Appearance and Shape Prior Alignments in Level Set Segmentation

We show a new segmentation technical method that takes shape and appearance data into account. It uses the level set technique. The algorithm merges the edge alignment and homogeneity terms with a shape dissimilarity measure in the segmentation task. Specifically, we make two contributions. In relation to appearance, we propose a new preprocessing step based on non-linear diffusion. The objective is to improve the edge detection and the region smoothing. The second and main contribution is an analytic formulation of the non-rigid transformation of the shape prior over the inertial center of the active contour. We have assumed gaussian density on the sample set of the shape prior and we have applied principal component analysis (PCA). Our method have been validated using 2D and 3D images, including medical images of the liver.