Evaluation of 3D correspondence methods for building point distribution models of the kidney

Statistical shape models (SSM) in general and point distribution models (PDM) in particular, are used for a variety of medical applications. A prerequisite condition to build a PDM is to establish the correspondences between landmarks across sample models in a given training population. Previously both image registration based correspondence building algorithms as well as surface parameterization based correspondence establishment algorithms have been proposed. In this paper, we present comparative studies on 25 kidney samples of two state-of-the-art correspondence methods. More specifically, we compare a diffeomorphic demons image registration algorithm based correspondence method to a surface parameterization based method where the correspondences are optimized with Minimal Description Length (MDL) principle. The results suggest that the studied image registration based correspondence method has a superior performance over the studied surface parameterization based method.