Global-to-Local Shape Matching for Liver Segmentation in CT Imaging
Kinda Anna Saddi, Christophe Chefd, Farida Chériet · 2007
Abstract. We propose a two-stage algorithm to segment the liver in CT images. First, we estimate the pose and global shape properties using a statistical shape model defined in the low dimensional space spanned by a training set of shapes. Then, we apply a template matching procedure to recover local deformations that were not present in the learning set. In both steps, we optimize the same image term: the likelihood of the intensity inside the region of interest and its background. The method requires a single seed point inside the liver for the initialization. We show that this global-to-local strategy is able to recover livers with peculiar shapes in arbitrary poses. 1