Towards Intelligent Deformable Models for Medical Image Analysis

Medical imaging continues to permeate the practice of medicine, but automated yet accurate segmentation and labeling of anatomical structures continues to be a major obstacle to computerized medical image analysis (MIA). Deformable models, with its profound roots in estimation theory, optimization, and physics-based dynamical systems, represent a powerful approach to the general problem of medical image segmentation. This Thesis presents a number of novel contributions to the field of deformable modeling, and includes theory as well as application. In the first part of the Thesis, a modified Active Contour Model (ACM), utilizing adaptive inflation reversal and damping, is applied to segmenting oral lesions in color images. In the second part, the amalgamation of Active Shape Models (ASM) and ACM into a technique, that harnesses the powers of both, is applied to locating the left ventricular boundary in echocardiographic images. The third part of the Thesis discusses the development of two methodological extensions for spatiotemporal image analysis: Optical flow-based contour deformations, applied to contrast agent tracking in echocardiographic image sequences, and deformable spatio-temporal shape models for extending 2D ASM to 2D+time. The fourth part describes the use of a new Hierarchical Regional Principal Component Analysis, and presents two methods for interactive and learned, localized and multiscale, controlled shape deformation: medial-based shape profiles and physics-based shape deformations. In the final part of the Thesis, we develop Deformable Organisms: a robust decision-making framework for MIA that combines bottom-up, data-driven deformable models with top-down, knowledge-driven processes in a layered fashion inspired by Artificial Life modeling concepts. We present different segmentation and labeling examples of various anatomical structures from medical images and conclude that deformable organisms represent a promising new paradigm for MIA.