Amr Abdel-Dayem, Fall 2005, "Computer Aided Diagnostic for Carotid Artery Ultrasound and Breast Mammogram Images", Computer Science Department, University of Western Ontario, Canada.

Ph.D. Thesis Abstract

Medical imaging is an important tool in diagnosing several diseases. Medical imaging is an interdisciplinary area that requires high collaboration between clinicians and computer scientists. Image processing, artificial intelligence, image understanding and classification techniques are the basic foundations towards a better understanding and analyzing medical images. The objectives of this thesis are to develop Computer Aided Diagnostic (CAD) systems to automatically extract the walls of the carotid artery (using ultrasound images) and to highlight suspicious regions on breast tissues (using mammogram images). The main contribution of this dissertation is the development of various promising CAD systems in both applications. For the carotid artery application, seven systems were introduced. These systems utilize morphological operations, watershed segmentation, fuzzy region growing, fuzzy c-Means, multiresolution analysis, and graph-cuts to accomplish the segmentation task. Experimental results demonstrate that the outputs produced by these systems are comparable to active contour models (which are very popular in clinical applications.). Unlike active contour models, our systems require minimal user interaction (i.e. they are more user friendly). Moreover, they do not suffer from either the collapse or the convergence problems, which are common problems in active contour models. However, our systems have shortcoming in dealing with images that have the shadowing effect. Further treatment is still needed to solve this special case.

For the second application, two systems for highlighting suspicious masses in digital mammogram images were introduced. The first one is based on minimizing the fuzzy entropy of the images. Whereas, the second system uses Bayesian decision theory to estimate an optimal threshold that can be used to extract the suspicious lesions. Moreover, we developed a new performance measure that can be used to compare the computer-segmented and the clinician-segmented images. The statistical analysis over a set of sample mammogram images shows that the proposed systems have high sensitivity and specificity ranges within both the 99% and the 95% confidence intervals. In future work, a classification stage will be added to the system to classify the segmented lesions into malignant and benign tissues.