ABSTRACT Title of Document X RAY IMAGE SEGMENTATION AND AN INTERNET BASED TOOL FOR MEDICAL VALIDATION Jing Cheng Master of Science 2006 Directed By Professor Ramalingam Chellappa Department of Electrical and Computer Engineering Segmentation of vertebrae in X ray images is a difficult task that requires an effective segmentation procedure Noise poor image contrast occlusions and shape variability are some of the challenges in many of the spine X ray images archived at the U S National Library of Medicine NLM In this thesis we propose a curvaturebased corner matching approach which exploits the posterior corners of the vertebra to estimate the location and orientation of the vertebrae The key advantage of the proposed approach is execution time roughly about one fifth of the previous approach that uses the generalized Hough transform when tested on a sizeable set of cervical spine images This thesis also presents the first ever effort to develop a prototype internetbased medical image segmentation and pathology validation tool which enables radiologists to validate computer generated image segmentations modify existing or create new segmentation in addition to identifying pertinent pathology data X RAY IMAGE SEGMENTATION AND AN INTERNET BASED TOOL FOR MEDICAL VALIDATION By Jing Cheng Thesis submitted to the Faculty of the Graduate School of the University of Maryland College Park in partial fulfillment of the requirements for the degree of Master of Science 2006 Advisory Committee Professor Ramalingam Chellappa Chair Professor K J Ray Liu Professor Adrianos Papamarcou Copyright by Jing Cheng 2006 Acknowledgements I am grateful to my advisor Professor Rama Chellappa for his scientific guidance unwavering support and constant encouragement throughout all of my research I would also take this opportunity to thank the U S National Library of Medicine for sponsoring this research work I would like to express my sincere thanks to Professor K J Ray Liu and Professor Adrian Papamarcou for serving on my thesis committee I would like to thank Dr Sameer Antani Rodney Long and Dr George Thoma at the U S national Library of Medicine for providing valuable inputs to this work Finally I would like to deeply thank my parents for always being there for me Thank you for all the love support and encouragement to go further Without you none of this would have been possible ii Table of Contents Acknowledgements ii Table of Contents iii List of Tables iv List of Figures v Chapter 1 Introduction 1 1 1 Motivation 1 1 2 Background 5 1 3 Thesis Organization 8 Chapter 2 Segmentation of Spine X ray Image 10 2 1 Spine Vertebra and Pathology Interests 10 2 2 Investigation of Contemporary Segmentation Approaches 12 2 2 1 Shape Operator 12 2 2 2 Active Contour Segmentation ACS 13 2 2 3 Generalized Hough Transform GHT 18 2 2 4 Steerable Filter 24 2 3 A Proposed Image Segmentation Scheme 26 2 3 1 Scheme Outline 26 2 3 2 Image Enhancement 28 2 3 3 Corner Detection 34 2 3 4 Definition of the Template 37 2 3 5 Template Corner Matching 38 2 3 6 Fine Segmentation 40 Chapter 3 Experiments and Results 42 3 1 Data Set and Ground Truth 42 3 2 Experiments and Performance Measurements 43 3 3 Computational Issue 46 Chapter 4 Image segmentation and Pathology Validation Tool 50 4 1 Design Considerations 50 4 2 Features and Capabilities 51 4 3 Lessons learned 56 Chapter 5 Conclusion and Future Perspective 57 Bibliography 59 iii List of Tables Table 2 1 A summary of snake implementation 17 Table 3 1 Comparisons of execution time for previous and current coarse segmentation scheme 47 iv List of Figures Figure 1 1 Illustration of a Cervical spine and b Lumbar spine in radiographic images 1 Figure 1 2 An example of cropped section from a lumbar X ray showing two vertebrae exhibiting double edges 4 Figure 1 3 a Anterior Osteophytes illustrated on vertebral body outline Points marked indicate regions of interest b Spinal X ray image showing segmented vertebra and a localized view showing inferior AO on vertebra with 36 boundary points superimposed 6 Figure 2 1 Human spine nomenclature 12 Figure 2 2 a A section of the original X ray image of cervical vertebra b Initial contour created by inputting 9 points in blue circle along the edges of the vertebral bodies c Orthogonal curves in defining the search grid used in Active Contour Segmentation d Final contour after deformation The image section has been enhanced to show the details in b c and d 14 Figure 2 3 a Initial contour created by inputting 9 points in blue circle along the edges of the vertebral bodies b Final contour after deformation c Orthogonal curves in defining the search grid used in Active Contour Segmentation The image section has been enhanced to show the details 15 Figure 2 4 Generalized Hough Transform a Illustration of the geometry for building the R table b The R table format 20 Figure 2 5 Example cervical spine shape template with location of the reference point marked in the template The minimum vertical bounding box is depicted in red rectangle 21 Figure 2 6 Example of GHT results a misplaced the template by one vertebra down b placed the template approximately close to ground truth Output of GHT is denoted in blue marks and the ground truth landmark points LPs is in red 23 Figure 2 7 Flow chart of the proposed segmentation scheme 27 Figure 2 8 a Example of original cervical image b Enhanced cervical image after adaptive histogram equalization c Example of original lumbar image d Enhanced lumbar image after adaptive histogram equalization 30 Figure 2 9 Unsharp masking applied after adaptive histogram equalization 31 Figure 2 10 a A section of original X ray image of cervical vertebra b Enhanced image after adaptive histogram equalization c Enhanced image after adaptive histogram equalization and unsharp masking d Enhanced image after applying 3rd Derivative of Gaussian separable steerable filters 0 09 1 5 33 Figure 2 11 a Region of interest of original cervical X ray image b Resulting image after de noising c Edge image after applying Canny detector on a d Edge image after applying the Canny detector on b 34 Figure 2 12 Illustration of a corner detection b template corner matching 39 Figure 2 13 Calculating the orientation of the object model from the corner points passing through the posterior corners of the cervical vertebrae 40 v Figure 3 1 Illustration of the ground truth morphometric points for a cervical spine C3 C6 b lumbar spine L1 L5 42 Figure 3 2