Active appearance model segmentation in medical image analysis

A model-based method for two-dimensional, dynamic two-dimensional, and three-dimensional image segmentation is developed and evaluated for the segmentation of volumetric cardiac magnetic resonance (MR) images. This thesis covers a comprehensive design of 2D Hybrid, 2D+Time, and 3D Active Appearance Models (AAM) based on extending the AAM framework introduced by Edwards, Taylor and Cootes.; Under the AAM technique, manually traced segmentation examples create a statistical model of appearance during an automated training stage. Information about the shape and appearance of similar objects is contained in a single model ensuring a spatially and/or temporally consistent segmentation. This technique segments images by optimizing this appearance model onto the target image.; In addition to segmentation, we develop a fully automated landmark placement technique to assign point corresponding landmarks onto to a set of training images. This is critical for the 3D AAMs because determining point correspondence is a time consuming and often ill-posed task.; AAM coefficients, generated in the segmentation process, capture the shape and appearance variations of the target object; therefore we hypothesize that AAM coefficients may be used for the classification of disease abnormalities. Classification techniques like Linear Discriminant Analysis, Kernel Discriminant Analysis, and Support Vector Machines showed tendency of disease prediction within a 9% misclassification error. This thesis demonstrates the clinical potential of AAM techniques in short-axis cardiac MR images. We assess the method's performance by comparing manual independent standards in 162 images for 2D, 25 image sequences for 2D+Time, and 56 volumes for 3D models. The methods showed good agreement with independent standards using quantitative indices of border positioning errors and endo- and epicardial volumes. An automated initialization method is included making the segmentation approach fully automated.