Segmenting articulated structures by hierarchical statistical modeling of shape, appearance, and topology

Abstract

This paper describes a general method for segmenting articulated structures. The method is based on statistical parametrical models, obtained by principal component analysis (PCA). The models, which describe shape, appearance, and topology of anatomic structures, are incorporated in a two- level hierarchical scheme. Shape and appearance models, describing plausible variations of shapes and appearances of individual structures, form the lower level, while the topological model, describing plausible topological variations of the articulated structure, forms the upper level. This novel scheme is actually a hierarchical PCA as the topological model is generated by the PCA of the parameters obtained at the lower level. In the segmentation process, we seek the configuration of the model instances that best matches the given image. For this purpose we introduce coarse and fine matching strategies for minimizing an energy function, which is a sum of a match measure and deformation energies of topology, shape, and appearance. The proposed method was evaluated on 36 X-ray images of cervical vertebrae by a leave-one-out test. The results show that the method well describes the anatomical variations of the cervical vertebrae, which confirms the feasibility of the proposed modeling and segmentation strategies.