Abstract:
The work of this dissertation focuses on a specifc computer aided diagnosis (CAD) problem, although the main concept can be generalized to similar problems. Our aim is to detect the presence of the spina bi da (open spine) neural tube defect that is evident for a physician when the fetal skull image of a subject is examined. The objective of applications performing automatic anomaly detection can be set in their original contexts. Such systems, as a second observer, may help avoid false diagnoses. Fetal skull shapes possess markers that signal the presence of spina bi da. That is why this thesis concentrates on exploiting features extracted from skull shapes obtained via ultrasound (US). Among the variety of shape representation and feature extraction schemes, we have implemented and experimented with two. Both the curvature scale space (CSS) and moment-based (i.e. Zernike moments) representations have proved to be robust in that the extracted features provide classi cation invariant under the similarity transformations of translation, rotation and scaling. Classi cation of shapes is commonly coupled with the problem of segmentation. Since the fully-automatic segmentation of US images is practically di cult, we have attempted to achieve segmentation semi-automatically after the manual marking of a small number of points on images, based on simple heuristics and the Active Shape Models (ASM). Our experiments use k-nearest neighbor (kNN) and Support Vector Machines (SVM) classi ers. The inherent problem of rarity of medical data sets is tackled with methods of undersampling and oversampling. The results, reported for ground truth segmentations, reveal the availability of optimal operating points serving particular objectives.
