of the seed regions. 3. The restriction of the dilated voxels to the area outlined by the isocontour.

PAGE 68

57 Figure 4-4. Region growing algorithm iteration. Shown in the upper left are two adjacent bone segments with corresponding seed regions. The first step progresses the threshold contour to a lower value (upper right). Then the seed regions are grown through morphological dilation (lower right). Finally, any highlighted voxels outside the threshold contour are removed (lower left). This basic cycle is initiated at the kernel-defining threshold value and repeated until all of the volume outlined by the lower threshold mark is filled. A couple parameters can be adjusted to discover the most effective sequence for region growing. The first is the number of threshold divisions, or the number of times that the isocontour value will step in order to reach the edge defining value. The height of the step will depend on the number of threshold divisions and the difference between the two boundary threshold values. The second definable parameter is the number of unit

PAGE 69

58 Figure 4-5. Algorithm input processing sequence. The steps requiring user input are outlined in green. The input image is histogram equalized over the boundary threshold ranges to regulate growth while preserving the geometric boundary conditions. dilations per threshold step. This determines the rate at which the available threshold-bound area will fill. Through repeated trials, a couple of basic trends were noted. Generally, a higher number of threshold divisions produced a more accurate segmentation. Secondly, the number of dilations required to properly fill a threshold step varied widely. In order to regulate this growth, an analysis was performed. It was determined that an intensity histogram could be used to determine the number of voxels

PAGE 70

59 Figure 4-6. Region growing. Once the input image is prepared, the algorithm instructs the labeled kernel regions to grow incrementally until the bone-defining threshold is filled. Object 4-2Region growing movie (object2-success.mpg, 25.2 MB) available to be filled per threshold step. Accordingly, the histogram can be equalized and applied to the input image to even the growth. This step has been added to the algorithm input processing sequence as shown in Figure 4-5. Once the preparation steps have been accomplished, the iterative region-growing algorithm is activated and continues until the total bone volume is assigned a label (Figure 4-6). The completed label set then defines the segmentation for the original image. Before the segmented dataset is used for a surgical application, it must be verified. This involves overlaying the segmented dataset on the original CT image for visual comparison. Manual post-processing steps may be necessary if the segmentation does not succeed in the proper isolation of adjacent vertebral bodies. The same tools mentioned for manual segmentation are used for this post-processing step. Algorithm Enhancement A few avenues were investigated to improve this region-growing method of image segmentation. The first was an examination of growth characteristics, and how they could be better controlled to suit the application. The second involves the modification of image information to produce more auspicious growth patterns. The third tests the

PAGE 71

60 Figure 4-7. Open and close operators are used to alter the evolution of the growth algorithm. On the left is the original image. In the center is an image with the open operator applied. The white regions highlight the areas removed by the operation. On the right is the image with a close operator applied. The white regions are the areas filled by the operation. inclusion of user-defined input information in addition to the initial kernel-laden segmentation dataset. During the region growing operations, the kernels tend to expand in finger-like fashion until they merge with other anatomically self-similar sections. The problem arises when one of these extends to a piece of unrelated anatomy, essentially allowing it to falsely populate. Two shape filters were inserted into the algorithm to modify growth behavior. One, the morphological open operator, was used to curb the extension of thin tendrils to false anatomy (Figure 4-7). The other, the morphological close operator, was used to facilitate growth into self-same anatomy. Orthogonal (6-connected) and diagonal (26-connected) structures were both tested. The trials with an orthogonal open operator applied tended toward the more favorable result by restricting finger growth, even though it also inhibited movement between self-same regions. The close operator, however, created too many false bridges to be effective. Image Variations Several variations of the input image were presented to the region-growing algorithm to determine if they could be used to improve the quality of the segmentation.

PAGE 72

61 In each of these trials, the kernels were defined based on the altered image and allowed to progress using the iterative scheme. However, the bone-defining threshold value selected using the original image was retained and acts as a barrier for growth. This was done because the intensity threshold is a good approximation of bone edge and these image variations would undoubtedly result in less appealing threshold-based boundary approximations. First, it was thought that a smoothing filter applied to the original CT image would weaken the intensity of anatomical cross-links and promote a more uniform, centralized geometry from which to evolve. The resulting anatomical edges were very clean and smooth, yet inappropriately labeled. This was attributed to smeared boundaries which lessened the inhibition of the kernels to leak into false anatomy. Next, the region-growing algorithm was tested on the gradient of the input image. The gradient is a common image filter that is used to highlight boundaries (Figure 4-8). A first note is that gradient kernels are much more disconnect at higher thresholds which makes the index assignment phase more laborious. The resulting segmentations showed similar success to those of the unaltered image, however, the borders were choppy and the areas of bleed-through were relocated. One noticeable property of the gradient image is that there are high intensity rings around the periphery of the vertebrae. This leaves the interior relatively uniform at a low intensity value. Therefore, an inverted gradient image would leave a strong centralized bulk kernel to feed into the iterative algorithm (Figure 4-8). The kernel assignment phase for this image variation, as predicted, was simple. The large internal areas were easy to label. The resulting segmentation, however, left

PAGE 73

62 Figure 4-8. Unique properties of the input CT image may be isolated or combined to produce a more favorable segmentation. On the left is the original CT scan. In the center, the gradient of the image highlights boundaries. The inverse of the gradient (right), in contrast, highlights the interior of the image. unpredictable, jagged boundaries. Example segmentations of the mentioned image variations are shown in Figure 4-9. The previous image variation tests, individually, did not provide wholly beneficial results. However, they each revealed advantageous properties which could be utilized to improve segmentations. Since this region-growing scheme is based on global intensity patterns, the various images were blended or combined to form a composite image with more suitable qualities. The most prosperous combination was a blend of original CT and inverse gradient images. The inverse gradient image served to add a more centralized interior which would quickly unify distant pieces of the vertebra. The original CT image contribution served to temper the erratic boundary formation that the inverse gradient image produced. This method had the potential to improve the quality of the segmentation, however, the optimal blending ratios were scan specific and tedious to determine. Also, boundary defects still existed which would foil the segmentation. Prelabeling The kernel assignment phase of the region growing process has a strong influence on the success of the segmentation. When circumstances require a high threshold level to

PAGE 74

63 Figure 4-9. Segmentation results using variations on the input CT image. The top row indicates the input kernels and the bottom row displays the result of the algorithm. Examples of smoothed image (left column), image gradient (center column), and inverse gradient (right column) are shown. create unique kernel elements, the algorithm may be left starved. Specifically, only a limited percentage of voxels in any one piece of anatomy may be available to seed growth. The consequence is a much smaller chance of the correct label reaching all other areas of the anatomy before borders are assigned. The third avenue pursued to improve the region-growing segmentation method is the inclusion of additional user-defined labels. These labels, set during the kernel assignment phase, force the assignment of unassigned voxels that are exposed during region growth. It is expected that these pre-labeled voxels will appropriately spread to adjacent areas and result in a better segmentation. The updated iterative action consists of these steps:

PAGE 75

64 1. The step to a lower threshold and resulting isocontour. 2. The pre-label of unassigned voxels. 3. The constrained orthogonal dilation of the labeled regions. 4. The restriction of the labeled voxels to the area outlined by the isocontour. The goal is to find a specific labeling technique that makes a strong contribution to segmentation while requiring a minimal amount of user input. The two methods tested were 2D image slice and sphere labeling. The CT image slice was decided as a good candidate for pre-labeling because it is a straightforward manual segmentation process. Sagittal image slices were chosen for the pre-labeling study because the vertebrae are fairly easy to visualize and a medial slice is likely to contain all of the boney sections. Between one and five slices were tested. Generally, a sparse kernel model or disruptive pathology forces the user to seek additional slices to label. The first problem noted is the pre-labeled slices tend to have little influence on adjacent axial growth patterns unless they happen to cross the labeled sagittal plane. The response was to attempt to position the pre-labeled slice over critical contact areas, such as the transarticular facets. This moderately improved the output; however, the determination of the most effective slice is a tedious process that often requires frequent references to adjacent sagittal image slices. In response to the inability of the 2D slice pre-labeling to effect growth in the axial direction, sphere labeling was developed. This is simply the creation of spherical regions given a user-defined point, radius, and label. Remarkably, this method showed little improvement in segmentation quality. The reason is that kernels tended to develop in a ring fashion a bit inside the edge of the vertebra before progressing to the border. Therefore, in order to effect kernel growth, the sphere must extend very close to the vertebral boundary. The definition of a sphere in this manner, however, requires strict

PAGE 76

65 attention to adjacent slices to assure that there is no violation into adjacent anatomy by the sphere. Any such occurrence would be very defeating to the overall result. The time-consuming verification process for sphere placement makes this avenue discouraging and impractical. Instrumentation The segmentation methods discussed rely on intensity profiles of the bones. When surgical instrumentation such as rods, screws, and wires are implanted into the patient, though, this profile is severely disrupted. Metallic implants cause auras of high intensity where they are positioned and serve as bridges incorrectly linking anatomy. Fortunately, the intensities produced are much higher than any occurring in the natural bone profile. This allows an intensity threshold mask to subtract the instrumentation out of the region growing algorithm. The iterative sequence is similar to pre-labeling; void regions are created in which the dilating kernels cannot travel. After a couple of tests, this method of instrumentation treatment proved very successful. The algorithm was able to navigate around the instrumentation and produce a segmentation (Figure 4-10). The only modification that was appended was an additional dilation or two of the instrumentation region. The first reason is that partial volume effect alters bone profile slightly outside the actual geometry of the instrumentation. The second reason is that implants that have resided in the body for a lengthy period of time will likely have bone growth that alters the natural bone profile. Results The process of segmentation with the region-growing algorithm was tested on a total of 31 clinical, high-resolution cervical CT scans intended for spinal surgery. Success was gauged by the amount of effort necessary to achieve an accurate

PAGE 77

66 Figure 4-10. Instrumentation subtraction. The instrumented area (purple) was successfully removed from the region growing algorithm, severing the false bridge and allowing the region-growing algorithm to naturally progress. visualization of the segmented vertebrae. Note that this may be more stringent than what is required for any specific registration method. Primarily, success was measured by the amount of image slices needing corrective manual post-processing alterations, with an awareness of input effort and algorithm runtime. There were 19 (61%) of the scans that required no or minimal post-processing. There were 7 (23%) that required manual processing, yet still remained a significant improvement over manual methods. The segmentations that showed no significant improvement over manual segmentation amounted to 5 (16%). The algorithm run time was roughly 3-5 minutes (Pentium 4 at 1.4 MHz) for 3-5 vertebral levels. The array size was a very significant factor in runtime since the routines operate on full rectangular matrices. Optimization schemes would significantly reduce processing time. There are a couple of key factors that affected the outcome of this segmentation method. First, the quality of the kernel regions had a significant contribution. If the seed

PAGE 78

67 Figure 4-11. Segmentation failure: boundary leakage. Seed regions defined distant to an anatomical boundary must carry the accurate label to the boundary before a competing label invades. The center image depicts the boundary leakage with the final segmentation on the right. Object 4-3Segmentation failure: boundary leakage movie (object3-failureleakage.mpg, 25.8 MB). regions were small and disconnected, then considerable weight was placed on the ability to branch and interconnect properly. The failure of those actions leads to boundary leakage (Figure 4-11). Threshold models that are created at a very high intensity produce these low quality kernels. Areas of high intensity contact will force the seed region selection to a high threshold. Severe posturing can change the proximity of the bones and bring high intensity areas into contact, usually at the transarticular surfaces. The second primary component to success of the segmentation process was image resolution. The cervical boundary profile requires a high level of detail. A low-resolution image will increase partial volume effects yielding incomplete intensity patterns and increase opportunity for failure. A similar effect is seen when posturing compresses the space between vertebrae. The last main ingredient to success of the segmentation was the condition of the bone intensity profile. Some patients had comparatively lower bone contrast with soft tissue, which created threshold models with significant gaps. These gaps effected the ability of the anatomic labels to spread. The intensity profile was also

PAGE 79

68 disrupted with pathology such as fusion. The most severe result observed was kernels that dilated with little distinction between correct anatomical locations (Figure 4-12). Figure 4-12. Segmentation failure: weak bone intensity profile. A CT scan that has low bone to soft tissue contrast or pathology that disrupts the clarity of the bone profile does not allow the dilating kernel regions to properly meet at the vertebral boundaries. Object 4-4Segmentation failure: weak bone intensity profile movie (object4-failureweak.mpg, 28.6 MB).

PAGE 80

CHAPTER 5 CONCLUSIONS AND FUTURE WORK A semi-automatic method has been developed to expedite the production of segmented CT images beyond that of manual methods. It promotes accurate visualization of bony components and encourages the use of image guidance with surgeries involving multiple levels. This region-growing hysteresis method has a couple of notable benefits. First, it is a relatively simple procedure to implement. It has no pre-defined datasets to build or maintain. It works well with fracture pathology. Plus, instrumentation can be easily subtracted without significantly affecting the performance of the algorithm. In contrast, this segmentation process has a few limitations. One is the potential for lengthy user intervention, such as during kernel assignment or post-processing, if required. Second, the threshold model does not completely define what is considered the interior of the anatomy. Some of the interior regions have the same intensity as the surrounding soft tissue and so are not included in the model. Region filling methods may be used to fill the interior; however, this is not a thorough solution. Note that image guidance registration may not require a segmentation with complete internal labeling to be successful. Third, the algorithm is susceptible to small intensity defects, especially those that bridge adjacent bones. A couple of directions exist which have the potential to improve upon the developed segmentation method, remedying some of its shortcomings. The region-growing hysteresis method relies strictly on global voxel-based operations. Additional 69

PAGE 81

70 elements, such as boundary or region-based components, if integrated, may provide the additional information necessary to improve the segmentation. For example, an elastic boundary concept and internal region statistics could be used to make the method more robust to noise and point defects. Also, any statistical geometric profile could be used to guide the evolving boundary to its correct location. One cost to these options, however, is the maintenance of a potentially cumbersome dataset. For the addition of these parameters to a method to be successful, it must be considered how well-suited the algorithm is to incorporate the information. For example, in the current algorithm, choosing a penalty term to curvature for a boundary-based representation may be difficult with regard to the chaotic nature of how the regions intersect. After review, we decided to pursue a method that was perhaps more suited to the integration of the mentioned voxel, boundary, and region-based components. Leventon (2000) presents a noteworthy approach to segmentation by calling upon curvature models, prior shape models, and statistical image-surface relationships to perform segmentations. One distinct advantage to this method is it blends these techniques into a level set framework, allowing the boundary to evolve over complex topography. This particular algorithm uses a signed distance map as the level set surface representation. It has several advantages. First, boundary and region-based representations are implicitly contained in the signed distance function. Second, the notion of holes in the boundary is seamlessly contained in the representation; a small hole in the object does not create an entire new set of features in the object. Third, it is robust to slight misalignment of features. One interesting aspect of Leventons method is the use of intensity/distance-to-boundary distributions. These profiles are incorporated as

PAGE 82

71 Figure 5-1. Intensity/distance-to-boundary profiles. The vertical axis is grayscale value, with the top having the highest intensity. The horizontal axis is distance from object boundary (vertical black stripe). Right indicates interior distance from the boundary and left indicates the exterior distance from the boundary. Color indicates frequency of pairs. The first three images on the left show the profile of vertebra C2 from three random individuals. The three images on the right show three different levels, C1, C2, and C3 (center to right) from one individual. training sets into a model that makes boundary estimations. Specifically, with this particular adaptation of the level set framework, distance maps are used to describe the higher-dimensional surface. The distribution of surface-intensity pairs, then, is a voxel match of the morphological distance map of the segmented object, and its associated CT image value. A cursory examination of intensity/distance-to-boundary distributions of spinal images was performed to determine if these profiles would be useful for segmentation. Six cervical spine CT images were manually segmented, and vertebrae from C1 to C4 were individually isolated. From that, 31 distinct distance-intensity maps were created (Figure 5-1). The images were interpolated to equivalent square voxels

PAGE 83

72 before the creation of the maps, so that the distance maps were accurate, and could be compared. Also, no images were selected that contained significant intensity artifacts. A visual comparison of the distance-intensity profiles shows many similarities. It is anticipated that with some training, an algorithm could use this information to properly locate a boundary. The next step would be to formulate a probability density model and implement it into a segmentation routine, most likely that presented by Leventon (2000).

PAGE 84

LIST OF REFERENCES D. Adalsteinsson & J. Sethian, The fast construction of extension velocities in level set methods, The Journal of Computational Physics, 148:2-22, 1999. L.P. Amiot, K. Lang, M. Putzier, H. Zippel & H. Labelle, Comparative results between conventional and computer-assisted pedicle screw installation in the thoracic, lumbar, and sacral spine, Spine, 25(5):606-614, 2000. R. Assaker, N. Reyns, M. Vinchon, X. Demondion & E. Louis, Transpedicular screw placement: image-guided versus lateral-view fluoroscopy: in vitro simulation, Spine, 26(19):2160-2164, 2001. M.S. Austin, A.R. Vaccaro, B. Brislin, R. Nachwalter, A.S. Hilibrand & T.J. Albert, Image-guided spine surgery: a cadaver study comparing conventional open laminoforaminotomy and two image-guided techniques for pedicle screw placement in posterolateral fusion and nonfusion models, Spine, 27(22):2503-2508, 2002. J. Bangham & S. Marshall, Image and signal processing with mathematical morphology, Electronics & Communication Engineering Journal, 10:117-128, 1998. S. Beucher & F. Meyer. The morphological approach of segmentation: the watershed transformation, Mathematical Morphology in Image Processing, E. Dougherty, Ed., chapter 12, 433-481, Marcel Dekker, New York, 1992. A. Blake & M. Isard, Active Contours, Springer-Verlag, London, 1998. A. Bleau & L. Leon, Watershed segmentation and region merging, Comp. Vision and Image Understanding, 77(3):317-370, 2000. G. Boehm, C.J. Knoll, V.G. Colomer, M. Alcaiz-Raya & S.E. Albalat, Three dimensional segmentation of bone structures in CT images, In Proc. SPIE, Medical Imaging, 3661:277-286, 1999. V. Caselles, F. Catte, T. Coll & F. Dibos, A geometric model for active contours in image processing, Numerische Mathematik, 66:1-31, 1993. V. Casellas, R. Kimmel & G. Sapiro, Geodesic active contours, Int. Journal of Computer Vision, 22(1):61-79, 1997. 73

PAGE 85

74 A. Chakraborty, L. Staib & J. Duncan, Deformable boundary finding in medical images by integrating gradient and region information, IEEE Trans. Medical Imaging, 15(6):859-870, 1996, http://noodle.med.yale.edu/staib/papers/chakrabtmi.pdf last accessed March 22, 2005. T. Chan & L. Vese, An active contour model without edges, Scale-Space Theories in Computer Vision, LNCS 1682:141-151, Springer-Verlag, London, 1999. Y. Chen, S. Thiruvenkadam, H. Tagare, F. Huang, D. Wilson & E. Geiser, On the incorporation of shape priors into geometric active contours, In Proc. 1st IEEE Workshop on Variational and Level Set methods in Computer Vision, pp.145-152, IEEE, Los Alamitos, 2001. G. Christensen, R. Rabbitt & M. Miller, Deformable templates using large deformation kinematics, IEEE Trans. on Image Processing, 5(10):1435-1447, 1996. G. Christensen, Consistent linear-elastic transformations for image matching, In Proc. Information Processing in Medical Imaging, 1613:224-237, Springer-Verlag, London, 1999. K. Cleary, Technical Requirements for Image-Guided Spinal Procedures, Workshop Report, April 1999, ISIS Center, Department of Radiology, Georgetown University Medical Center, Washington, D.C., http://www.visualization.georgetown.edu/Spine_Workshop/report/report.html last accessed March 22, 2005. L. Cohen, On active contour models and balloons, CVGIP: Image Understanding, 53(2):211-218, 1991. T. Cootes & C. Taylor, Active shape models smart snakes, In Proc. British Machine Vision Conference, pp. 266-275, Springer-Verlag, London, 1992. T. Cootes, C. Taylor, D. Cooper & J. Graham, Training models of shape from sets of examples, In Proc. British Machine Vision Conference, pp. 9-18, Springer-Verlag, London, 1992. T. Cootes, C. Taylor, D. Cooper & J. Graham, Active shape models their training and application, Computer Vision and Image Understanding, 61:38-59, 1995. T. Cootes & C. Taylor, A mixture model for representing shape variation, Image and Vision Computing, 17(8):567-574, 1999. T. Cootes, C. Beeston, G. Edwards & C. Taylor, A unified framework for atlas matching using active appearance models, In Proc. Information Processing in Medical Imaging, 1613:322-333, Springer-Verlag, London, 1999. T. Cootes, G. Edwards & C. Taylor, Active appearance models, 5th European Conference on Computer Vision, 2:484-498, Springer-Verlag, Berlin, 1998.

PAGE 86

75 D. Cremers, F. Tischhuser, J. Weickert & C. Schnrr, Diffusion snakes: introducing statistical shape knowledge into the Mumford-Shah functional, International Journal of Computer Vision, 50(3): 295-313, 2002. O. Cuisenaire, Distance transformations: fast algorithms and applications to medical image processing, Ph.D. thesis, Universit cotholique de Louvain, Laboratoire de Telecommunications et Teledetection, 1999, http://www.tele.ucl.ac.be/PEOPLE/OC/these.pdf last accessed March 22, 2005. C. Davatzikos & J. Prince, Convergence analysis of the active contour model with applications to medical images, In Proc. SPIE, Visual Communications and Image Processing, 1818:1244-1255, 1992. A. Dempster, N. Laird & D. Rubin, Maximum likelihood from incomplete data via the EM algorithm, J. Royal Statistical Soc., Ser. B, 39(1):1-38, 1977. C.A. Dickman, Point of view, Spine, 25:2674, 2000. E. Dougherty, Mathematical Morphology in Image Processing, M. Dekker, New York, 1993. I. Dryden & K. Mardia, Statistical Shape Analysis, Wiley, Chichester, 1998. K.T. Foley, D.A. Simon & Y.R. Rampersaud, Virtual fluoroscopy: computer-assisted fluoroscopic navigation, Spine, 26:347-351, 2001. G.E. Healthcare Technologies, Waukesha, WI, U.S.A., www.gehealthcare.com/index.html 2004, last accessed March 22, 2005. S. Geman & D. Geman, Stochastic relaxation, Gibbs, and the Bayesian restoration of images, IEEE Trans. on PAMI, 6(6):721-741, November, 1984. N. Glossop & R. Hu, Assessment of vertebral body motion during spine surgery, Spine, 22:903-909, 1997. P. Golland, W.E.L. Grimson & R. Kikinis, Statistical shape analysis using fixed topology skeletons: corpus callosum study, In Proc. Inf. Proc. in Med. Imaging (IPMI), 1613:382-387, Springer-Verlag, 1999. P. Golland, W.E.L. Grimson, M. Shenton & R. Kikinis, Small sample size learning for shape analysis of anatomical structures, In Proc. MICCAI 2000: Medical Image Computing and Computer-Assisted Intervention, LNCS 1935:72-82, Springer-Verlag, 2000. Gray's Anatomy, http://education.yahoo.com/reference/gray/ last accessed March 22, 2005.

PAGE 87

76 M. Jones & T. Poggio, Multidimensional morphable models: a framework for representing and matching object classes, International Journal of Computer Vision, 29(2):107-131, 1998. T. Kaneko, L. Gu & H. Fujimoto, Abdominal organ recognition using 3D mathematical morphology, In Proc. 15th International Conference on Pattern Recognition, 2:263-266, 2000. M. Kass, A. Witkin & D. Terzopoulos, Snakes: active contour models, Int. Journal on Computer Vision, 1:321-331, 1987. A. Kichenassamy, A. Kumar, P. Olver, A. Tannenbaum & A. Yezzi, Gradient flows and geometric active contour model, In Proc. IEEE Int. Conf. Computer Vision, pp. 810-815, IEEE, Los Alamitos, 1995. R. Kikinis, S.K. Warfield & C.F. Westin, High performance computing (HPC) in medical image analysis (MIA) at the surgical planning laboratory (SPL), In Proc. High Performance Computing Asia, pp. 290-297, Singapore, 1998, http://splweb.bwh.harvard.edu:8000/pages/papers/kikinis/hpc/HPCfinal.pdf last accessed March 31, 2005. S. Lavalle, J. Troccaz, P. Sautot, B. Mazier, P. Cinquin, P. Merloz & J.P. Chirossel, Computer-assisted spinal surgery using anatomy-based registration, Computer-Integrated Surgery: Technology and Clinical Applications, R. Taylor, S. Lavallee, G. Burdea & R. Mosges, Eds., pp. 425-449, MIT Press, Cambridge, 1996. R. Leahy, T. Herbert & R. Lee, Application of Markov random fields in medical imaging, Information Processing in Medical Imaging, J. Llacer & D. Ortendahl, Eds., pp. 1-14, Wiley-Liss, NY, 1991. M. Leventon, Statistical models for medical image analysis, Ph.D. Thesis, Massachusetts Institute of Technology, 2000. J.A. Little, D.L.G. Hill & D.J. Hawkes, Deformations incorporating rigid structures, Computer Vision and Image Understanding, 66(2):223-232, 1997. C. Lorenz & N. Krahnstover, D statistical shape models for medical image segmentation, In Proc. Int. Conf. 3-D Imaging and Modeling, pp. 414-423, IEEE, Los Alamitos, 1999. R. Malladi, J. Sethian & B. Vemuri, Shape modeling with front propagation: a level set approach, IEEE Trans. Pattern Anal. Machine Intell., 17:158-175, 1995. G. Matheron, Random Sets and Integral Geometry, John Wiley & Sons, New York, 1975.

PAGE 88

77 T. McInernery & D. Terzopoulos, Medical image segmentation using topologically adaptable surfaces, In Proc. Conf. Computer Vision, Virtual Reality, and Robotics in Medicine and Medical Robotics and Computer-Assisted Surgery (CVRMed-MRCAS), 1205:23-32, Springer-Verlag, Berlin, 1997. T. McInerney & D. Terzopoulos, Deformable models in medical image analysis: A survey, Medical Image Analysis, 1(2):91-108, 1996, http://mrl.nyu.edu/~dt/papers/mia96/mia96.pdf last accessed March 22, 2005. Medtronic Surgical Navigation Technologies, StealthStation, Louisville, CO, www.stealthstation.com last accessed March 22, 2005. F. Meyer & S. Beucher, Morphological segmentation, Journal of Visual Comm. and Image Rep., 1(1):21-46, 1990. M. Miller, A. Banerjee, G. Christensen, S. Joshi, N. Khadeja, U. Grenander & L. Matejic, Statistical methods in computational anatomy, Statistical Methods in Medical Research, 6(3):267-99, 1997. D. Mumford & J. Shah, Optimal approximation by piecewise smooth functions and associated variational problems, Comm. Pure and Applied Math., 42:577-684, 1989. A. Neumann & C. Lorenz, Statistical shape model based segmentation of medical images, Computerized Medical Imaging and Graphics, 22:133-141, 1998. H. Nguyen, M. Worring & R. van den Boomgaard, Watersnakes: energy-driven watershed segmentation, IEEE Trans. on Pattern Analysis and Machine Intelligence, 25(3): 330-342, 2003. Odin Medical Technologies, Inc., Newton, MA, www.odinmed.com last accessed March 22, 2005. K. Ohmori, Y. Kawaguchi, M. Kanamori, H. Ishihara, H. Takagi & T. Kimura, Image-guided anterior thoracolumbar corpectomy: a report of three cases, Spine, 26:1197-1201, 2001. S. Osher & J. Sethian, Fronts propagating with curvature dependent speed: Algorithms based on Hamilton-Jacobi formulations, Journal of Computational Physics, 72:12-49, 1988. N. Paragios & R. Deriche, Geodesic active regions for supervised texture segmentation, In Proc. IEEE Int. Conf. Comp. Vision, pp. 926-932, IEEE, Los Alamitos, 1999. A. Pentland & S. Sclaroff, Closed-form solution for physically based shape modeling and recognition, IEEE Trans. on Pattern Analysis and Machine Intelligence, 13(7):715-729, 1991.

PAGE 89

78 T.M. Peters, Image-guided surgery: from x-rays to virtual reality, Comput. Methods Biomech. Biomed. Engin., 4(1):27-57, 2000. R. Pichumani, Construction of 3D geometric models for segmentation and visualization of cervical spine images, Ph.D. thesis, Stanford University, November 1997. Philips Medical Systems, Achieva I/T Interventional MR, Bothell, WA, http://www.medical.philips.com last accessed March 22, 2005. J. Prewitt, Object enhancement and extraction, Picture Processing and Psychopictorics, B. Lipkin & A. Rosenfeld, Eds., pp. 75-149, Academic Press, New York, 1970. Y.R. Rampersaud, K.T. Foley, A.C. Shen, S. Williams & M. Solomito, Radiation exposure to the spine surgeon during fluoroscopically assisted pedicle screw insertion, Spine, 25:2637-2645, 2000. Y.R. Rampersaud, D.A. Simon & K.T. Foley, Accuracy requirements for image-guided spinal pedicle screw placement, Spine, 26:352-359, 2001. L. Roberts, Machine perception of three-dimensional solids, Optical and Electro-optical Information Processing, J. Tippett, Ed., pp. 159-197, MIT Press, Cambridge, 1965. S. Sclaroff & A. Pentland, Modal matching for correspondence and recognition, IEEE Trans. on Pattern Analysis and Machine Intelligence, 17(6):545-561, 1995. J. Serra, Image Analysis and Mathematical Morphology, Academic Press, New York, 1982. J. Sethian, Level Set Methods: Evolving Interfaces in Geometry, Fluid Mechanics, Computer Vision, and Material Science, Cambridge University Press, New York, NY, 1996. J. Sethian, A fast marching level set method for monotonically advancing fronts, Proc. Nat. Acad. Sci., 93(4):1591-1595, 1996. J. Sethian, Fast marching methods, SIAM Review, 41(2):199-235, 1999. Siemens Medical Solutions USA, Siemens SIREMOBIL Iso-C3D, Inc., Iselin, NJ, http://energized.esiemenshealthcare.com/SIREMOBIL/iso3d.htm last accessed March 22, 2005. I. Sobel, Camera models and machine perception, Technical Report AIM-21, Stanford Artificial Intelligence Lab, Palo Alto, California, May 1970. L. Staib & J. Duncan, Boundary finding with parametrically deformable models, IEEE Transactions on Pattern Analysis and Machine Intelligence, 14:161-175, 1992.

PAGE 90

79 G. Szekely, A. Kelemen, C. Brechbuhler & G. Gerig, Segmentation of 2-D and 3-D objects from MRI volume data using constrained elastic deformations of flexible Fourier contour and surface models, Med Image Anal., 1(1):19-34, 1996. H. Tek & B. Kimia, Image segmentation by reaction diffusion bubbles, Proc. Int. Conf. Computer Vision, pp. 156-162, IEEE, Los Alamitos, 1995. A.R. Vaccaro & K. Singh, Computer-assisted image-guided surgery: application in cervical spine surgery, Current Opinion in Orthopaedics, 12:245-250, 2001. L. Vincent & P. Soille, Watershed in digital spaces: an efficient algorithm based on immersion simulations, IEEE Trans. on PAMI, 13(6),583-598, 1991. Y. Wang & L. Staib, Boundary finding with correspondence using statistical shape models, In Proc. IEEE Conf. Comp. Vision and Patt. Recog., pp. 338-345, IEEE, Los Alamitos, 1998. Y. Wang & L. Staib, Elastic model based non-rigid registration incorporating statistical shape information, In Proc. Medical Image Computing and Computer-Assisted Intervention (MICCAI), LNCS 1496:1162-1173, Springer-Verlag, London, 1998. S. Warfield, A. Robatino, J. Dengler, F. Jolesz & R. Kikinis, Nonlinear registration and template driven segmentation, In Brain Warping, Arthur W. Toga, Ed., Ch.4:67-84, Elsevier Science & Technology Books, St. Louis, 1998. A. Weidner, M. Whler, S.T. Chiu & C.G. Ullrich, Modification of C1-C2 transarticular screw fixation by image-guided surgery, Spine, 25:2668-2674, 2000. W.C. Welch, B.R. Subach, I.F. Pollack & G.B. Jacobs, Frameless stereotactic guidance for surgery of the upper cervical spine, Neurosurgery, 40:958-964, 1997. C.F. Westin, S. Warfield, A. Bhalerao, L. Mui, J. Richolt & R. Kikinis, Tensor controlled local structure enhancement of CT images for bone segmentation, Int. Conf on Medical Image Computing and Computer-Assisted Intervention (MICCAI'98), LNCS 1496:1205-1212, Springer-Verlag, London, 1998. C.F. Westin, J. Richolt, V. Moharir & R. Kikinis, Affine adaptive filtering of CT data, Medical Image Analysis, 4:161-177, 2000, http://splweb.bwh.harvard.edu:8000/pages/papers/westin/media2000/westinMedia2000.pdf last accessed March 22, 2005. A. Yezzi, A. Tsai & A. Willisky, A statistical approach to image segmentation for bimodal and trimodal imagery, In Proc Int. Conf. Comp. Vision, pp. 898-903, IEEE, Los Alamitos, 1999. A.S. Youkilis, D.J. Quint, J.E. McGillicuddy & S.M. Papadopoulos, Stereotactic navigation for placement of pedicle screws in the thoracic spine, Neurosurgery, 48:771-779, 2001.

PAGE 91

80 S.C. Zhu & A. Yuille, Region competition: unifying snakes, region growing, and Bayes/MDL for multiband image segmentation, IEEE Trans. on PAMI, 18(9): 884-900, 1996.

PAGE 92

BIOGRAPHICAL SKETCH Matthew James Williams, born April 13, 1975, received a Bachelor of Science in Engineering in mechanical engineering, with concentrations in design and biomedical applications, from the University of Michigan in 1998. During his undergraduate years, he was blessed with many rewarding experiences and opportunities that have helped to shape his life. Some of these include internships at a variety of manufacturing and technical facilities, capped with a cooperative at the Denso Tech Center in 1997. After graduation, Matthew was hired as a project management consultant at ACM, Inc. Throughout his career, he has been cementing his resolve to explore engineering opportunities in medicine. This led him to the University of Florida Biomedical Engineering Program; and in 2000, he migrated to Gainesville. Now that he has finished the requirements for his masters degree, Matthew has elected to continue taking advantage of the cornucopia of opportunities at the University of Florida through the study of the regenerative medicine aspect of biomedical engineering. Also, Matthew continues the pursuit of the performing arts, nourishing creativity in all aspects of his life. 81


Permanent Link: http://ufdc.ufl.edu/UFE0010401/00001

Material Information

Title: Segmentation in 3D Virtual Spine Modeling for Assistance in Surgical Planning and Guidance
Physical Description: Mixed Material
Copyright Date: 2008

Record Information

Source Institution: University of Florida
Holding Location: University of Florida
Rights Management: All rights reserved by the source institution and holding location.
System ID: UFE0010401:00001

Permanent Link: http://ufdc.ufl.edu/UFE0010401/00001

Material Information

Title: Segmentation in 3D Virtual Spine Modeling for Assistance in Surgical Planning and Guidance
Physical Description: Mixed Material
Copyright Date: 2008

Record Information

Source Institution: University of Florida
Holding Location: University of Florida
Rights Management: All rights reserved by the source institution and holding location.
System ID: UFE0010401:00001

Full Text












SEGMENTATION IN 3D VIRTUAL SPINE MODELING FOR ASSISTANCE IN
SURGICAL PLANNING AND GUIDANCE















By

MATTHEW JAMES WILLIAMS


A THESIS PRESENTED TO THE GRADUATE SCHOOL
OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT
OF THE REQUIREMENTS FOR THE DEGREE OF
MASTER OF ENGINEERING

UNIVERSITY OF FLORIDA


2005

































Copyright 2005

by

Matthew James Williams















ACKNOWLEDGMENTS

There are many people that I am pleased to recognize who have had a very

positive influence on my journey at the University of Florida. First, I would like to thank

Dr. Frank Bova for his guidance and patience, and for making possible a very clinically

interactive biomedical engineering experience. It is the kind of project I had hoped for

when I began my studies at UF. Dr. Wesley Bolch and Dr. David Gilland both have my

sincere appreciation for their commitment to come on board at such a late stage to help

out. Literally, I could not have finished without them. Also, I would like to extend

gratitude to Dr. Bernard Guiot, for introducing me to neurosurgery; Dr. Lionel Bouchet,

for his tutelage and example of tireless dedication; Drs. Yunmei Chen and Bernard Mair,

for lending their ears for mathematical reflection; and Drs. Christopher Batich and

Anthony Brennan, for their continued support and words of motivation. In addition, I

would like to acknowledge the financial contributions of the Department of

Neurosurgery, the Department of Biomedical Engineering, and the Whitaker Foundation.

I have encountered and have had the pleasure of knowing many friendly

personalities in the lab, in the operating room, in the classroom, and in various other

nooks and crannies, and each has helped to make this a memorable experience; and for

that I give thanks. Finally, I would like to extend a heartfelt appreciation to Alan

Wineman, Jennifer Kadlowec, and Jeanette Clute for their belief in my abilities; my

fraternity brothers, for their companionship; and my exceptional friends Matt, JP, Vishal,

Christina, Chris, Tom, and Kent for their encouragement and distracting adventures. And









most of all, I am indebted to my brother, my parents, and the rest of my loving family as

they are the foundation of my strength; and to Jennifer, the delightful blend of sunshine

and spice that has helped me to believe in all of the wonderful things in life.
















TABLE OF CONTENTS

page

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

L IST O F F IG U R E S .... ...... ................................................ .. .. ..... .............. vii

LIST OF OBJECTS ......... ......................... ...... ...... ............. ix

A B ST R A C T .......... ..... ...................................................................................... x

CHAPTER

1 IMAGE-GUIDED SPINE SURGERY..................................... ........................1

Preoperative Im age G uidance........................................................... ............... 2
Intraoperative Im aging ........... ...... ...................................... .............. ................ .5
Fluoroscopy ................................ ................................. ........ 5
C om puted T om ography .............................................................. .....................6
M agnetic R esonance Im aging ........................................ .......... ............... 7
U ltrasou n d ................................................ .......................... 7
Com puter-assisted Surgery System s....................................... .......................... 8
V irtual Surgical M odel .......................................................................8
Su rg ical P lan n in g ............ ............................................................ .. .... .. .. .. .. 10
R registration and Tracking .............................................................. ... ............ 11
Surgical G guidance ........................................ .. .. .... ........ ......... 14
Surgical Application .................. ......................... .. ....... ................. 14
S y stem s .................. ................................................... ................ 17
Segmentation Introduction.......................................... 18

2 SEGMENTATION BACKGROUND....................................................................20

Im age Processing ...................................................................... ......... 21
M them atical M orphology .............................................................. .....................22
B inary O perations......... .............................................................. ..... .... .. 24
G rayscale O perations........... ................................................ ..................... 28
Voxel-based Segm entation M ethods ........................................ ...................... 29
T h re sh o ld in g ................................................. ............... ................. 3 0
M orphological Segm entation ........................................ ......................... 32
E dge-based Segm entation ............................................................................ ...... 34
R egion-based Segm entation ............................................... ............................ 39


v










Shape-based Segm entation ............................................................. .............. 40
A tlas W arping................................................... 40
M odeling.................................................. 4 1
M e th o d s .....................................................................................................4 2

3 SEGM EN TA TION ................................. .....................................44

P problem E valuation ..............................................................45
M ethod Selection .................................................................. ... ......... 47
R research T ools.....................................................4 8
Cursory Exam nation ............................................................................ 50

4 R E G IO N G R O W IN G ............................................................................. 53

Kernel Definition .................................................................. ... ......... 53
S h ap e C o rre latio n .................................................................................................. 5 3
Hy steresis .............. ........ ........... .................. 56
Algorithm Enhancement ............... ......... .......... ......... 59
Im ag e V ariatio n s .................................................................................................... 6 0
P re lab e lin g ......................................................................................................6 2
Instrumentation ............... ......... .......................65
Results ............. ..................... ............. ...............65

5 CONCLUSIONS AND FUTURE WORK .... .................... ........69

L IST O F R E F E R E N C E S .............. ..... ............ ........................................................... 73

B IO G R A PH IC A L SK E T C H ........................................................................................ 81
















LIST OF FIGURES

Figure p

1-1 Change in anatomical orientation occurring during transarticular screw
p lacem en t. .......................................................... ................ .. 4

1-2 Segm entation of a surgical m odel ........................................ ......................... 9

1-3 Surgical plan ................... .............. .... ........ ........................ 11

1-4 Registration of the surgical model to the patient................................. ...............12

1-5 A ctive surgical guidance .......................................................... ............... 14

2-1 Intensity histograms and histogram equalization....................................................21

2-2 Morphological structuring elements and neighborhood configurations ..................24

2-3 Binary m orphological dilation ........................................... .......................... 25

2-4 Compound morphological operations ........................................... ............... 27

2-5 M orphological distance transform s.................................. ..................................... 28

2-6 Threshold m modeling .................. ..................................... ... ............ 31

2-7 Binary im age segm entation ......................................................... .............. 33

2-8 Singularity in contour evolution........................................ ........................... 37

2-9 Level set application for contour evolution.........................................................38

3-1 B building a threshold m odel ........................................................... .....................45

3-2 M annual segm entation and display .................................... ............. ........ ....... 46

3-3 Surface characteristics of vertebrae....................................................................... 47

3-4 Toolset GUI .................. ................................... ........... .. ............ 49

3-5 3D intensity gradient ........................ ........ .. ........51









4-1 M anu al kernel selection ........................................ .............................................54

4-2 Segm entation using distance m aps.................................... ......................... 55

4-3 High threshold values result in diminished kernels .............................................56

4-4 Region growing algorithm iteration .............................................. ................. 57

4-5 Algorithm input processing sequence ........................................... ............... 58

4-6 R region grow ing ............... ........ ............................................... ............. 59

4-7 Open and close operators are used to alter the evolution of the growth algorithm..60

4-8 Unique properties of the input CT image may be isolated or combined to
produce a m ore favorable segm entation........................................ ............... 62

4-9 Segmentation results using variations on the input CT image...............................63

4-10 Instrumentation subtraction....................................................................... 66

4-11 Segmentation failure: boundary leakage..................................... ............... 67

4-12 Segmentation failure: weak bone intensity profile.................. ................68

5-1 Intensity/distance-to-boundary profiles................................................................ 71
















LIST OF OBJECTS


Object page

4-1 High threshold movie (objectl-highthreshold.mpg, 27.6 MB).............................56

4-2 Region growing movie (object2-success.mpg, 25.2 MB) ............... .....................59

4-3 Boundary leakage movie (object3-failureleakage.mpg, 25.8 MB)......................67

4-4 Weak bone intensity profile movie (object4-failureweak.mpg, 28.6 MB)...............68















Abstract of Thesis Presented to the Graduate School
of the University of Florida in Partial Fulfillment of the
Requirements for the Degree of Master of Engineering

SEGMENTATION IN 3D VIRTUAL SPINE MODELING FOR ASSISTANCE IN
SURGICAL PLANNING AND GUIDANCE

By

Matthew James Williams

May 2005

Chair: Frank J. Bova
Major Department: Biomedical Engineering

Surgical procedures of the spine are technically demanding, requiring precise

navigation to avoid critical vascular and nerve structures. Often, image guidance systems

are employed to improve accuracy of surgical tool placement and increase the likelihood

of a successful outcome. The most commonly used systems generate a patient-specific

virtual 3D model. This model is used to create a surgical plan that, during the operation,

safely guides the surgeon in placement of the surgical tools. Unfortunately, the current

modeling technique limits the application of image-guided surgery.

The typical surgical model is created using the thresholding technique performed

on a pre-operatively acquired diagnostic CT image. The result is a single rigid body

model that appears as a surface-rendered outline of the involved bones. During the

procedure, this model is registered or aligned with the patient. The drawback to the rigid

model is its inability to account for the mechanical flexibility of the spine. Since the

patient is likely to change positions between image acquisition and placement for









surgery, errors are introduced into the guidance. In order for the surgical model to

properly deform, the virtual bones must be segmented and individually registered to the

vertebrae. The current method of segmentation is very time-consuming and not

clinically feasible on a routine basis. The unfortunate consequence is that, given

accuracy restrictions, image-guided procedures of the spine are generally limited to one

or two vertebral levels. It is the purpose of this research to develop a method that

provides a segmented model for use in image-guided surgery. It is desired that this

method be easy to work with, provide swift results, and require minimal user

intervention.

The segmentation method developed here utilizes a simple region-growing scheme.

It dilates seed regions that are uniquely assigned to the underlying bones. This allows the

growing regions to retain their identity while expanding into the full bone profile. The

method was tested on 31 high-resolution clinical scans. Over 80% of the resulting

segmentations showed significant improvement over the standard manual methods.

Success was gauged by the amount of time and effort required to achieve a segmentation

using the method as compared to manual methods. The factors contributing to

segmentation failures are attributed to poor bone resolution and inadequate starting

information. This method showed marked success in the segmentation of the CT-derived

models used in guided spine surgery. This method may be employed to significantly

reduce segmentation time and facilitate surgical image-guided applications to multi-level

spinal procedures.














CHAPTER 1
IMAGE-GUIDED SPINE SURGERY

Surgical procedures of the spine are technically demanding. They often involve

placement of rigid instrumentation in close proximity to critical vascular and nerve

structures and accordingly, demand a high degree of accuracy. In order to avoid

complications, the neurosurgeon or orthopedic surgeon must accurately plan the

trajectories for placement of this instrumentation. This requires visualization of the

involved anatomy in three dimensions. Physical inspection allows the surgeon to make

an initial assumption of specific anatomical orientation. However, imaging systems

verify this assumption without unnecessary surgical exposure. In addition, they provide

crucial information such as condition of vasculature or other hidden structures or extent

of pathology such as fracture dimensions and tumor location.

Image-guided surgery, at its simplest, is the use of an imaging modality to facilitate

surgical intervention. Systems dedicated to image guidance, though, actually allow non-

invasive evaluation and planning through a virtual surgical environment. In addition,

during surgery, these systems direct the surgeon's tools to the proper trajectories. The

use of image guidance systems in surgery, through increased accuracy and precision, has

reduced the risk of damage to critical areas and increased the ability of the surgeon to

handle technically difficult surgeries. In addition to a reduction in the duration of the

operation, these advancements have led to significant reductions in patient morbidity and

mortality (Cleary, 1999).









The most common image guidance systems utilize a model of the vertebrae

generated from a diagnostic Computed Tomography (CT) image. It appears as a surface-

rendered image of the bone profile. Unfortunately, this model is not able to deform, or

account for any of the natural motions that may occur with the spine. Since motion of the

patient is likely to occur after image acquisition, the model of the vertebrae constructed

may not adequately reflect the orientation of the vertebrae presented during surgery. This

limits the accuracy of the model and its utility in guiding surgical procedures. In order to

create a spine model that can properly reorient or deform each of the vertebrae must be

isolated. The process of identifying these individual vertebrae in the model is called

segmentation. The current method of segmenting spinal models is a labor-intensive,

time-consuming procedure, which is infeasible on a routine clinical basis. The purpose

of this research is to create a more user-friendly method for segmenting the models

intended for image-guided surgery. The goal is to provide results in a more suitable time

frame while requiring minimal user intervention.

Preoperative Image Guidance

The standard of care is the use of a presurgical scan for image-guided surgery.

Typically, plain radiographs are supplemented with an additional modality such as

myelography, computed tomography, or magnetic resonance imaging (MRI). Each

modality provides another layer of information that the surgeon can utilize. However, the

use of both CT and MRI is discouraged because of cost constraints (Cleary, 1999). The

preoperative scans that are at the surgeon's disposal are generally used to mentally

construct a 3D representation of the anatomy. This envisioned model is associated with

the anatomy exposed during the operation. It allows the surgeon to proceed with more

confidence without relying on any anatomical statistical convention. This is especially









true in cases of severe pathologic deformation. Overall, the greater the knowledge of

anatomical positioning, the less invasive a procedure must be to assess the correct

orientation of the involved elements (Lavallee et al., 1996). The accurate determination

of vertebral alignment is essential for a successful surgical outcome. Of the modalities

available for presurgical imagery, CT and MRI are typically used for spinal applications.

They provide multiple 2D slices that can be mentally stacked to form a 3D model. CT

has excellent bone and soft tissue contrast, yet requires ionizing radiation (Cleary, 1999).

MRI, on the other hand, provides soft tissue contrast without the ionizing radiation and

bone flare usually associated with CT. They are both susceptible to artifacts, such as

non-uniformity distortions in MRI and starburst patterns in CT.

The primary drawback of preoperative imaging used for surgery is that it does not

account for any anatomical alignment changes that occur after imaging. Any changes of

positioning that occur detract from the accuracy of the representation. This change of

positioning can be isolated from two sources. One is the movement of the patient

between the presurgical scan and the operative positioning. The presurgical scan is taken

in the supine position to minimize breathing artifacts. A process called "reduction"

whereby the surgeon optimally orients the patient for the procedure determines the

operative position. The second source of position change occurs during the operation.

Throughout the procedure the vertebrae can experience significant surgeon-induced

motion, especially in the case of spinal instability (Glossop & Hu, 1997). Patients with

unstable spines may exhibit abrupt translations of spinal segments, and these movements

are difficult to model and predict (Cleary, 1999). Accounting for these alignment

changes is especially critical in cases where two vertebral bodies are to be fixed by the









Table 1-1. Change of orientation between the vertebrae during a surgical procedure.
Presurgical and postsurgical CT scans were compared, and the change of the
positions of Cl relative to C2 was recorded. The translation and rotation
between the two vertebrae in orthogonal planes are shown for three patients.

Patient 1 Patient 2 Patient 3
AP Lat. SI AP Lat. SI AP Lat. SI
Trans. (mm) 2.7 -6.1 3.7 -9.3 3.8 7.4 -6.7 0.7 10.0
Rot. (degree) 3.0 -5.4 -6.5 0.8 -14.0 9.7 -1.9 -18.0 2.4


Figure 1-1. Change in anatomical orientation occurring during transarticular screw
placement. A profile of vertebral positioning is taken from the presurgical CT
scan (left) and overlaid onto the postsurgical CT scan (right). A shift in
position can be seen between Cl and C2.

same screw (Figure 1-1). Table 1-1 quantifies the changes of position that occur after

imaging by comparing preoperative and postoperative CT scans. Even though this

particular example only involves two vertebral levels, it is apparent that the use of pre-

operative imaging for image guidance has a limited accuracy. There are, however, two

possible solutions for managing the problem of spinal motion after image acquisition.

The first is intraoperative imaging and the second is the use of a guidance system that

tracks each individual vertebral body.









Intraoperative Imaging

Imaging systems can be used to acquire scans during the operation and provide an

instant indication of anatomical positioning and surgical tool location. Despite this

advantage, however, there are certain considerations such as cost, intraoperative image

quality, and impact on the operating room environment that must be considered before

the investment in such a system.

Fluoroscopy

Fluoroscopy is the dominant intraprocedural imaging modality for spinal surgery.

The unit is a common piece of equipment that easily fits around the table and provides

minimum disruption to the operative field or operative technique. It is a low-cost, useful

and familiar technology. It is also easily accessible and portable (Foley et al., 2001;

Cleary, 1999). In the application of spinal surgery, fluoroscopy is commonly used to

verify vertebral alignment and instrumentation and tool positioning. Fluoroscopy units

provide high resolution, large field of view (FOV) scans in real time that have excellent

bone to soft tissue contrast (Cleary, 1999). There are some drawbacks however, that

limit the utility of x-ray fluoroscopy. The primary one is fluoroscopy only acquires a 2D

image in one plane, and hence only gives one aspect of 3D positioning. It is impractical

to gauge depth in the picture, as it is a display of overlapping tissues. In order to image

in additional planes, the fluoroscope must be repositioned repeatedly (Foley et al., 2001).

Also, fluoroscopic images have poor soft tissue discrimination, which makes the

visualization of vascular and nerve structures difficult (Cleary, 1999). Finally, the

capability of imaging in real time is at the expense of constant x-ray radiation exposure.

In the case of spine surgery, where there may be long periods of imaging activity, there is









significant radiation exposure to the patient and operating team (Foley et al., 2001;

Rampersaud et al., 2000; Cleary, 1999).

Computed Tomography

Computed tomography is an imaging modality in which a three dimensional image

is constructed from a series of plane cross-sectional radiographic images made along an

axis. The scans are high resolution and have excellent bone to soft tissue contrast.

Intraoperatively, CT provides an excellent view of bony structures and their relationships,

and can accurately localize the tip of interventional instruments (Cleary, 1999).

However, the imaging is not real time and subjects the patient and those in the scanning

field to ionizing radiation.

There are three main types of CT scanning systems that are used during surgery.

The first is spiral CT. It is very fast with a large bore and excellent image quality, yet

requires high capital and maintenance costs (Cleary, 1999). In addition, it is a large,

fixed machine that has limited accessibility in a surgical environment. Mobile CT, on the

other hand, is smaller, portable, and has a comparatively lower radiation dose. The

disadvantages to this imaging system are slower acquisitions, decreased tube capacity,

and lower image quality, which can lead to registration difficulties (Cleary, 1999).

Mobile CT is costly as well. The last type of CT imaging systems uses a sweeping

fluoroscope to generate images. This fluoro-CT has the advantage of quick

reconstruction and display for 3D imaging as well as easy patient access and targeting. It

offers a low cost, low patient dose alternative to spiral CT or mobile CT. However,

fluoro-CT has comparatively poor tissue contrast which results in minimally acceptable

bone reconstruction. Two examples of Fluoro-CT systems are the SIREMOBIL Iso-C3D

(Siemens Medical Solutions USA, Inc., Iselin, NJ) and FluoroCATTM (G.E. Healthcare









Technologies, Waukesha, WI). Given these options, there is still one primary limiting

factor in the use of intraoperative CT for spinal alignment determination. Scans from the

typical surgical position are subject to breathing artifacts, and the resulting images are

unsuitable for accurate image guidance.

Magnetic Resonance Imaging

Magnetic resonance imaging is a non-ionizing imaging modality that can be

employed intraoperatively. The images are high-resolution with excellent soft tissue

contrast. Safety considerations with the magnetic fringe field however, make it difficult

to adapt to the surgical environment. The need for specialized tools and equipment that

can operate in the intense magnetic field and radio frequency-rich environment make it a

costly endeavor. This is in addition to the intrinsic cost of the device. Also, specific to

guidance, there is insufficient tool tip viewing accuracy. MRI images have poor

definition of bony structures, making edge interpretation difficult. The correct

determination of bone boundaries is critical in spinal surgery. Finally, MRI has its own

handful of potential artifacts that must be considered (Cleary, 1999). An example of an

MR system optimized for surgical application is the Achieva I/T Interventional MR

(Philips Medical Systems, Bothell, WA). Two other examples that allow intraoperative

use are the Signa SP (G.E. Healthcare Technologies, Waukesha, WI) and the

MAGNETOM Concerto (Siemens Medical Solutions USA, Inc., Iselin, NJ). Similar to

computed tomography, this imaging modality is also susceptible to the breathing artifacts

that plague scans performed in standard operative position.

Ultrasound

Ultrasound is an inexpensive, easily portable, real-time imaging modality that does

not rely on ionizing radiation. However, the poor image quality, weak discrimination of









certain critical spinal tissues, and reliance on operator skill, does not make it a candidate

for precise surgical visualization. Also, the use of ultrasound for surgery often requires

increased intrasurgical access, which should be avoided if at all possible (Cleary, 1999).

This modality is mentioned because it has utility in being used for image to patient

registration, a process necessary for computer-assisted surgery.

Computer-assisted Surgery Systems

The problem of realizing spinal motion after pre-operative image acquisition is a

primary concern for accurate surgical intervention. In response, computer-assisted

surgical guidance systems have been developed which actively track spinal movement

and relay real-time positioning of the vertebrae to the surgeon. Unlike the intraoperative

imaging systems, which typically operate over discrete periods of time with some degree

of surgical field interruption, these systems relay positioning information throughout the

critical portion of the surgical intervention. Image-guided surgical systems accomplish

their task by overlaying a virtual surgical field with the patient's operative field. The

computer-generated virtual field is created preoperatively and includes a geometric model

of the involved anatomy with visual indications of intended surgical pathways and targets

of instrumentation. Once this virtual field is matched or registered intraoperatively to the

physical surgical field, the surgeon can compare his surgical tool placement with the

plan. The ability of the model to properly represent the spinal anatomy has a direct effect

on the accuracy of the surgical guidance.

Virtual Surgical Model

The most preferred anatomical model used for spine surgery is one that best

approximates the mechanical flexibility of the spine. The spine should be modeled as a

series of rigid non-deformable bodies, vertebrae, connected by deformable tissue, muscle,


























Figure 1-2. Segmentation of a surgical model. A rigid model of the cervical vertebrae is
created using the thresholding technique (left). An applied segmentation
method produces a model that properly represents the anatomy (right).

cartilage, and ligaments. The model used for guided surgery, however, only needs to

include each individual vertebra as isolated bodies, so that they can slide, or translate and

rotate relative to each other. Such a model best reflects the motion seen during the

operation.

The computer-based anatomical model is typically created from a presurgically

acquired CT image. As indicated by Peters (2000), a high-resolution, three-dimensional

image is best suited to clearly represent the patient's anatomy. For example, a CT scan

intended for use in spine surgery has a .6 to 1.2 mm slice thickness with a 0.2 0.3 mm

in plane pixel size. These voxel dimensions (0.3 x 0.3 x 0.3 mm) equate to an

approximate 18 cm field of view with 512 x 512 pixel images.

Currently, thresholding is the scheme most commonly used to create the anatomic

virtual model used in spine surgery. It is a convenient procedure that takes little of the

clinician's time. Thresholding is a process that creates a boundary at disjoint intensity

ranges such as between bone and soft tissue on CT images. It is easy to indicate the









proper bone boundary using this method, however, it often leaves adjacent bones as one

contiguous mass (Kikinis et al., 1998). The resulting single rigid body model, once

registered to the patient, is ill suited to account for any movement of the spine that may

occur during the surgical procedure, which potentially introduces errors into the surgical

guidance. This challenges the efficacy of the guidance system and presents the need for a

segmented model. The single rigid body model created through thresholding can be

converted into a model that allows for flexibility. This process is called segmentation, or

the division of the single spinal bone model into its constituent vertebrae (Figure 1-2).

The resulting composite model, if properly matched to the patient, will allow the

orientation of the vertebrae to be tracked intraoperatively. Currently, segmentation is a

time-consuming process in which the clinician must use basic processing tools to

manually contour the boundaries of each vertebra; there is to date no automatic, reliable,

and robust method of doing so. Therefore, the development of real-time image

processing techniques for model creation is of primary importance for image-guided

surgery (Cleary, 1999).

Surgical Planning

Once an adequate virtual model is created, it can be used to aid treatment selection

(Cleary, 1999). The surgeon will use the model to visualize inside structures, and

establish optimal surgical pathways. Preparation involves the placement of virtual tools,

which represent the actual instruments used in surgery. The surgeon can plan the

incision, define resection margins, and determine the appropriate orientation for

instrumentation (Welch et al., 1997). This includes identifying working corridors that

provide adequate access while minimizing the risk of damage to fragile tissues (Cleary,

1999). An example of tool placement is shown in Figure 1-3. After the planning phase


































Figure 1-3. Surgical plan. Pedicle screw trajectory is planned using a virtual tool. Note
the surgical model in the lower-right corer.

of surgical treatment, the virtual model must be matched, or registered to the patient to

allow for guidance.

Registration and Tracking

Registration is the mapping of coordinates between any two spaces. In a guided

surgery application, it is the process that links the coordinate frame of the virtual

computer field to the surgical field. Registration is done so the displayed computer

model of the involved anatomy and overlaid virtual surgical tools accurately represents

the physical operation. A popular system at the University of Florida accomplishes

registration using a stereotactic, optical camera system. It tracks precise positioning of

the anatomy and surgical tools in three-dimensional space using reference frames. These

frames have attached light emitting diodes (LEDs) that allow the infrared (IR) camera to






















Figure 1-4. Registration of the surgical model to the patient. The reference frames
attached to the probe and vertebra are used by the IR camera system to track
stereotactic positioning. The surgeon uses a probe to touch various points on
an exposed vertebra for surface registration (left). The corresponding points
are reflected on the virtual model (right).

track their position and pose. There are several steps taken to perform registration for

this system. First, a reference frame is attached to an exposed spinal process and to each

of the tools involved in the guided surgery. Since the thresholding process creates a

single rigid-body model, only one vertebra can be registered. Therefore, for best

utilization, the model is registered to the most significant or important vertebral level.

Next, the computer-built virtual anatomic model must be registered or geometrically

associated to the patient's selected vertebra. Registration is required because the

computer is unaware of the actual position of the anatomy relative to the attached

reference frame. Registration for the tracked tools is much simpler because of

predefined attachment sites for the reference arrays.

The primary registration method for the spinal system is point matching. A set of

points, usually landmarks, are identified on the virtual vertebrae and then matched to

corresponding points on the patient's vertebrae using a tracked probe. Once point

matching is completed, and an estimate of the association between the surgical model and

exposed vertebrae is achieved, an additional registration step, surface matching, may be









employed. In order to surface match, the surgeon touches a multitude of points on the

surface of the exposed vertebrae (Figure 1-4). These points are used to create a surface

profile, which is then aligned to the surface of the virtual model. This step may improve

the accuracy of the registration. Once registration is complete, whichever vertebra the

reference frame is attached to is tracked dynamically to account for changes in the

patient's vertebral orientation. Registration assures that the virtual surgical tools, which

correspond to the real tools, are properly overlaid upon the spine model as the operation

is performed.

Registration is usually limited to one vertebral level. This is due to both current

modeling restrictions and the ability of the tracking system to easily recognize multiple

rigid bodies. Therefore, to preserve accuracy, the use of surgical navigation is generally

reserved for those procedures that only involve a couple of vertebral levels. An increase

in availability of segmented models, however, will encourage registration and tracking of

multiple vertebral levels. This will allow the benefits of surgical guidance to be applied

to a broader scope of procedures.

Alternate techniques exist for tracking and registration of the spine. One

noteworthy tracking system uses electromagnetic (EM) markers to locate anatomy.

These point localizers minimize signal blockage commonly associated with the

cumbersome IR reference frames, but may be susceptible to EM interference and the

presence of magnetic or ferrous objects. Registration, alternatively, can be accomplished

with a localized imaging system. Specifically, ultrasound can be used to create a surface

profile for use in the surface-matching method of registration. One novel development

presented by Medtronic is the FluoroMerge software (Medtronic Surgical Navigation





















Figure 1-5. Active surgical guidance. A targeting screen (right) guides the surgeon to
the proper trajectory of the surgical drill guide (left).

Technologies, Louisville, CO). It can perform the registration of a preoperative CT

image using only two fluoroscopic images. This not only provides a quick, "hands-free"

registration method, but may also be useful for multiple rigid body registration.

Surgical Guidance

Once all of the steps of preparation (model creation, planning, registration) have

been completed, the system is ready to provide the surgeon with the feedback necessary

to perform a successful operation. A visual terminal in the operating room allows the

surgeon to simultaneously view surgical tool placement relative to both the virtual model

and the patient. Also included is a targeting screen that compares the trajectories of tools

relative to the surgical plan (Figure 1-5). This feedback assures proper placement of the

instrumentation during the surgical procedure.

Surgical Application

There are several benefits to the use of image guidance in surgery. First, it

provides a multidimensional view of anatomic relationships in the operative field,

including extent of bone and soft tissue resection (Welch et al., 1997). It is especially

useful when traditional surgical landmarks are obscured or altered, as may be the case

with pathology or bony fusion (Austin et al., 2002). Image guidance reduces the need for









exposure and increases the confidence of the surgeon during the procedure (Welch et al.,

1997). It also improves accuracy of the surgical tool placement, which leads to the

reduced risk of structurally significant violations such as neurovascular injury.

Ultimately, image guidance lessens the technical difficulty of spine surgery and allows

for more safe and effective outcomes (Assaker et al., 2001; Ohmori et al., 2001; Youkilis

et al., 2001; Welch et al., 1997).

Currently, there are three general classifications of spinal surgery that take

advantage of image-guidance. The first is decompression, or removal of anatomical

pressure on the spinal cord. This category has the highest volume of cases. Stabilization

is the category with the next highest volume of procedures. Surgical intervention is

almost always required when instability occurs below the level of C2. The final group is

deformity correction, which has the highest risk of undesirable outcomes (Cleary, 1999).

The hard tissues manipulated in image-guided surgical procedures often include the

vertebral body, facet joints, iliosacral joint, and intervertebral disc. The pathology

addressed during the procedure varies from fracture to inflammation and may involve the

spinal cord and other adjacent soft tissues (Cleary, 1999). The use of computer-assisted

surgery has especially seen use in cervical procedures, given the need for a high level of

accuracy. Vaccaro & Singh (2001) discusses various applications in cervical spine

surgery. He states the primary use for computer-assisted surgery is placement of C2-C1

transarticular screws for atlantoaxial fusion in order to minimize the risk of injury to the

vertebral artery. Other infrequent uses include transoral odontoid resections, cervical

subaxial pedicle screw placement, and anterior cervical corpectomies.









Spinal surgery often includes placement of instrumentation such as screws, rods,

hooks and wires. Screws can be used alone to repair fracture or instability as in C1-C2

transarticular screw fixation, or as an anchor for further instrumentation as in the case of

pedicle screw fixation. Pedicle screws are often used in the treatment of pathological

conditions such as arthritic deformity (spondylosis), fractures (iliosacral or vertebral), and

can be used to support bony fusion (arthrodesis) (Cleary, 1999). In certain levels of the

spine, extremely high accuracy is needed for the placement of pedicle screws to avoid

perforation of the pedicle wall (Rampersaud et al., 2001). In many cases, misplacement

of the screw can result in vertebral artery injury (Weidner et al., 2000). Cervical

procedures are even more technically difficult as the screw trajectory is in very close

proximity to the spinal canal, vertebral artery, and spinal nerve root. Image-guided

surgery has been shown to improve the placement of pedicle screws and reduce the risk

of screw misplacement (Austin et al., 2002; Youkilis et al., 2001; Weidner et al., 2000;

Amiot et al., 2000; Henderson et al., 1996).

There are a few considerations in the use of image-guidance for spinal surgery.

First, navigation systems eliminate the need for repetitive intraoperative fluoroscopy for

tool placement, dramatically reducing radiation exposure (Foley et al., 2001; Welch et al.,

1997). Conversely, the use of fluoroscopy or other intraoperative modality can be used to

verify a CT-based image guidance system and avoid complications resulting from

registration errors, modeling errors, shifting of the reference frame, or untracked

intraoperative shifting of anatomy (Dickman, 2000). As far as how guided-surgery

effects the overall duration of the surgery, there are arguments for no appreciable effect









and increased time consumption (Assaker et al., 2001; Weidner et al., 2000; Henderson et

al., 1996).

Systems

There are a number of image guidance systems that are used in spine surgery. Each

of them relies on the data from an imaging modality and a tracking system to relay

vertebral orientation to the surgeon. Medtronic offers three configurations that use IR

tracking, yet utilize different imaging modalities (Medtronic Surgical Navigation

Technologies, Louisville, CO). The first, the StealthStation, uses a traditional CT-

derived model to perform navigation. The second system uses 3D fluoro-CT images by

interfacing the StealthStation with a Siemens SIREMOBIL Iso-C3D, an isocentric,

automated fluoroscopy system (Siemens Medical Solutions USA, Inc., Iselin, NJ).

Fluoroscopic-CT images generally have excellent spatial accuracy, yet their poor image

contrast results in a 3D image reconstruction with poor tissue resolution and

differentiation (Foley et al., 2001). These qualities hinder the construction of suitable

virtual models for use in surgical guidance. The third pertinent system from Medtronic is

the FluoroNav virtual fluoroscopy system. It facilitates real-time navigation using C-

arm fluoroscopy. This system employs a two-plane display, yet the drawback is similar

to standard fluoroscopy in that 2D views do not give an accurate appreciation of 3D

anatomy. G.E. Healthcare Technologies (Waukesha, WI) offers surgical guidance

system configurations that use EM tracking, such as the OEC 9800 FluoroTrakTM, which

couples a high-resolution fluoroscopic imaging system with surgical navigation. A

software upgrade to the FluoroTrakTM Surgical Navigation allows the use of 3D

Fluoroscopic-CT images.









In addition to CT and Fluoro-CT, there are 3D image-based guidance systems that

use MR technology. An example is the PoleStarTM Intraoperative MR Image Guidance

System (Odin Medical Technologies, Inc., Newton, MA). It uses an infrared navigation

system similar to CT-based systems; however, this system allows intraoperative image

acquisition. The primary drawbacks to use of this system are very high cost and small

size which prohibits spinal operations on the trunk.

Considering the options for tracking intraoperative spinal motion during surgical

procedures, the cost of intraoperative MR and CT, the difficulty in resolving breathing

artifacts, the lack of dimensionality of fluoroscopy, and poor 3D image construction from

fluoro-CT, it is most clinically appealing to perform surgical guidance using a navigation

system that utilizes preoperatively acquired CT images. These images provide a suitable

platform from which to create an accurate anatomical model, plan the surgery, and

program the navigation. However, the primary obstacle is that multi-level spinal

procedures require the registration of a segmented model to be accurate.

Segmentation Introduction

Segmentation, as it is considered in medical image application, is the division of an

image into anatomically labeled sections. A clinician experienced in the modality can

easily recognize and outline the pertinent anatomy, however, a similar computational

recognition or identification scheme is challenging to develop. Algorithms that

automatically segment images have the potential to significantly reduce involvement of

the clinician, encouraging the beneficial application of segmented models. These models

can not only be used for visualization, surgical planning, and surgical navigation, but as a

medium to study structural information particular to the modality.









In order for a segmentation algorithm to be successful, many factors must be

considered. First, the time of the clinician is at a premium, so the algorithm must operate

quickly (in accordance with current computational technology) and with minimum user

input. It also must be flexible to adjustment; as such models will always be subject to the

fine-tuning of a trained eye. Also, the algorithm must be accurate and robust despite the

varied pathologies presented such as crushed vertebrae, fractures, scoliosis, and

involvement of tumors. Artifacts are problematic to any imaging modality, and a

segmentation algorithm must be tolerant to these instances. In the specific case of CT

scans, instrumentation can cause such artifacts. Finally, the performance of the algorithm

must be consistent and reliable, optimally eliminating the variability introduced through

human error. These parameters, combined with a thorough review of methodology, will

appropriately guide research of automatic segmentation algorithms.

Image segmentation is acknowledged as the most difficult and prohibitive step in

the modeling of anatomical data. Despite this impediment, segmentation of images has a

clear benefit to spinal surgery. The proper registration of an extended-level surgical plan

to a patient diminishes the inaccuracies introduced during intraoperative movement. It is

the focus of this research to develop and test methods to automate the process of vertebral

segmentation of 3D spine models constructed from CT image slices. This will encourage

image-guided spine surgery involving multiple vertebral levels, ultimately improving

patient outcomes.














CHAPTER 2
SEGMENTATION BACKGROUND

Segmentation is the division of a whole image into a subset of connected pixel

regions that have some common property. Segmentation in medical image processing,

however, includes the implied step of anatomical labeling. That is the association of the

image information to the appropriate anatomy for analysis. Computer algorithms that

accurately segment medical images can be challenging to develop. However, there are

many potential benefits such as enhanced visualization and ability to perform complex

surgical planning and navigation procedures.

There are several approaches to the problem of medical image segmentation. Each

of them relies on different information and can give somewhat different results,

depending on the application. The first, voxel classification, uses globally defined

characteristics to determine segmentation. For example, the use of an intensity threshold

to identify bony anatomy qualifies as voxel classification. The next approach to

segmentation involves the creation of a boundary concept that is used to mark the

division between regions. This concept relies on the identification of discontinuities

between regions. The contrast to this edge-based segmentation is region-based

segmentation. This approach uses regional characteristics such as common intensity

patterns to identify clusters. The final approach to image segmentation is the use of

deformable atlases. The atlas is a general model that is associated to a particular

anatomic structure. To perform segmentation, the general atlas is deformed to match an














3404



0 64 128 192 255
68043








: 1 64 128 192 255

Figure 2-1. Intensity histograms and histogram equalization. An intensity histogram
(upper right) is calculated from the sample image (upper left). The horizontal
axis shows grayscale value and the vertical axis shows frequency. Histogram
equalization improves the contrast over the most densely populated intensity
ranges. A histogram equalization (lower right) is performed on the sample
image (lower left).

individual image. Often, these techniques for medical image segmentation can be

integrated in an algorithm to produce a more favorable result.

Image Processing

Certain methods can be used to extract information from an image for use in a

segmentation algorithm or to enhance an image to better suit a segmentation algorithm.

Examples include histogram equalization, boundary detection, and filtering.

An intensity histogram of an image is a frequency distribution of pixel intensities.

The resultant graph indicates which intensity ranges have the most pixels (Figure 2-1).

One use of the information provided in a histogram is to allow contrast normalization.

This normalization process, called histogram equalization (Figure 2-1), maps pixels to









new intensity values to approximate a flat histogram. This process creates no new

intensity values.

Boundary detection is a feature detection method used to identify boundary pixels

in an image. These boundary pixels correspond to the region between different tissue

types and are usually defined by a high intensity gradient (Roberts, 1965; Sobel, 1970;

Prewitt, 1970). The gradient of a three-dimensional image with image intensity I is

defined as:


VI(x, y,z) = I(x, y, z) + J I(x, y, z)+k I(x, y, z)
ax Dy Dz

Various tracking systems can use the gradient information to make an edge trace.

However, ambiguous or discontinuous edge data can produce errors. A drawback to the

use of gradient information is that the image often must be smoothed because gradient

calculations are susceptible to noise.

Filters are used to enhance the qualities of an image in accordance with some

desired characteristic. In one instance, multi-dimensional adaptive filters are used to

resample the image data to reduce partial volume effects and noise. They also handle the

low off-plane resolution of CT images (Westin et al., 2000). Another filter of note is

specifically designed to allow for a more robust image segmentation for use in guided

surgery. It proposes to enhance separation of joint spaces in a CT scan, while allowing

the retention of important edge information (Westin et al., 1998).

Mathematical Morphology

Mathematical morphology is a geometrical approach to signal processing

(Matheron, 1975; Serra, 1982). It performs many image processing tasks using object

quantification and easily deals with attributes such as shape and size, connectivity, and









contrast. Also, edge information is preserved during boundary manipulation. Common

applications include noise reduction, texture analysis, and shape changing such as

thickening, pruning, or skeletonization. These characteristics make mathematical

morphology well suited to the task of image segmentation.

Mathematical morphology uses set theory as a foundation for many of its functions.

In accordance, its primary operations use binary structures which are defined with an

object and complementary background:

Object: A = {a property(a) == TRUE}

Background: Ac = {ala a A

A common technique for the creation of such a structure is the threshold mask. The

threshold mask TM is a binary overlay that indicates which voxels in a three-dimensional

image I are included within a prescribed intensity range [to, ti ] as shown:

TM =i, k,1) (i, j, k, g) e I, ge [t,,, ]}

Simple operations such as reflection and translation form the basis of more complex set

functions, and can be applied to this binary image structure. Reflection of set A is

indicated by A:

A = {w = -a, Va e A}

In other words, A is the set of elements, w, such that w is formed by multiplying each of

the coordinates of the elements, a, of set A by -1. Note that the elements or voxels in an

image are considered vectors, so in a three-dimensional image in Z3 space

a = (a,, a2, a ). Another basic set operation is translation:

(A) = (cc = a + x, Va e A










111

1 1 1

111

Figure 2-2. Morphological structuring elements and neighborhood configurations. A flat
3x3 structuring element (FSE) (left) is equivalent to a pixel with its eight
nearest neighbors (8NN configuration) (middle). On the right is a structuring
element composed of a pixel and its four nearest neighbors (4NN).

The equation says that the shifting of set A by vector x is the set of pixels c such that c is

equal to a+x for all pixels that are a member of A.

Binary Operations

Two operations, dilation and erosion, are part of a core of image processing

algorithms used in mathematical morphology. They produce a result by passing a

structuring element over the image. This structuring element is analogous to the

convolution kernel used in linear filter theory and it must have the same dimension as the

image. Technically, in this application, the structuring element can be considered an

image. For example, in an algorithm for a two-dimensional image, a flat structuring

element (FSE) is commonly used. It has each element in the structuring array set to

"true" or "1" (Figure 2-2). Other types of structuring elements can be defined according

to their neighborhood configuration. For example, equivalent to the 3x3 FSE is a

structure consisting of a pixel and its adjacent horizontal and diagonal pixels. These

adjacent pixels are known as the eight nearest neighbors (8NN). Another common

structure is a central pixel with its adjacent orthogonal pixels or four nearest neighbors

(4NN). These neighborhood configurations are shown in Figure 2-2.

Nearest neighbors are defined by their connectivity patterns; the arrangements that

determine if adjoining pixels are part of the same object. In the 4NN structure example























Figure 2-3. Binary morphological dilation. A binary image (left) dilated by a structure
(middle) results in an expanded object (right)

above, all of the four nearest neighbors are 4-connected to the central pixel. On the other

hand, pixels are 8-connected if they are connected in the orthogonal or diagonal direction,

as in the 8NN structure. Connectivity arrangements for three-dimensional structures

include:

1. 6-connected: voxels are connected by their faces
2. 18-connected: voxels are connected by their faces or edges
3. 26-connected: voxels are connected by their faces, edges, or corners

One of the simplest morphological functions to implement is dilation. It serves to expand

the binary image structure and is analogous to convolution (Figure 2-3). In typical

notation, A is the image set and B is the structuring element. The set theory formula for

dilation is given by:

A B= {z(B),nAc-A}

That is, all coordinates z where the translation of the reflected structure B by z intersects

with the binary image A. The implementation of dilation on a computer, also known as

Minkowski addition, is in a slightly different form. It is the union of the sets where the









binary image is translated according to each of the points in the dilating structure:

A)B=UAb
bEB

Erosion, another simple morphological image processing operation has the effect of

shrinking a binary object. The set operation for erosion as well as Minkowski subtraction

is given by:


A 0 B={x:(B)x A}= A

The set equation simply states that the erosion of A by B is the set of points x such that B

translated by x is contained in A. The Minkowski variant computes the erosion by taking

the intersection of all of the sets from the result of A translated by each element of B. The

two simple morphological operations dilation and erosion can be combined in series to

form the compound morphological operations open and close. The open operation serves

several functions such as smoothing of object contours, breaking of narrow isthmuses,

eliminating thin protrusions, and acts as a filter to remove background noise:

AoB= (AOB) B

The close operation, on the other hand, fills gaps in contours, fuses breaks, eliminates

small holes, and acts as a filter to remove foreground noise:

A.B=(AOB) 0 B

Opening and closing have the advantage of being idempotent, which means that repeated

applications will not further change the signal. Examples of opening and closing are

shown in Figure 2-4.

There are a handful of other binary morphological operations that are useful for

segmentation. One is boundary extraction. A boundary is extracted from a binary object




















S 4 I i


Figure 2-4. Compound morphological operations. Comparison of open (middle) and
close (right) morphological operations performed on a binary image (left)
using a 3x3 Flat Structuring Element.

by subtracting the result of an erosion operation performed on that object. The

connectivity of this boundary is determined by the structure used for the erosion. Region

filling is another useful operation. It utilizes the background or complement of a binary

image as the area to be filled. A point Xo is selected in the background and repeatedly

dilated until it achieves the desired fill result:

Xk = (Xk- B)rAC

This technique is especially successful for filling internal gaps. Finally, distance

transforms can be integrated into mathematical morphology (Cuisenaire, 1999). Distance

transforms create distance maps from binary image objects (Figure 2-5). Distance maps

are images in which the intensity of a pixel p is an indication of proximity or nearest

distance to the object 0:

D(p) = min{distM (p, q), q e O}

Distance maps have a smooth surface and an even gradient making them desirable as

shape representations. In addition, distance maps have an interesting property. The

result of an object dilated by a spherical structure can be expressed as the threshold of a

distance map of that same object. B is a spherical structure with radius d, created by the


I I I I I I


I I I I I I










0 0 0 0 0 0 2.2 2 2 2.8 3.6

0 0 0 0 0 02.2 3.2

0 1 1 0 0 0 1 2 3

0 1 1 0 0 0 1 2 3

0 0 1 0 0 0 14 2 3

0 0 0 0 0 0 1 I 4 2.2 3.2

Figure 2-5. Morphological distance transforms. A Euclidean distance transform is used
to form a distance map (right) from a binary image object (left). The pixel
intensity, as indicated by the numerical value, relays the distance of the pixel
to the object.

selection of all points that are less than or equal to a certain distance d away from point

(0,0,0):

B = {b dist (b, (0,0,0)) d}

And an object Xdilated by the structure B is equivalent to the threshold of the distance

transform DT(x) of the object at distance d:

X G B = {x DT(x) d}

Grayscale Operations

Binary morphological image processing methods can be extended to grayscale

(Bangham & Marshall, 1998). Dilation and erosion become useful filtering functions.

Dilation extends an object by using the maximum filter to remove low-valued regions.

Accordingly, erosion contracts an object by using the minimum filter to remove high

valued regions. These operations both have the effect of smoothing an image. Grayscale

morphological dilation assigns to each pixel the maximum of the sum of the local region

and the structuring element. In the special case that this structuring element contains all












zeroes, this operation is equivalent to the maximum filter:

A B= max (Ax+, +,z+k)
(i,j,k)GB I

And the dual operation for erosion using the minimum operator:

AOB = min (Ax+z,+,z+k
(I,jk)B Y ,

These operations, as with their binary analogues, can also be combined to form the

compound operations open and close. Opening acts as a high intensity point filter and

closing acts as a low intensity point filter. Another useful grayscale operator is the

morphological gradient. Given an image I, the morphological gradient is given by the

difference between the respective dilation and erosion:

g(I) = (I 9 B) (IB)

This intensity gradient is very useful for boundary extraction. The advantage of using

these morphological operations is that they provide useful image-processing features that

are easy to implement, and integrate into a segmentation routine.

Voxel-based Segmentation Methods

The simplest of the segmentation methods utilizes global image information to

assign the memberships of voxels into particular anatomical regions. Boundaries of

regions are then implicitly determined from a complete label map. Given a set of

anatomical structures contained in an image {co0... Ok}, a label map L(x) overlays the

image I(x) and L(x)= coi where oi represents the anatomical structure at I(x). The

assignment of voxels takes into account such factors as intensity value, neighboring pixel

classification, and relative distance of neighboring pixels. However, if the chosen factors

result in an overlap of neighboring anatomical structures, the global representation makes









additional helpful information such as shape and geometric relationships difficult to

incorporate. On the other hand, voxel methods are quickest in speed and do not require a

cumbersome training model to achieve a result.

Thresholding

Thresholding is a simple, extensively used image processing technique that isolates

a region of an image based solely on intensity criteria. It is a computationally

inexpensive and fast low-level technique. In addition, the result may be used as an input

to a higher-level segmentation model. It is defined as:

G(i,j,k)= 1 for I(i,j,k) > T


0 for I(i,j,k)
G is the resultant three-dimensional binary image of a grayscale image I thresholded at

intensity value T (Figure 2-6). Semithresholding is a similar technique which masks out

the image background leaving gray level information present in the objects:

G(i, j, k)= I(i,j,k) for I(i, k) > T

0 for I(i, j, k)
An upper bound may also be incorporated into these thresholding forms. It serves to

isolate a specific intensity band. Overall, this class of voxel methods is typically applied

in cases where particular anatomy or tissue can be identified within a certain intensity

range. The most apparent example is the segmentation of bone, where all included

voxels are for the most part at a higher intensity than adjoining tissues. The selection of a

threshold value can be done through manual inspection or probability estimation. To

perform the estimation, a tissue model is created that predicts which voxels belong to

which structures based on probability distributions of intensities. The probability of





















6808


3404



0 64 128 192 255

Figure 2-6. Threshold modeling. A binary image (upper right) is created by thresholding
the sample image (upper left). The red line on the histogram of the sample
image (bottom) indicates the threshold intensity value. Each of the pixels
with an intensity value higher than the threshold value is highlighted in the
binary image.

intensity value x, estimated at each tissue class aoi is based on a set of training

data P(xco,). Histograms are commonly used to collect training data for this technique.

Care must be taken in application of these voxel methods if intensity ranges are not

disjoint. For example, in spinal CT images, it is not uncommon to have certain intensity

values coinciding with both bone and cartilage or ligament. Also, thresholding or tissue

class estimation is susceptible to imaging noise and artifacts that cause intensity overlap

and unclear boundaries between tissues. A popular method for probability estimation is

called Expected Maximization (Dempster et al., 1977). It is especially effective at

handling incomplete data.









Tissue class estimators can actually use criteria other than intensity to perform

segmentation of tissues. One example, which makes a classification according to regular

neighborhood intensity patterns, uses a technique called statistical clustering (Leahy et

al., 1991). Texture patterns are typically measured though a co-occurrence of distance

and intensity.

Morphological Segmentation

Methods in mathematical morphology can be used to segment an image

(Dougherty, 1993). And although morphology is a language for shape representation and

manipulation, its basic segmentation methods can be considered voxel-based. The first

concept in morphological segmentation is seeded region growing. A seed is a point voxel

or a group of voxels. The seed is expanded by checking to see if neighboring boundary

voxels are within specified criteria. A common criterion is if the absolute value of the

intensity difference of a seed and its neighboring voxel is beneath a threshold. A very

simple version of region growing is found in the watershed transform (Vincent & Soille,

1991; Beucher & Meyer, 1992).

The watershed transform is a segmentation method for grayscale images. It

interprets the topology of an image and assigns watershed lines to boundaries between

catchment basins. Catchment basins are the areas of local minima on the topographical

map and are analogous to the depressions that would collect drops of water. The

watershed lines are the crests between these basins. The common input to the watershed

transform is a morphological gradient of an image so that watershed lines would

correspond to the areas of strong edge evidence and divide the original image into

homogenous regions. In application, however, simply taking local minima can result in

oversegmentation, especially if there is a large noise contribution from false minima.






















Figure 2-7. Binary image segmentation. The distance transform can be used to create a
topographical image (center) from a binary image (left). The binary image is
a silhouette of three overlapping shapes. Utilizing the topographical image,
the watershed transform can then be called to approximate a segmentation
(right).

The problem of oversegmentation can be handled in a couple of ways. The first is to

apply a merging scheme, which combines adjacent areas according to some guideline.

This guideline is usually based on statistical gray level properties. The second way to

deal with an oversegmented image is to use marker selection. First, a new analogy

should be introduced, and that is the gradual flooding of the image topography using a

rising water table. Given this, marker selection is the decision of which local minima or

seed regions will flood (Beucher & Meyer, 1992; Meyer & Beucher, 1990). The result is

regions that are selected will flood the regions that are not, eliminating oversegmentation.

Caution must be used in attempting to use smoothing operators or other such filters to

reduce the false minima from noise contributions as they have the potential to remove

areas of strong edge information. In modem application, the watershed transform has

been used to segment out maxillofacial bone in CT (Bohm et al., 1999). The cited

method also uses a tissue classification scheme to label the segmented regions. The

watershed transform has applications on binary images (Figure 2-7). It can use a









topographical image created by the distance transform to separate overlapping objects

(Cuisenaire, 1999).

The last morphological medical image segmentation method of note uses Recursive

Erosion (RE) and Geodesic Influence (GI) (Kaneko et al., 2000). The recursive erosion

is used generate candidate seeds while the geodesic reconstruction recovers the separated

organs.

Edge-based Segmentation

Some segmentation problems can be solved using boundary localization. This

concept involves the creation of an edge or surface model that is designed to converge on

an object boundary. An accurate convergence will then describe the border of the

segmented object. Given a three-dimensional image, closed surfaces are defined

{Si ... Sk} with all points inside surface Si corresponding to an anatomical structure omi.

Also, all points that represent wo, are contained by S,. The information for the model is

gathered by relating the edge representation to its associated image information. This

includes image gradient, texture discontinuities, or any other useful measure that can be

geometrically associated with the boundary. One indirect scheme for the segmentation of

3D objects involves the unification of multiple segmentations in two dimensions. An

example is the segmentation of bone in CT images. Once the proper segmentation of one

image slice is accomplished, it can be used to direct the result of adjacent slices. The

combination of all of these individual slices results in a segmented three-dimensional

image.

Active contours are dynamic deformable models used for edge-based segmentation

(Blake & Isard, 1998; Caselles et al., 1993; Malladi et al., 1995; Tek & Kimia, 1995).









One such model is called the snake (Kass et al., 1987). It makes use of an energy

function to guide the evolution of the boundary. The snake itself is a spline function

parameterized by a set of node points. The cubic polynomial is a common choice for the

spline function:

x(u) = axu3 + bxu2 + cxu + d

y(u) = ayu3 + bu2 + cu + dy

The energy equation is then used to direct the movement of the nodes. The terms in the

equation are a balance of forces that pull the spline toward the desired edge features. A

proper segmentation is achieved through the minimization of this energy equation:

Etotal = [J mt ernal + ;mage Econstra mt


This equation includes a term for internal energy, image energy and constraint energy.

Internal energy is solely dependent on the shape of the spline. It includes parameters for

stretching and flexing, which are optimized according to the geometric knowledge of the

object to be segmented. The image energy is based on the image values along the path of

the spline and can include attraction to strong gradients or regions of light or dark

intensity. Constraint energy is intended to capture the higher level knowledge about the

image and features. Example constraint energy terms are manually defined attraction or

repulsion fields such as volcanoes and springs. A specific snake model designed for

image segmentation was introduced by Kass et al. (1987):

E(C) = f C'(q) 2dq- J VJI(C(q))dq

It includes internal and external energy terms for the curve C. Internal energy is a

regulating force that keeps the curve smooth by penalizing high curvature. This also









makes the curve robust against noise. External energy is image energy and is designed to

act as an attraction force to high gradients. The coefficients of these terms are

empirically adjusted according to the specific application. One consideration in the

optimization of snakes is the dynamic adjustment of nodes. If the number of nodes

remains consistent, then snake expansion will result in decreased resolution and snake

contraction will result in unnecessary computational expense. Algorithms that effectively

adjust number and spacing of nodes will maximize accuracy and speed.

Snakes have the advantage of being conveniently autonomous in their search for a

minimal energy state. Also, since the integral operator is an inherent noise filter, they are

insensitive to noise and other image ambiguities. These ambiguities include spatial

aliasing and sampling artifacts that can cause boundaries to be indistinct and

disconnected (Mclnemery & Terzopoulos, 1996). In contrast, snakes often overlook

minute features in the process of minimizing the energy over the entire path of their

contours.

Initiation of the snake is critical to the success of the algorithm and accordingly

there are a few points to consider. First, the snake must be initialized close to the

expected boundary for good performance since they can get stuck in lower minima states.

This is a concern since contours are generally difficult to initialize around the region of

interest. For best results, the snake should be initiated depending on the uniformity of the

intensity distribution, either inside or outside the boundary. Caution must be observed if

the curve is initialized on the inside of the expected boundary in a place where the image

information has little influence as the curvature penalty will cause the spline to shrink to

a point. The addition of an outward pointing force addresses the shrinking problem














Figure 2-8. Singularity in contour evolution. A self-intersecting curve creates a
singularity which can be described as in the process of either splitting or
merging. Grass burning is one method used to solve this ambiguity (right). It
determines the state by marking the path (texture) traveled by the curve (blue).
In this case, the contour is in the process of splitting.

inherent in the internal energy term of the snake. This edge-based active contour is called

a balloon (Cohen, 1991). The balloon function includes stability aids such as the

normalizing of image force to restrict boundary movement and interpolation to avoid

discretization errors.

A particular concern with the evolution of contours occurs when image topography

causes the contour to self-intersect. The intersection point can be in one of two states,

splitting or merging. The nodes can be reparameterized to adjust, as in establishing two

separate contours in the case of splitting (Figure 2-8). However, the state must be

established. Several solutions exist to tackle this problem. One is to use a grass-burning

assumption (Sethian, 1996a). This creates an entropy condition where the evolving front

leaves a "burnt" or irreversibly marked path as it travels. This will determine what

direction the contour was moving. Another solution is to use T-snakes, which use a

triangulation of the embedded space to determine new node points (McInemey &

Terzopoulos, 1997). The Level Set Method is yet another method for tracking the

evolution of interfaces (Osher & Sethian, 1998).

Level set methods offer highly robust and accurate methods for tracking interfaces

moving under complex motions (Osher & Sethian, 1988). They handle topologically

breaking and merging naturally as in the creation of channels in a surface contour.


C::::






















Figure 2-9. Level set application for contour evolution. The contour (red) is embedded
as a zero level set of a higher dimensional surface (blue). This allows the
smooth adaptation of the contour to topographical changes.

Details about theory, implementation, and application of level set methods are in

Sethian's Level Set Methods (1996a). It relies on two central embeddings. One, the

contour is embedded as the zero level set of a higher dimensional function (Figure 2-9).

Two, the velocity of the contour is integrated into the higher level function. This is called

the extension velocity, and is curvature dependent. A companion technique to Level Set

Methods is Fast Marching Methods (Sethian, 1999, 1996b). They are numerically

efficient methods for evolving a front traveling in one direction. They can be used to

construct a distance map and provide appropriate extension velocities for level sets

(Adalsteinsson & Sethian, 1999). There are several associated active contour methods of

note that are used for image segmentation. The first uses a geodesic or minimal path

formulation for active contours in the level set approach and applies it to segmentation of

medical imagery data (Caselles et al., 1997). The second uses gradient flows to direct the

curve evolution (Kichenassamy et al., 1995). The last integrates other geometric

techniques for image segmentation (Malladi et al., 1995).









Much of the current research in active contour models deals with generalizing the

form of the contours and overcoming the convergence and stability problems encountered

during the energy minimization process (Davatzikos & Prince, 1992).

Region-based Segmentation

In applications such as medical imaging, adequate edge information such as strong

gradients may not always be present for an edge-based segmentation approach to be

consistently accurate. In these cases region-based models may be employed. In the

region-based segmentation approach, the boundaries of the deformable model are

determined by statistics inside and outside a region or cluster of voxels. These statistics

measure properties such as unique texture patterns, homogeneity of intensity, or some

other pixel-based statistic. The goal in evolving the region is that variation of these

properties is less inside a region then between regions. Similar to edge-based methods,

region based methods are evolved through the minimization of an energy term. This

global energy term, however, is defined for the entire area of the region rather than only

the boundary. There are a variety of region-based segmentation methods that include

Bayesian segmentation (Geman & Geman, 1984), piecewise constant energy (Mumford

& Shah, 1989), region competition with balloons (Zhu & Yuille, 1996), and an energy-

based watershed (Bleau & Leon, 2000; Nguyen et al., 2003). One technique uses the

level set formulation to evolve the boundary of a region based on texture statistics (Yezzi

et al., 1999).

Region-based segmentation algorithms have several advantages. First, they have a

greater capture range and are not as dependent on initialization as edge-based methods.

Also, they are not reliant on high frequency information and are not as susceptible to









noise (Chan & Vese, 1999). Care must be taken in application, though, as the

determination of regional parameters can be computationally expensive.

As a final note, efforts have been made to integrate region-based and edge-based

modeling techniques (Chakraborty et al., 1996; Paragios & Deriche, 1999). The addition

of a boundary regulation term is used to control smoothness in a region-based

segmentation model. Also, geometric information, such as a shape prior, is much more

easily incorporated into a boundary formulation. However, the tradeoff of boundary

regularization comes at a cost of strength of region description.

Shape-based Segmentation

The final approach to medical image segmentation is to incorporate shape

information in a deformable modeling algorithm. This is accomplished through the

construction of an atlas. An atlas is a collection of prior information used to direct the

deformation of the algorithm. It can be a set of points, finite elements, flow fields, or

other similar parameter, but is most commonly a shape descriptor. Once defined, the

atlas or standard template molds itself to the target image, imprinting an inherent label

map. The data associated with the atlas can also be used to direct the mentioned edge or

region-based segmentation strategies.

Atlas Warping

The classic warp consists of a representative image scan that has an assigned label

map L*(x). Given a new image, some transformation Tmust be computed that deforms

the representative scan I*(T(x)) to correspond with I(x). The same transformation

applied to the label map L*(T(x)) will then describe the label map for the new image

L(x). There are several examples of this template-driven segmentation (Pentland &

Sclaroff, 1991; Sclaroff& Pentland, 1995; Wang & Staib, 1998b; Warfield et al., 1998).









One specific method computes the deformation field in a way that allows a course to fine

solution (Christensen et al., 1996; Christensen, 1999). Another method begins with a

rigid transformation and iteratively progress in plasticity to find a solution (Miller et al.,

1997). A technique especially useful in the atlas-based segmentation of composite boney

structures incorporates the rigidity of the tissue class in the deformation model (Little et

al., 1997). More examples of atlas warping include works by Cootes and Taylor (1992a),

Jones and Poggio (1998), and Pichumani (1997).

Modeling

There are several factors to consider in the development of a shape-based model.

First, for automatic interpretation, it is essential to have a model that not only describes

the size, shape, location and orientation of the target object but that also permits expected

variations in these characteristics. In order to properly account for this object variability

seen in application, a statistical analysis must be performed (Dryden & Mardia, 1998;

Neumann & Lorenz, 1998). Casting the fitting process of deformable modeling into a

probabilistic framework allows incorporation of prior statistics as well as an inherent

measure of uncertainty (Mclnemey & Terzopoulos, 1996). These statistical models have

the advantage of being flexible to ambiguity and noise. Since the development of a

model and its associated training data set can be data intensive and computationally

laborious, techniques are used to minimize the number of statistically significant

parameters. One technique in particular is Principle Component Analysis (PCA). This

process identifies the primary modes of variation and reduces the dimensionality of the

data set, optimizing the framework (Golland et al., 2000; Lorenz & Krahnstover, 1999).

Multiple strategies for shape description have been developed. Feature detection is

a common approach. A feature is a geometric property of the object to be segmented that









can be easily identified by an algorithm. The algorithm performs a segmentation by

matching the set of features that is extracted from the unlabeled image to the training set.

Depending on the application, the association of these features should be defined with

certain invariances such as translation, rotation, or homogeneity. One method of feature

extraction uses medial axis or skeletonization. The geometric features produces are

scale-invariant, which allow detection at different resolutions (Dougherty, 1993). A

drawback to the use of skeletons is that original size and edge information are lost. An

improvement on skeletonization uses fixed topology skeletons. They have the advantage

of being robust to the noise and quantization errors that traditional skeletons are

susceptible to (Golland et al., 1999). Mathematical morphology can be used to build a

shape classification strategy as well (Dougherty, 1993). The binary object of interest is

probed with an array of simple shape primitives, from which a statistically appropriate

feature set can be collected. In another method, distance maps are used as a shape

descriptor (Golland et al., 2000). Distance maps provide smooth and wide minima for a

matching algorithm.

Methods

There are a variety of shape-based segmentation algorithms. One particular group

uses Active Shape Models (ASM) (Cootes & Taylor, 1992a; Cootes et al., 1992b; Cootes

et al., 1999b). The ASM is a parametric deformable model that is represented in the

image as an n-point polygon. The algorithm then deforms the points of the polygon to

match the landmark points in the target image object. The points are evolved according

to a point density model (PDM), which is constructed using PCA. The PDM provides a

statistical distribution of global shape variation. Once the model is deformed, an

accompanying label atlas identifies the segmented object. The ASM is significant









because it is fast and accurate and can describe complex shapes with minimal parameters.

In application, models have been developed to account for both Gaussian and non-

Gaussian data (Cootes et al., 1995; Cootes & Taylor, 1999a). In one instance, ASMs are

incorporated into a Bayesian formulation for boundary location (Wang & Staib, 1999a).

ASMs have also been extended to accommodate grayscale information as Active

Appearance Models (Cootes & Taylor, 1998).

Shape-based segmentation approaches have been integrated into edge and region-

based algorithms to retain the benefits of those approaches. For example, contours are

well suited to handle shape information, and such information can easily be incorporated

as priors. One method uses statistical priors on the Fourier coefficients of the contour to

represent shape (Staib & Duncan, 1992). A related method combines snakes with a

shape-based Fourier parameterization method (Szekely et al., 1996). Chen et al. (2001),

uses shape priors with a level-set evolved geometric active contour model. Leventon

(2000) tests a method that incorporates distance-intensity profiles and shape contour with

a probability-based segmentation approach. It uses a level set implementation of

geodesic active contours for shape evolution. Also, PCA is implemented to reduce the

dimensionality of the training set. A couple of region-based shape methods are also of

note. Cremers et al. (2002) incorporate statistical shape knowledge into the Mumford-

Shah functional to perform segmentation. In application of cervical spine segmentation,

Pichumani (1997) uses a finite-element shape model integrated into a region-based

approach.














CHAPTER 3
SEGMENTATION

Currently, to create a segmented model appropriate for guided surgery, the clinician

follows a few basic steps. First, he reviews the volumetric CT scan and selects an

intensity value that defines the edge of the bone. Isocontours are displayed on the

terminal to assist in this task. Typically, voxels of bone have much higher intensity

values than adjacent soft tissues, and the boundary value can be easily identified. The

selected intensity is used to define a threshold mask. The mask is a binary overlay which

identifies all voxels that are associated with bone (Figure 3-1). The highlighted voxels

describe an object which can usually be seen with accompanied 3D visualization

software.

There are a few considerations in the creation of a threshold model. First, it is

crucial that the edge of the bone is accurately defined, as many of the registration

techniques use surface points to do the matching. A model created with a threshold value

that is low will include non-bony soft tissue such as cartilage and ligaments. A model

created with a threshold value that is high will diminish the volume of the bone, shrinking

the edge and withering the interior. The process of using a threshold to describe an area

of the image can also be utilized to isolate an object from high intensity artifacts or noise.

Instead of a just a single lower intensity threshold, an upper intensity bound is used as

well.

The model created by thresholding rarely results in a structure that is segmented or

appropriately sectioned to match the underlying anatomy. Therefore, the model must be





















Figure 3-1. Building a threshold model. On the left is a sagittal slice of a cervical CT
image. On the right is an isocontour (red) scribed at the boundary of the bone,
with its implied threshold mask area highlighted in blue. It can be seen that
several of the bones are inappropriately grouped together.

manually segmented. This is a tedious and time-consuming process whereby the

clinician flips through each of the series of CT image slices and uses simple pixel

highlighting tools to identify each bone. The result is a segmented model that can be

used for guided surgery (Figure 3-2). However, even though this method produces an

acceptable result, the time commitment involved in this manual process restricts the

routine use of guided-surgery for multi-level applications. Therefore, the focus of this

research is to develop methods to automate the segmenting process and facilitate these

procedures.

Problem Evaluation

Several guidelines were established to direct this segmentation research. First, the

time spent by the clinician to prepare a segmented model must be minimized. This

includes increasing simplicity and minimizing user input. In contrast, however, there

must be enough prior knowledge provided by the user to direct the algorithm to a correct

solution. Another contributor to the involved time of the clinician is the speed or

computational efficiency of the algorithm, but this optimization is secondary at this stage

of the research. Second, there must be an accurate approximation of bone edge for the





















Figure 3-2. Manual segmentation and display. The image on the left is a 3-D view of a
contiguous model formed using the thresholding method. The image on the
right is the same model that has been manually segmented. The goal of this
research is to reduce the involved time to produce segmentation.

registration methods to be successful. Third, the segmentation process must be robust to

the assorted image variations. These include pathologies and artifacts, such as those

induced with instrumentation.

The initial step in development of the segmentation algorithm is to assess a random

sample of images used for image guided surgery. Currently, most of the image-guided

surgical procedures are implemented on the cervical spine. This is because there is a very

small tolerance available for tool placement to avoid critical vascular and nerve

structures. A series of the high-resolution cervical CT scans used for these surgeries

were evaluated. Upon analysis, it was apparent that the general shape and intensity

distribution of the vertebrae were similar. Some patients had an overall higher bone

density, with clear distinction of boundaries. Others had bone density that was close to

that of soft tissue. In some of the most severe cases, the individual bones were difficult

to recognize. Some of the pathologies observed include fractures and fusion, each which

showed unique discontenuities and intensity patterns. Instrumentation, such as rod and

screw assemblies or wires used to stabilize bones, produce unique artifact patterns in the









lAnfffiwO tbrclrR ofe


t ; iSatelf/r Anterior juverE e,.;

SrImr tser pcess'
8u 'Jr Sulcu F n ostnora s ubersle o8


Figure 3-3. Surface characteristics of vertebrae. The typical cervical vertebra has very
complex topology, with areas of high curvature and channels. These
properties make it a troublesome candidate for edge-based or shape-based
algorithms. (Left) Superior view. (Right) Lateral view. (Gray's Anatomy).

CT scan and disrupt the normal intensity patterns. A careful review of image variations

was helpful in selecting the approach to algorithm design for this research.

Method Selection

Several avenues exist for the development of a segmentation algorithm. One

common approach is to use boundary localization techniques. These involve the creation

of a boundary concept that will deform until it reaches a minimum energy state at an

appropriate edge. They have the capacity to handle small gaps in information, but also

have the potential to arrive at false minima. These are erroneous minimal energy states

that locate an incorrect boundary. A primary difficulty in the use of boundary

localization techniques is that the typical cervical vertebra has very complex topological

characteristics including areas of high curvature and channels. This makes it a difficult

surface to evolve (Figure 3-3). In addition, the peculiar shape would exacerbate the

initialization problems that have the potential to plague these methods.

Another avenue for image segmentation is to use region-based methods. These

methods have the advantage of not being as dependent on initialization as edge-based

methods and are less affected by the gaps in information or high frequency errors that

may cause boundary methods to reach a false minima. A drawback, however, is that









region-based parameters are often computationally expensive. In addition, region-based

methods may not be best suited to handle the complex boundaries and topology presented

with a typical vertebra. One advantage to a region-based approach, however, is that

shape information can be more easily incorporated. An atlas, if properly developed,

could capture variance across a population and be used not only to direct segmentation,

but also to identify abnormality or pathology. The incorporation of shape information,

however, can be costly. First, reliable features must be identified. And then, a

comprehensive database must be put together and maintained. Given the number of

pathologic conditions to account for, this could be a very time consuming process.

The third avenue to develop a segmentation algorithm is to use voxel-based

techniques, which are independent of any boundary or region components. This

significantly reduces complexity, yet has the drawback of not being able to easily account

for geometric information. Thresholding is a proven, if inefficient, voxel-based modeling

technique, and makes a good approximation of the bone surface. Therefore, it provides a

convenient platform from which to launch other voxel-based modeling techniques.

Given that this is the simplest approach, it shall be investigated first. To facilitate this

work, morphological operators will be used, as they are computationally efficient, work

in 2D and 3D and preserve edge information.

Research Tools

A programming language and accompanying toolset were selected to facilitate the

segmentation research. IDL (Research Systems Inc., Boulder, CO) was chosen because

of its strong image processing toolset, assortment of visualization methods, a GUI

interface, and ability to process large arrays. First, a GUI was created to visualize

volumetric CT images. Through it, any orthogonal image slice could be viewed and











-Inlxl


File Tools Work 3D Rendering Image Processing Information


F r OI "*,-,, F|l |.-r-l

&Fr,, |tj I TE T
















anatomical association.
manipulated. Plus, a function for overlaying intensity contours to aid in He selection of
















was constructed, multiple groups of[2D and 3D image processing functions were added to
























environment that mimicked the manual segmentation process. Before this could be
-r l AI1j',' Ij ii'.*




4j 1 r'i4
I ti '. i HI,
5u




,I /,121 2II
S1, 1 "7 5, li i



r C r'G Fri rI A o.,
-_J



.'1l.d l[' .l-"" r H, -i i l L, vl['


Figure 3-4. Toolset GUI. Multiple functions are easily arranged to facilitate
segmentation. Shown is an axial CT slice with an overlaid indexed array
(yellow highlight), indicating the segmentation boundary and corresponding
anatomical association.

manipulated. Plus, a function for overlaying intensity contours to aid in the selection of

threshold values was included. Since a 2D view does not give an accurate account of 3D

connectivity, several 3D visualization schemes were also setup. Once a viewing system

was constructed, multiple groups of 2D and 3D image processing functions were added to

the menu structure, such as the morphological functions open, close, and gradient.

The next step in toolset development was the addition of functions to the program

environment that mimicked the manual segmentation process. Before this could be

implemented, however, a data representation of the segmentation model had to be









chosen. We decided on an indexed overlay. This is an integer data array equal in size to

the image array with an index number assigned to each voxel. This index identifies that

voxel as being associated to a specific bone. Once created, this array could be blended

with the original picture for visualization (Figure 3-4). This data representation for

segmentation is certainly inefficient, yet is easy to associate with any voxel-based

manipulation. Next, several options were added to the GUI to allow the creation of the

index array based on threshold techniques. Several tools were created to assign or

change index assignments. They include 2D single pixel and polygon and 3D sphere and

rectangular prism assignment options. In addition, continuity functions in 2D and 3D

were also added. They can be used to "paint" interior spaces or to fill holes. Throughout

the research, functionality of the interface was expanded accordingly.

Cursory Examination

In order to test the toolset interface, a few simple, preliminary examinations were

made. The primary failure of segmentation by thresholding is that the process does not

properly isolate each of the bones, leaving one contiguous clump of bone. These initial

experiments, then, attempt to use a few simple techniques to accomplish improved

separation. There are two approaches to test. One is to cut the threads that leave the

bones connected. The other is to identify the boundary and subtract it from the

contiguous model, leaving separated regions.

The morphological open operator is a simple function that can be used to effect a

break of thin connections on a binary object. Hypothetically, this could be used to

separate bones on a threshold model. Several tests were performed using the open

function on sample images. Symmetric structuring elements were applied in an order of

increasing width to see what size would effectively break the connections. The result is






















Figure 3-5. 3D intensity gradient. The morphological gradient is constructed from a CT
volume. On the left is an axial slice of a cervical vertebra. On the right is an
equivalent slice from the gradient map. Notice the interior areas of strong
edge-information.

that the size of structure required was thicker than some of the natural boney structures,

and the application of which significantly altered the surface of the vertebra.

The intensity gradient is a very commonly used indicator of edge strength. Several

3D gradient intensity maps were calculated from sample CT scans and analyzed. A

majority of the boundaries were highlighted in these trials; however, there were multiple

gaps or areas of low gradient. Therefore, without further modification, the gradient map

could not be directly used to make a boundary subtraction and foster segmentation.

However, an interesting property of the vertebrae is that there is a strong gradient on the

interior of the bone (Figure 3-5). One observation is that, given these circumstances, an

edge-based algorithm solely dependent on gradient information would have difficulty

localizing the correct boundary. The interior areas of strong gradient would tempt the

curve during evolution.

As a result of these preliminary experiments, a rectangular region of interest (ROI)

tool was setup to allow the application of image processing algorithms on a defined area.

It was theorized that this would allow unique image faults to be addressed without






52


affecting the whole image. For example, it could be used to address specific gaps in the

gradient map. After a few trials, it was apparent that ROI definition in accordance with a

specific region in 3D was a time-consuming process subject to burdensome iterations and

is something that should be avoided.














CHAPTER 4
REGION GROWING

Kernel Definition

The fundamental problem in using a threshold model created at the edge defining

value for bone is that it often connects adjacent anatomy as one contiguous model.

However, it has been observed that a regression of the isocontour to a higher threshold

value results in a disjoint model that can be uniquely associated with the underlying

anatomy. It is hypothesized that these regions can be used to create a successful

segmentation.

A method has been developed to create and label this kernel-based threshold model

using IDL. First, an isocontour is selected that allows the creation of disjoint regions.

This is an iterative process that involves using the visualization software to check for

contiguous regions. Once an intensity level is selected, a binary threshold is created in

the label array. The proper IDs are assigned through point selection and continuity

labeling tools (Figure 4-1). Since the goal is to minimize the intervention of the clinician,

not every unassigned region is labeled. It is assumed that the user will use knowledge of

the segmentation method to label the regions that will have the most profound effect on

the results.

Shape Correlation

It is observed in the creation of the disjoint threshold model that the shapes of the

kernel regions bear a resemblance to their associated boundaries. It is reasoned that if

















Figure 4-1. Manual kernel selection. A threshold model created at the bone-defining
intensity value of a sagittal slice of a cervical spine CT image (left). The
threshold value is raised, resulting in a model with multiple discontinuous
bony elements (middle). Each area is labeled to correspond to the underlying
anatomy (right).

enough boundary information is preserved in the shape of the kernels, then they can be

used for segmentation.

One simple voxel-based method to test this theory uses both the labeled region

kernel model and the contiguous boundary threshold model. They are compared, and

each of the highlighted voxels in the boundary model are assigned according to the

closest labeled kernel region. The assumption is that there is a high likelihood that an

unassigned voxel has the same label as the closest pre-labeled region. This can be

accomplished through a system of constrained morphological dilations. Specifically, the

kernels of adjacent anatomies are dilated equally until they touch. The voxel IDs

assigned in this process cannot be overwritten, so the boundary created is fixed. The

process of repeated dilation continues until all of the voxels in the boundary threshold

model are assigned. Expanding on this idea, distance maps may be used to approximate

this process. A distance map is an image whereby each of the voxels is assigned an

intensity value corresponding to the Euclidean distance to an object. This value, as it

turns out, is approximately equal to the number of unit dilations necessary to reach that

voxel from the object. Further, if a distance map is created for two sets of objects, the






































Figure 4-2. Segmentation using distance maps. Unique, discontinuous kernel regions are
created from a cervical CT scan (upper left). Then, two distance maps are
created. One represents the kernels associated with C2 (yellow) and the other
represents the kernels associated with C3 (Cyan). The difference of the
resulting maps (upper right) approximates a boundary between C2 and C3
(lower left). A segmentation results when a boundary threshold mask is
applied (lower right).

difference of the two maps scribes a zero line exactly equidistant between the object sets.

Figure 4-2 illustrates this process enacted upon two sets of adjacent kernels.

This segmentation method is a good way to quickly approximate a boundary.

However, there is one failing that made further pursuit inadvisable. The amount of shape

information contained in the labeled kernels rapidly deteriorates as the threshold level is

















Figure 4-3. High threshold values result in diminished kernels. The threshold contour
(red) is increased from the boundary defining value (left) to one that allows
proper label associations. The resulting high threshold and label (right) shows
kernels that retain very little boundary shape information.

Object 4-1 .High threshold movie (objectl-highthreshold.mpg, 27.6 MB)

increased (Figure 4-3). Therefore, circumstances that warrant a high threshold to achieve

isolated kernel regions produced the least accurate boundaries.

Hysteresis

The previous study showed that the shapes of the segmentation kernels are not

infallibly correlated to boundaries. More information is needed to direct the growth of

these kernels to arrive at the appropriate segmented boundary. In the kernel labeling

process, an intensity threshold is used to withdraw the isocontour to the interior high

intensity areas of the bone. This threshold information, if appropriately utilized, can

assist the kernels to revert back to the bone edge.

An algorithm was developed using a combination of morphologic dilation and

thresholding functions to progress labeled kernel regions to a fully segmented model.

The inputs to this algorithm are the CT image, the user-determined bone boundary

threshold intensity value, and the kernel region dataset. Each iteration of the region-

growing algorithm involves three basic actions (Figure 4-4):

1. The step to a lower threshold and resulting isocontour.
2. The constrained orthogonal dilation of the seed regions.
3. The restriction of the dilated voxels to the area outlined by the isocontour.







































Figure 4-4. Region growing algorithm iteration. Shown in the upper left are two
adjacent bone segments with corresponding seed regions. The first step
progresses the threshold contour to a lower value (upper right). Then the seed
regions are grown through morphological dilation (lower right). Finally, any
highlighted voxels outside the threshold contour are removed (lower left).

This basic cycle is initiated at the kernel-defining threshold value and repeated until all of

the volume outlined by the lower threshold mark is filled.

A couple parameters can be adjusted to discover the most effective sequence for

region growing. The first is the number of threshold divisions, or the number of times

that the isocontour value will step in order to reach the edge defining value. The height

of the step will depend on the number of threshold divisions and the difference between

the two boundary threshold values. The second definable parameter is the number of unit











Region Growing Input
r-----------------------------------------------------

1. Selection of intensity
threshold which defines the
glob al bone boundary

IP I

2. Selection of intensity --------------------------------------
threshold which allows the Tser hpllut
unique identification of
desired underlying anatomy


4. Image is equalized
over the upper and lower
threshold bounds to
regulate region growth


3. Identification of
seed areas using a 3D
continuity scheme



5. The labeled seed
regions are overlaid on
tl equalized image to
start the iteratve
procedure

L-----------------------------

Figure 4-5. Algorithm input processing sequence. The steps requiring user input are
outlined in green. The input image is histogram equalized over the boundary
threshold ranges to regulate growth while preserving the geometric boundary
conditions.

dilations per threshold step. This determines the rate at which the available threshold-

bound area will fill. Through repeated trials, a couple of basic trends were noted.

Generally, a higher number of threshold divisions produced a more accurate

segmentation. Secondly, the number of dilations required to properly fill a threshold step

varied widely. In order to regulate this growth, an analysis was performed. It was

determined that an intensity histogram could be used to determine the number of voxels



















Figure 4-6. Region growing. Once the input image is prepared, the algorithm instructs
the labeled kernel regions to grow incrementally until the bone-defining
threshold is filled.

Object 4-2 .Region growing movie (object2-success.mpg, 25.2 MB)

available to be filled per threshold step. Accordingly, the histogram can be equalized

and applied to the input image to even the growth. This step has been added to the

algorithm input processing sequence as shown in Figure 4-5. Once the preparation steps

have been accomplished, the iterative region-growing algorithm is activated and

continues until the total bone volume is assigned a label (Figure 4-6). The completed

label set then defines the segmentation for the original image.

Before the segmented dataset is used for a surgical application, it must be verified.

This involves overlaying the segmented dataset on the original CT image for visual

comparison. Manual post-processing steps may be necessary if the segmentation does

not succeed in the proper isolation of adjacent vertebral bodies. The same tools

mentioned for manual segmentation are used for this post-processing step.

Algorithm Enhancement

A few avenues were investigated to improve this region-growing method of image

segmentation. The first was an examination of growth characteristics, and how they

could be better controlled to suit the application. The second involves the modification

of image information to produce more auspicious growth patterns. The third tests the



















figure 4-/. upen and close operators are used to alter the evolution ot the growth
algorithm. On the left is the original image. In the center is an image with the
open operator applied. The white regions highlight the areas removed by the
operation. On the right is the image with a close operator applied. The white
regions are the areas filled by the operation.

inclusion of user-defined input information in addition to the initial kernel-laden

segmentation dataset.

During the region growing operations, the kernels tend to expand in finger-like

fashion until they merge with other anatomically self-similar sections. The problem

arises when one of these extends to a piece of unrelated anatomy, essentially allowing it

to falsely populate. Two shape filters were inserted into the algorithm to modify growth

behavior. One, the morphological open operator, was used to curb the extension of thin

tendrils to false anatomy (Figure 4-7). The other, the morphological close operator, was

used to facilitate growth into self-same anatomy. Orthogonal (6-connected) and diagonal

(26-connected) structures were both tested. The trials with an orthogonal open operator

applied tended toward the more favorable result by restricting finger growth, even though

it also inhibited movement between self-same regions. The close operator, however,

created too many false bridges to be effective.

Image Variations

Several variations of the input image were presented to the region-growing

algorithm to determine if they could be used to improve the quality of the segmentation.









In each of these trials, the kernels were defined based on the altered image and allowed to

progress using the iterative scheme. However, the bone-defining threshold value selected

using the original image was retained and acts as a barrier for growth. This was done

because the intensity threshold is a good approximation of bone edge and these image

variations would undoubtedly result in less appealing threshold-based boundary

approximations.

First, it was thought that a smoothing filter applied to the original CT image would

weaken the intensity of anatomical cross-links and promote a more uniform, centralized

geometry from which to evolve. The resulting anatomical edges were very clean and

smooth, yet inappropriately labeled. This was attributed to smeared boundaries which

lessened the inhibition of the kernels to leak into false anatomy.

Next, the region-growing algorithm was tested on the gradient of the input image.

The gradient is a common image filter that is used to highlight boundaries (Figure 4-8).

A first note is that gradient kernels are much more disconnect at higher thresholds which

makes the index assignment phase more laborious. The resulting segmentations showed

similar success to those of the unaltered image, however, the borders were choppy and

the areas of bleed-through were relocated. One noticeable property of the gradient image

is that there are high intensity rings around the periphery of the vertebrae. This leaves the

interior relatively uniform at a low intensity value. Therefore, an inverted gradient image

would leave a strong centralized bulk kernel to feed into the iterative algorithm (Figure

4-8). The kernel assignment phase for this image variation, as predicted, was simple.

The large internal areas were easy to label. The resulting segmentation, however, left



















Figure 4-8. Unique properties of the input CT image may be isolated or combined to
produce a more favorable segmentation. On the left is the original CT scan.
In the center, the gradient of the image highlights boundaries. The inverse of
the gradient (right), in contrast, highlights the interior of the image.

unpredictable, jagged boundaries. Example segmentations of the mentioned image

variations are shown in Figure 4-9.

The previous image variation tests, individually, did not provide wholly beneficial

results. However, they each revealed advantageous properties which could be utilized to

improve segmentations. Since this region-growing scheme is based on global intensity

patterns, the various images were blended or combined to form a composite image with

more suitable qualities. The most prosperous combination was a blend of original CT

and inverse gradient images. The inverse gradient image served to add a more

centralized interior which would quickly unify distant pieces of the vertebra. The

original CT image contribution served to temper the erratic boundary formation that the

inverse gradient image produced. This method had the potential to improve the quality of

the segmentation, however, the optimal blending ratios were scan specific and tedious to

determine. Also, boundary defects still existed which would foil the segmentation.

Prelabeling

The kernel assignment phase of the region growing process has a strong influence

on the success of the segmentation. When circumstances require a high threshold level to

































Figure 4-9. Segmentation results using variations on the input CT image. The top row
indicates the input kernels and the bottom row displays the result of the
algorithm. Examples of smoothed image (left column), image gradient (center
column), and inverse gradient (right column) are shown.

create unique kernel elements, the algorithm may be left starved. Specifically, only a

limited percentage of voxels in any one piece of anatomy may be available to seed

growth. The consequence is a much smaller chance of the correct label reaching all other

areas of the anatomy before borders are assigned. The third avenue pursued to improve

the region-growing segmentation method is the inclusion of additional user-defined

labels. These labels, set during the kernel assignment phase, force the assignment of

unassigned voxels that are exposed during region growth. It is expected that these pre-

labeled voxels will appropriately spread to adjacent areas and result in a better

segmentation. The updated iterative action consists of these steps:









1. The step to a lower threshold and resulting isocontour.
2. The pre-label of unassigned voxels.
3. The constrained orthogonal dilation of the labeled regions.
4. The restriction of the labeled voxels to the area outlined by the isocontour.

The goal is to find a specific labeling technique that makes a strong contribution to

segmentation while requiring a minimal amount of user input. The two methods tested

were 2D image slice and sphere labeling.

The CT image slice was decided as a good candidate for pre-labeling because it is a

straightforward manual segmentation process. Sagittal image slices were chosen for the

pre-labeling study because the vertebrae are fairly easy to visualize and a medial slice is

likely to contain all of the boney sections. Between one and five slices were tested.

Generally, a sparse kernel model or disruptive pathology forces the user to seek

additional slices to label. The first problem noted is the pre-labeled slices tend to have

little influence on adjacent axial growth patterns unless they happen to cross the labeled

sagittal plane. The response was to attempt to position the pre-labeled slice over critical

contact areas, such as the transarticular facets. This moderately improved the output;

however, the determination of the most effective slice is a tedious process that often

requires frequent references to adjacent sagittal image slices.

In response to the inability of the 2D slice pre-labeling to effect growth in the axial

direction, sphere labeling was developed. This is simply the creation of spherical regions

given a user-defined point, radius, and label. Remarkably, this method showed little

improvement in segmentation quality. The reason is that kernels tended to develop in a

ring fashion a bit inside the edge of the vertebra before progressing to the border.

Therefore, in order to effect kernel growth, the sphere must extend very close to the

vertebral boundary. The definition of a sphere in this manner, however, requires strict









attention to adjacent slices to assure that there is no violation into adjacent anatomy by

the sphere. Any such occurrence would be very defeating to the overall result. The time-

consuming verification process for sphere placement makes this avenue discouraging and

impractical.

Instrumentation

The segmentation methods discussed rely on intensity profiles of the bones. When

surgical instrumentation such as rods, screws, and wires are implanted into the patient,

though, this profile is severely disrupted. Metallic implants cause auras of high intensity

where they are positioned and serve as bridges incorrectly linking anatomy. Fortunately,

the intensities produced are much higher than any occurring in the natural bone profile.

This allows an intensity threshold mask to subtract the instrumentation out of the region

growing algorithm. The iterative sequence is similar to pre-labeling; void regions are

created in which the dilating kernels cannot travel. After a couple of tests, this method of

instrumentation treatment proved very successful. The algorithm was able to navigate

around the instrumentation and produce a segmentation (Figure 4-10). The only

modification that was appended was an additional dilation or two of the instrumentation

region. The first reason is that partial volume effect alters bone profile slightly outside

the actual geometry of the instrumentation. The second reason is that implants that have

resided in the body for a lengthy period of time will likely have bone growth that alters

the natural bone profile.

Results

The process of segmentation with the region-growing algorithm was tested on a

total of 31 clinical, high-resolution cervical CT scans intended for spinal surgery.

Success was gauged by the amount of effort necessary to achieve an accurate



























Figure 4-10. Instrumentation subtraction. The instrumented area (purple) was
successfully removed from the region growing algorithm, severing the false
bridge and allowing the region-growing algorithm to naturally progress.

visualization of the segmented vertebrae. Note that this may be more stringent than what

is required for any specific registration method. Primarily, success was measured by the

amount of image slices needing corrective manual post-processing alterations, with an

awareness of input effort and algorithm runtime. There were 19 (61%) of the scans that

required no or minimal post-processing. There were 7 (23%) that required manual

processing, yet still remained a significant improvement over manual methods. The

segmentations that showed no significant improvement over manual segmentation

amounted to 5 (16%). The algorithm run time was roughly 3-5 minutes (Pentium 4 at 1.4

IMHz) for 3-5 vertebral levels. The array size was a very significant factor in runtime

since the routines operate on full rectangular matrices. Optimization schemes would

significantly reduce processing time.

There are a couple of key factors that affected the outcome of this segmentation

method. First, the quality of the kernel regions had a significant contribution. If the seed




















Figure 4-11. Segmentation failure: boundary leakage. Seed regions defined distant to an
anatomical boundary must carry the accurate label to the boundary before a
competing label invades. The center image depicts the boundary leakage with
the final segmentation on the right.

Object 4-3. Segmentation failure: boundary leakage movie (object3-failureleakage.mpg,
25.8 MB).

regions were small and disconnected, then considerable weight was placed on the ability

to branch and interconnect properly. The failure of those actions leads to boundary

leakage (Figure 4-11). Threshold models that are created at a very high intensity produce

these low quality kernels. Areas of high intensity contact will force the seed region

selection to a high threshold. Severe posturing can change the proximity of the bones and

bring high intensity areas into contact, usually at the transarticular surfaces. The second

primary component to success of the segmentation process was image resolution. The

cervical boundary profile requires a high level of detail. A low-resolution image will

increase partial volume effects yielding incomplete intensity patterns and increase

opportunity for failure. A similar effect is seen when posturing compresses the space

between vertebrae. The last main ingredient to success of the segmentation was the

condition of the bone intensity profile. Some patients had comparatively lower bone

contrast with soft tissue, which created threshold models with significant gaps. These

gaps effected the ability of the anatomic labels to spread. The intensity profile was also









disrupted with pathology such as fusion. The most severe result observed was kernels

that dilated with little distinction between correct anatomical locations (Figure 4-12).











Figure 4-12. Segmentation failure: weak bone intensity profile. A CT scan that has low
bone to soft tissue contrast or pathology that disrupts the clarity of the bone
profile does not allow the dilating kernel regions to properly meet at the
vertebral boundaries.

Object 4-4. Segmentation failure: weak bone intensity profile movie (object4-
failureweak.mpg, 28.6 MB).














CHAPTER 5
CONCLUSIONS AND FUTURE WORK

A semi-automatic method has been developed to expedite the production of

segmented CT images beyond that of manual methods. It promotes accurate

visualization of bony components and encourages the use of image guidance with

surgeries involving multiple levels. This region-growing hysteresis method has a couple

of notable benefits. First, it is a relatively simple procedure to implement. It has no pre-

defined datasets to build or maintain. It works well with fracture pathology. Plus,

instrumentation can be easily subtracted without significantly affecting the performance

of the algorithm.

In contrast, this segmentation process has a few limitations. One is the potential for

lengthy user intervention, such as during kernel assignment or post-processing, if

required. Second, the threshold model does not completely define what is considered the

interior of the anatomy. Some of the interior regions have the same intensity as the

surrounding soft tissue and so are not included in the model. Region filling methods may

be used to fill the interior; however, this is not a thorough solution. Note that image

guidance registration may not require a segmentation with complete internal labeling to

be successful. Third, the algorithm is susceptible to small intensity defects, especially

those that bridge adjacent bones.

A couple of directions exist which have the potential to improve upon the

developed segmentation method, remedying some of its shortcomings. The region-

growing hysteresis method relies strictly on global voxel-based operations. Additional









elements, such as boundary or region-based components, if integrated, may provide the

additional information necessary to improve the segmentation. For example, an elastic

boundary concept and internal region statistics could be used to make the method more

robust to noise and point defects. Also, any statistical geometric profile could be used to

guide the evolving boundary to its correct location. One cost to these options, however,

is the maintenance of a potentially cumbersome dataset. For the addition of these

parameters to a method to be successful, it must be considered how well-suited the

algorithm is to incorporate the information. For example, in the current algorithm,

choosing a penalty term to curvature for a boundary-based representation may be difficult

with regard to the chaotic nature of how the regions intersect. After review, we decided

to pursue a method that was perhaps more suited to the integration of the mentioned

voxel, boundary, and region-based components.

Leventon (2000) presents a noteworthy approach to segmentation by calling upon

curvature models, prior shape models, and statistical image-surface relationships to

perform segmentations. One distinct advantage to this method is it blends these

techniques into a level set framework, allowing the boundary to evolve over complex

topography. This particular algorithm uses a signed distance map as the level set surface

representation. It has several advantages. First, boundary and region-based

representations are implicitly contained in the signed distance function. Second, the

notion of "holes" in the boundary is seamlessly contained in the representation; a small

"hole" in the object does not create an entire new set of features in the object. Third, it is

robust to slight misalignment of features. One interesting aspect of Leventon's method is

the use of intensity/distance-to-boundary distributions. These profiles are incorporated as





























Figure 5-1. Intensity/distance-to-boundary profiles. The vertical axis is grayscale value,
with the top having the highest intensity. The horizontal axis is distance from
object boundary (vertical black stripe). Right indicates interior distance from
the boundary and left indicates the exterior distance from the boundary. Color
indicates frequency of pairs. The first three images on the left show the
profile of vertebra C2 from three random individuals. The three images on the
right show three different levels, Cl, C2, and C3 (center to right) from one
individual.

training sets into a model that makes boundary estimations. Specifically, with this

particular adaptation of the level set framework, distance maps are used to describe the

higher-dimensional surface. The distribution of surface-intensity pairs, then, is a voxel

match of the morphological distance map of the segmented object, and it's associated CT

image value. A cursory examination of intensity/distance-to-boundary distributions of

spinal images was performed to determine if these profiles would be useful for

segmentation. Six cervical spine CT images were manually segmented, and vertebrae

from Cl to C4 were individually isolated. From that, 31 distinct distance-intensity maps

were created (Figure 5-1). The images were interpolated to equivalent square voxels






72


before the creation of the maps, so that the distance maps were accurate, and could be

compared. Also, no images were selected that contained significant intensity artifacts.

A visual comparison of the distance-intensity profiles shows many similarities. It

is anticipated that with some training, an algorithm could use this information to properly

locate a boundary. The next step would be to formulate a probability density model and

implement it into a segmentation routine, most likely that presented by Leventon (2000).















LIST OF REFERENCES


D. Adalsteinsson & J. Sethian, "The fast construction of extension velocities in level set
methods," The Journal of Computational Physics, 148:2-22, 1999.

L.P. Amiot, K. Lang, M. Putzier, H. Zippel & H. Labelle, "Comparative results between
conventional and computer-assisted pedicle screw installation in the thoracic,
lumbar, and sacral spine," Spine, 25(5):606-614, 2000.

R. Assaker, N. Reyns, M. Vinchon, X. Demondion & E. Louis, "Transpedicular screw
placement: image-guided versus lateral-view fluoroscopy: in vitro simulation,"
Spine, 26(19):2160-2164, 2001.

M.S. Austin, A.R. Vaccaro, B. Brislin, R. Nachwalter, A.S. Hilibrand & T.J. Albert,
"Image-guided spine surgery: a cadaver study comparing conventional open
laminoforaminotomy and two image-guided techniques for pedicle screw
placement in posterolateral fusion and nonfusion models," Spine, 27(22):2503-
2508, 2002.

J. Bangham & S. Marshall, "Image and signal processing with mathematical
morphology," Electronics & Communication Engineering Journal, 10:117-128,
1998.

S. Beucher & F. Meyer. "The morphological approach of segmentation: the watershed
transformation," Mathematical Morphology in Image Processing, E. Dougherty,
Ed., chapter 12, 433-481, Marcel Dekker, New York, 1992.

A. Blake & M. Isard, Active Contours, Springer-Verlag, London, 1998.

A. Bleau & L. Leon, "Watershed segmentation and region merging," Comp. Vision and
Image Understanding, 77(3):317-370, 2000.

G. Boehm, C.J. Knoll, V.G. Colomer, M. Alcafiiz-Raya & S.E. Albalat, "Three
dimensional segmentation of bone structures in CT images," In Proc. SPIE,
Medical Imaging, 3661:277-286, 1999.

V. Caselles, F. Catte, T. Coll & F. Dibos, "A geometric model for active contours in
image processing," Numerische Mathematik, 66:1-31, 1993.

V. Casellas, R. Kimmel & G. Sapiro, "Geodesic active contours," Int. Journal of
Computer Vision, 22(1):61-79, 1997.









A. Chakraborty, L. Staib & J. Duncan, "Deformable boundary finding in medical images
by integrating gradient and region information," IEEE Trans. Medical Imaging,
15(6):859-870, 1996, http://noodle.med.yale.edu/staib/papers/chakrabtmi.pdf, last
accessed March 22, 2005.

T. Chan & L. Vese, "An active contour model without edges," Scale-Space Theories in
Computer Vision, LNCS 1682:141-151, Springer-Verlag, London, 1999.

Y. Chen, S. Thiruvenkadam, H. Tagare, F. Huang, D. Wilson & E. Geiser, "On the
incorporation of shape priors into geometric active contours," In Proc. 1st IEEE
Workshop on Variational and Level Set methods in Computer Vision, pp.145-152,
IEEE, Los Alamitos, 2001.

G. Christensen, R. Rabbitt & M. Miller, "Deformable templates using large deformation
kinematics," IEEE Trans. on Image Processing, 5(10):1435-1447, 1996.

G. Christensen, "Consistent linear-elastic transformations for image matching," In Proc.
Information Processing in Medical Imaging, 1613:224-237, Springer-Verlag,
London, 1999.

K. Cleary, "Technical Requirements for Image-Guided Spinal Procedures", Workshop
Report, April 1999, ISIS Center, Department of Radiology, Georgetown University
Medical Center, Washington, D.C.,
http://www.visualization.georgetown.edu/SpineWorkshop/report/report.html, last
accessed March 22, 2005.

L. Cohen, "On active contour models and balloons," CVGIP: Image Understanding,
53(2):211-218, 1991.

T. Cootes & C. Taylor, "Active shape models 'smart snakes'," In Proc. British
Machine Vision Conference, pp. 266-275, Springer-Verlag, London, 1992.

T. Cootes, C. Taylor, D. Cooper & J. Graham, "Training models of shape from sets of
examples," In Proc. British Machine Vision Conference, pp. 9-18, Springer-Verlag,
London, 1992.

T. Cootes, C. Taylor, D. Cooper & J. Graham, "Active shape models their training and
application," Computer Vision and lmage Understanding, 61:38-59, 1995.

T. Cootes & C. Taylor, "A mixture model for representing shape variation," Image and
Vision Computing, 17(8):567-574, 1999.

T. Cootes, C. Beeston, G. Edwards & C. Taylor, "A unified framework for atlas matching
using active appearance models," In Proc. Information Processing in Medical
Imaging, 1613:322-333, Springer-Verlag, London, 1999.

T. Cootes, G. Edwards & C. Taylor, "Active appearance models," 5th European
Conference on Computer Vision, 2:484-498, Springer-Verlag, Berlin, 1998.









D. Cremers, F. Tischhauser, J. Weickert & C. Schnorr, "Diffusion snakes: introducing
statistical shape knowledge into the Mumford-Shah functional," International
Journal of Computer Vision, 50(3): 295-313, 2002.

O. Cuisenaire, "Distance transformations: fast algorithms and applications to medical
image processing," Ph.D. thesis, Universite cotholique de Louvain, Laboratoire de
Telecommunications et Teledetection, 1999,
http://www.tele.ucl.ac.be/PEOPLE/OC/these.pdf, last accessed March 22, 2005.

C. Davatzikos & J. Prince, "Convergence analysis of the active contour model with
applications to medical images," In Proc. SPIE, Visual Communications and Image
Processing, 1818:1244-1255, 1992.

A. Dempster, N. Laird & D. Rubin, "Maximum likelihood from incomplete data via the
EM algorithm," J. Royal Statistical Soc., Ser. B, 39(1):1-38, 1977.

C.A. Dickman, "Point of view," Spine, 25:2674, 2000.

E. Dougherty, Mathematical Morphology in Image Processing, M. Dekker, New York,
1993.

I. Dryden & K. Mardia, Statistical h\lq',e Analysis, Wiley, Chichester, 1998.

K.T. Foley, D.A. Simon & Y.R. Rampersaud, "Virtual fluoroscopy: computer-assisted
fluoroscopic navigation," Spine, 26:347-351, 2001.

G.E. Healthcare Technologies, Waukesha, WI, U.S.A.,
www.gehealthcare.com/index.html, 2004, last accessed March 22, 2005.

S. Geman & D. Geman, "Stochastic relaxation, Gibbs, and the Bayesian restoration of
images," IEEE Trans. on PAMI, 6(6):721-741, November, 1984.

N. Glossop & R. Hu, "Assessment of vertebral body motion during spine surgery," Spine,
22:903-909, 1997.

P. Golland, W.E.L. Grimson & R. Kikinis, "Statistical shape analysis using fixed
topology skeletons: corpus callosum study," In Proc. Inf. Proc. in Med. Imaging
(IPMI), 1613:382-387, Springer-Verlag, 1999.

P. Golland, W.E.L. Grimson, M. Shenton & R. Kikinis, "Small sample size learning for
shape analysis of anatomical structures," In Proc. MICCAI' 2000: Medical Image
Computing and Computer-Assisted Intervention, LNCS 1935:72-82, Springer-
Verlag, 2000.

Gray's Anatomy, http://education.yahoo.com/reference/gray/, last accessed March 22,
2005.









M. Jones & T. Poggio, "Multidimensional morphable models: a framework for
representing and matching object classes," International Journal of Computer
Vision, 29(2):107-131, 1998.

T. Kaneko, L. Gu & H. Fujimoto, "Abdominal organ recognition using 3D mathematical
morphology," In Proc. 15th International Conference on Pattern Recognition,
2:263-266, 2000.

M. Kass, A. Witkin & D. Terzopoulos, "Snakes: active contour models," Int. Journal on
Computer Vision, 1:321-331, 1987.

A. Kichenassamy, A. Kumar, P. Olver, A. Tannenbaum & A. Yezzi, "Gradient flows and
geometric active contour model," In Proc. IEEE Int. Conf. Computer Vision, pp.
810-815, IEEE, Los Alamitos, 1995.

R. Kikinis, S.K. Warfield & C.F. Westin, "High performance computing (HPC) in
medical image analysis (MIA) at the surgical planning laboratory (SPL)," In Proc.
High Performance Computing Asia, pp. 290-297, Singapore, 1998,
http://splweb.bwh.harvard.edu:8000/pages/papers/kikinis/hpc/HPCfinal.pdf, last
accessed March 31, 2005.

S. Lavallee, J. Troccaz, P. Sautot, B. Mazier, P. Cinquin, P. Merloz & J.P. Chirossel,
"Computer-assisted spinal surgery using anatomy-based registration," Computer-
Integrated Surgery: Technology and Clinical Applications, R. Taylor, S. Lavallee,
G. Burdea & R. Mosges, Eds., pp. 425-449, MIT Press, Cambridge, 1996.

R. Leahy, T. Herbert & R. Lee, "Application of Markov random fields in medical
imaging," Information Processing in Medical Imaging, J. Llacer & D. Ortendahl,
Eds., pp. 1-14, Wiley-Liss, NY, 1991.

M. Leventon, "Statistical models for medical image analysis," Ph.D. Thesis,
Massachusetts Institute of Technology, 2000.

J.A. Little, D.L.G. Hill & D.J. Hawkes, "Deformations incorporating rigid structures,"
Computer Vision and Image Understanding, 66(2):223-232, 1997.

C. Lorenz & N. Krahnstover, "3D statistical shape models for medical image
segmentation," In Proc. Int. Conf. 3-D Imaging and Modeling, pp. 414-423, IEEE,
Los Alamitos, 1999.

R. Malladi, J. Sethian & B. Vemuri, "Shape modeling with front propagation: a level set
approach," IEEE Trans. Pattern Anal. Machine Intell., 17:158-175, 1995.

G. Matheron, Random Sets and Integral Geometry, John Wiley & Sons, New York, 1975.









T. McInernery & D. Terzopoulos, "Medical image segmentation using topologically
adaptable surfaces," In Proc. Conf. Computer Vision, Virtual Reality, and Robotics
in Medicine and Medical Robotics and Computer-Assisted Surgery (CVRMed-
MRCAS), 1205:23-32, Springer-Verlag, Berlin, 1997.

T. McInerney & D. Terzopoulos, "Deformable models in medical image analysis: A
survey," Medical Image Analysis, 1(2):91-108, 1996,
http://mrl.nyu.edu/-dt/papers/mia96/mia96.pdf, last accessed March 22, 2005.

Medtronic Surgical Navigation Technologies, StealthStation, Louisville, CO,
www.stealthstation.com, last accessed March 22, 2005.

F. Meyer & S. Beucher, "Morphological segmentation," Journal of Visual Comm. and
Image Rep., 1(1):21-46, 1990.

M. Miller, A. Banerjee, G. Christensen, S. Joshi, N. Khadeja, U. Grenander & L. Matejic,
"Statistical methods in computational anatomy," Statistical Methods in Medical
Research, 6(3):267-99, 1997.

D. Mumford & J. Shah, "Optimal approximation by piecewise smooth functions and
associated variational problems," Comm. Pure and Applied Math., 42:577-684,
1989.

A. Neumann & C. Lorenz, "Statistical shape model based segmentation of medical
images," Computerized Medical Imaging and Graphics, 22:133-141, 1998.

H. Nguyen, M. Worring & R. van den Boomgaard, "Watersnakes: energy-driven
watershed segmentation," IEEE Trans. on Pattern Analysis and Machine
Intelligence, 25(3): 330-342, 2003.

Odin Medical Technologies, Inc., Newton, MA, www.odinmed.com, last accessed March
22, 2005.

K. Ohmori, Y. Kawaguchi, M. Kanamori, H. Ishihara, H. Takagi & T. Kimura, "Image-
guided anterior thoracolumbar corpectomy: a report of three cases," Spine,
26:1197-1201, 2001.

S. Osher & J. Sethian, "Fronts propagating with curvature dependent speed: Algorithms
based on Hamilton-Jacobi formulations," Journal of Computational Physics,
72:12-49, 1988.

N. Paragios & R. Deriche, "Geodesic active regions for supervised texture
segmentation," In Proc. IEEE Int. Conf. Comp. Vision, pp. 926-932, IEEE,
Los Alamitos, 1999.

A. Pentland & S. Sclaroff, "Closed-form solution for physically based shape modeling
and recognition," IEEE Trans. on Pattern Analysis and Machine Intelligence,
13(7):715-729, 1991.









T.M. Peters, "Image-guided surgery: from x-rays to virtual reality," Comput. Methods
Biomech. Biomed. Engin., 4(1):27-57, 2000.

R. Pichumani, "Construction of 3D geometric models for segmentation and visualization
of cervical spine images," Ph.D. thesis, Stanford University, November 1997.

Philips Medical Systems, Achieva I/T Interventional MR, Bothell, WA,
http://www.medical.philips.com, last accessed March 22, 2005.

J. Prewitt, "Object enhancement and extraction," Picture Processing and
Psychopictorics, B. Lipkin & A. Rosenfeld, Eds., pp. 75-149, Academic Press,
New York, 1970.

Y.R. Rampersaud, K.T. Foley, A.C. Shen, S. Williams & M. Solomito, "Radiation
exposure to the spine surgeon during fluoroscopically assisted pedicle screw
insertion," Spine, 25:2637-2645, 2000.

Y.R. Rampersaud, D.A. Simon & K.T. Foley, "Accuracy requirements for image-guided
spinal pedicle screw placement," Spine, 26:352-359, 2001.

L. Roberts, "Machine perception of three-dimensional solids," Optical and Electro-
optical Information Processing, J. Tippett, Ed., pp. 159-197, MIT Press,
Cambridge, 1965.

S. Sclaroff & A. Pentland, "Modal matching for correspondence and recognition," IEEE
Trans. on Pattern Analysis and Machine Intelligence, 17(6):545-561, 1995.

J. Serra, Image Analysis and Mathematical Morphology, Academic Press, New York,
1982.

J. Sethian, Level Set Methods: Evolving Interfaces in Geometry, Fluid Mechanics,
Computer Vision, andMaterial Science, Cambridge University Press, New York,
NY, 1996.

J. Sethian, "A fast marching level set method for monotonically advancing fronts," Proc.
Nat. Acad. Sci., 93(4):1591-1595, 1996.

J. Sethian, "Fast marching methods," SIAMReview, 41(2):199-235, 1999.

Siemens Medical Solutions USA, Siemens SIREMOBIL Iso-C3D, Inc., Iselin, NJ,
http://energized.esiemenshealthcare.com/SIREMOBIL/iso3d.htm, last accessed
March 22, 2005.

I. Sobel, "Camera models and machine perception," Technical Report AIM-21, Stanford
Artificial Intelligence Lab, Palo Alto, California, May 1970.

L. Staib & J. Duncan, "Boundary finding with parametrically deformable models," IEEE
Transactions on Pattern Analysis and Machine Intelligence, 14:161-175, 1992.









G. Szekely, A. Kelemen, C. Brechbuhler & G. Gerig, "Segmentation of 2-D and 3-D
objects from MRI volume data using constrained elastic deformations of flexible
Fourier contour and surface models," Med mage Anal., 1(1): 19-34, 1996.

H. Tek & B. Kimia, "Image segmentation by reaction diffusion bubbles," Proc. Int. Conf
Computer Vision, pp. 156-162, IEEE, Los Alamitos, 1995.

A.R. Vaccaro & K. Singh, "Computer-assisted image-guided surgery: application in
cervical spine surgery," Current Opinion in 01 ilhqpaedlik 12:245-250, 2001.

L. Vincent & P. Soille, "Watershed in digital spaces: an efficient algorithm based on
immersion simulations," IEEE Trans. on PAMI, 13(6),583-598, 1991.

Y. Wang & L. Staib, "Boundary finding with correspondence using statistical shape
models," In Proc. IEEE Conf. Comp. Vision andPatt. Recog., pp. 338-345, IEEE,
Los Alamitos, 1998.

Y. Wang & L. Staib, "Elastic model based non-rigid registration incorporating statistical
shape information," In Proc. Medical Image Computing and Computer-Assisted
Intervention (MICCAI), LNCS 1496:1162-1173, Springer-Verlag, London, 1998.

S. Warfield, A. Robatino, J. Dengler, F. Jolesz & R. Kikinis, "Nonlinear registration and
template driven segmentation," In Brain Warping, Arthur W. Toga, Ed.,
Ch.4:67-84, Elsevier Science & Technology Books, St. Louis, 1998.

A. Weidner, M. Wahler, S.T. Chiu & C.G. Ullrich, "Modification of C-C2 transarticular
screw fixation by image-guided surgery," Spine, 25:2668-2674, 2000.

W.C. Welch, B.R. Subach, I.F. Pollack & G.B. Jacobs, "Frameless stereotactic guidance
for surgery of the upper cervical spine," Neurosurgery, 40:958-964, 1997.

C.F. Westin, S. Warfield, A. Bhalerao, L. Mui, J. Richolt & R. Kikinis, "Tensor
controlled local structure enhancement of CT images for bone segmentation," Int.
Conf on Medical Image Computing and Computer-Assisted Intervention
(MICCAI'98), LNCS 1496:1205-1212, Springer-Verlag, London, 1998.

C.F. Westin, J. Richolt, V. Moharir & R. Kikinis, "Affine adaptive filtering of CT data,"
Medical Image Analysis, 4:161-177, 2000,
http://splweb.bwh.harvard.edu:8000/pages/papers/westin/media2000/westinMedia2
000.pdf, last accessed March 22, 2005.

A. Yezzi, A. Tsai & A. Willisky, "A statistical approach to image segmentation for
bimodal and trimodal imagery," In Proc Int. Conf. Comp. Vision, pp. 898-903,
IEEE, Los Alamitos, 1999.

A.S. Youkilis, D.J. Quint, J.E. McGillicuddy & S.M. Papadopoulos, "Stereotactic
navigation for placement of pedicle screws in the thoracic spine," Neurosurgery,
48:771-779,2001.






80


S.C. Zhu & A. Yuille, "Region competition: unifying snakes, region growing, and
Bayes/MDL for multiband image segmentation," IEEE Trans. on PAMI,
18(9): 884-900, 1996.















BIOGRAPHICAL SKETCH

Matthew James Williams, born April 13, 1975, received a Bachelor of Science in

Engineering in mechanical engineering, with concentrations in design and biomedical

applications, from the University of Michigan in 1998. During his undergraduate years,

he was blessed with many rewarding experiences and opportunities that have helped to

shape his life. Some of these include internships at a variety of manufacturing and

technical facilities, capped with a cooperative at the Denso Tech Center in 1997. After

graduation, Matthew was hired as a project management consultant at ACM, Inc.

Throughout his career, he has been cementing his resolve to explore engineering

opportunities in medicine. This led him to the University of Florida Biomedical

Engineering Program; and in 2000, he migrated to Gainesville.

Now that he has finished the requirements for his master's degree, Matthew has

elected to continue taking advantage of the cornucopia of opportunities at the University

of Florida through the study of the regenerative medicine aspect of biomedical

engineering. Also, Matthew continues the pursuit of the performing arts, nourishing

creativity in all aspects of his life.