Establishing Local Correspondences towards Compact Representations of Anatomical Structures

Computer-aided diagnosis is often based on comparing a structure of interest with prior models. Such a comparison requires automatic techniques in determining prior models from a set of examples and establishing local correspondences between the structure and the model. In this paper we propose a variational technique for solving the correspondence problem. The proposed method integrates a powerful representation for shapes (implicit functions), a state-of-the-art criterion for global registration (mutual information) and an efficient technique to recover local correspondences (free form deformations) that guarantees one-to-one mapping. Local correspondences can then be used to build compact representations for a structure of interest according to a set of training examples. The registration and statistical modeling of Systolic Left Ventricle shapes in Ultrasonic images demonstrate the potential of the proposed technique.