Directional averages for motion segmentation in discontinuity preserving image registration

Abstract

The registration of abdominal images is central in the analysis of motion patterns and physiological investigations of abdominal organs. Challenges which arise in this context are discontinuous changes in correspondence across sliding organ boundaries. Standard regularity criteria like smoothness, are not valid in such regions. In this paper, we introduce a novel regularity criterion which incorporates local motion segmentation in order to preserve discontinuous changes in the spatial mapping. Based on local directional statistics of the transformation parameters it is decided which part of a local neighborhood influences a parameter during registration. Thus, the mutual influence of neighboring parameters which are located on opposing sides of sliding organ boundaries is relaxed. The motion segmentation is performed within the regularizer as well as in the image similarity measure and is thus implicitly updated throughout the optimization. In the experiments on the 4DCT POPI dataset we achieve competitive registration performance compared to state-of-the-art methods.