Date of Award
Doctor of Philosophy (PhD)
Paul A. Yushkevich
James C. Gee
In biomedical image analysis, shape information can be utilized for many purposes. For example, irregular shape features can help identify diseases; shape features can help match different instances of anatomical structures for statistical comparison; and prior knowledge of the mean and possible variation of an anatomical structure's shape can help segment a new example of this structure in noisy, low-contrast images. A good shape representation helps to improve the performance of the above techniques. The overall goal of the proposed research is to develop and evaluate methods for representing shapes of anatomical structures. The medial model is a shape representation method that models a 3D object by explicitly defining its skeleton (medial axis) and deriving the object's boundary via "inverse-skeletonization". This model represents shape compactly, and naturally expresses descriptive global shape features like "thickening","bending", and "elongation". However, its application in biomedical image analysis has been limited, and it has not yet been applied to the heart, which has a complex shape. In this thesis, I focus on developing efficient methods to construct the medial model, and apply it to solve biomedical image analysis problems. I propose a new 3D medial model which can be efficiently applied to complex shapes. The proposed medial model closely approximates the medial geometry along medial edge curves and medial branching curves by soft-penalty optimization and local correction. I further develop a scheme to perform model-based segmentation using a statistical medial model which incorporates prior shape and appearance information. The proposed medial models are applied to a series of image analysis tasks. The 2D medial model is applied to the corpus callosum which results in an improved alignment of the patterns of commissural connectivity compared to a volumetric registration method. The 3D medial model is used to describe the myocardium of the left and right ventricles, which provides detailed thickness maps characterizing different disease states. The model-based myocardium segmentation scheme is tested in a heterogeneous adult MRI dataset. Our segmentation experiments demonstrate that the statistical medial model can accurately segment the ventricular myocardium and provide useful parameters to characterize heart function.
Sun, Hui, "Statistical Medial Model dor Cardiac Segmentation and Morphometry" (2010). Publicly Accessible Penn Dissertations. 97.