Application of a Probabilistic Statistical Shape Model to Automatic Segmentation

In this paper, we propose a method for applying a probabilistic statistical shape model (SSM) to automatic segmentation. We use a point-represented SSM which is based on correspondence probabilities instead of point-to-point correspondences as commonly used. In order to combine the a priori knowledge of the SSM with the image information during the segmentation, we employ a deformable surface whose deformation depends on on the shape prior given by the SSM on the one hand and on the image information on the other hand. We formulate this problem as an alternated minimization of an external energy term integrating the image information and an internal energy term integrating the SSM probabilities. In order to robustify the segmentation, we add statistical knowledge about typical organ intensities. This method is applied to the segmentation of the left kidney in noisy CT images with breathing artefacts and evaluated in comparison to the results of an active shape model (ASM).