The Probabilistic Active Shape Model: From Model Construction to Flexible Medical Image Segmentation

Automatic processing of three-dimensional image data acquired with computed tomography or magnetic resonance imaging plays an increasingly important role in medicine. For example, the automatic segmentation of anatomical structures in tomographic images allows to generate three-dimensional visualizations of a patient’s anatomy and thereby supports surgeons during planning of various kinds of surgeries. Because organs in medical images often exhibit a low contrast to adjacent structures, and because the image quality may be hampered by noise or other image acquisition artifacts, the development of segmentation algorithms that are both robust and accurate is very challenging. In order to increase the robustness, the use of model-based algorithms is mandatory, as for example algorithms that incorporate prior knowledge about an organ’s shape into the segmentation process. Recent research has proven that Statistical Shape Models are especially appropriate for robust medical image segmentation. In these models, the typical shape of an organ is learned from a set of training examples. However, Statistical Shape Models have two major disadvantages: The construction of the models is relatively difficult, and the models are often used too restrictively, such that the resulting segmentation does not delineate the organ exactly. This thesis addresses both problems: The first part of the thesis introduces new methods for establishing correspondence between training shapes, which is a necessary prerequisite for shape model learning. The developed methods include consistent parameterization algorithms for organs with spherical and genus 1 topology, as well as a nonrigid mesh registration algorithm for shapes with arbitrary topology. The second part of the thesis presents a new shape model-based segmentation algorithm that allows for an accurate delineation of organs. In contrast to existing approaches, it is possible to integrate not only linear shape models into the algorithm, but also nonlinear shape models, which allow for a more specific description of an organ’s shape variation. The proposed segmentation algorithm is evaluated in three applications to medical image data: Liver and vertebra segmentation in contrast-enhanced computed tomography scans, and prostate segmentation in magnetic resonance images.