Shape recognition with application to medical imaging

Recognition is one of the main tasks involved in human and computer vision. We provide a novel framework for 3D shape recognition with application to medical imaging. We apply our algorithms to differentiate cancerous structures from normal anatomical tissues in the human colon.; We address the recognition process by a combination of two tasks: feature estimation and statistical classification. The feature estimation process computes descriptors of objects. A classifier first builds a database of descriptors for previously seen objects, and then maps descriptors of novel images to categories corresponding to previously seen objects. We provide algorithms for both of these stages in addition with a feedback framework between these stages.; We designed a new 3D shape representation, called Random Orthogonal Shape Sections (ROSS) method for the feature estimation stage. ROSS method utilizes a random set of mutually orthogonal triples of shape sections. The statistics of triples is obtained in the form of histograms and is given as the shape signature. The ROSS signatures are fed to Support Vector Machine (SVM) Classification. SVM obtains the optimal differentiating hypersurface between the two classes of objects to be separated. In addition, SVM provides the data samples—called support vectors—that carry the differentiating characteristics between the two classes of objects. We provide a feedback framework, called Distinctive Component Analysis (DCA), which combines support vector samples with linear discriminant analysis to map the features of clustered support vectors to a lower dimensional space, where the two classes of objects are optimally separated.; Our recognition system improves the specificity of the previous methods at the same sensitivity level. The ROSS method does not assume anything about the nature of shapes, and is potentially applicable to any 3D shape. We also provide experiments with synthetic shapes and lung nodules.; Our contribution to shape recognition is not limited to medical applications. In this thesis, we also provide algorithms for tracking human face in 3D. The tracked features are used to recognize facial expressions. The system provides very good accuracy regardless of the pose of the face.