Multi-object deformable templates dedicated to the segmentation of brain deep structures

Abstract

We propose a new way of embedding shape distributions in a topological deformable template. These distributions rely on global shape descriptors corresponding to the 3D moment invariants. In opposition to usual Fourier-like descriptors, they can be updated during deformations at a relatively low cost. The moment-based distributions are included in a framework allowing the management of several simultaneously deforming objects. This framework is dedicated to the segmentation of brain deep nuclei in 3D MR images. The paper focuses on the learning of the shape distributions, on the initialization of the topological model and on the multi-resolution energy minimization process. Results are presented showing the segmentation of twelve brain deep structures.