Directionally selective regularization for sliding preserving medical image registration

Medical image registration aims to recover the physical deformations between source and target images. As an ill-posed optimization problem, regularization energies are often used to guide the registration towards physically desirable resolutions. One challenge in registering volumetric images deformed by respiration-induced motion is to produce motion fields that are globally smooth, yet to allow for sparing physical sliding. In this study, we propose to utilize available partial prior knowledge about the sliding sites as well as their directions. We provide a general recipe for relaxed regularization that incorporates such prior knowledge. The proposed relaxation is both simple and flexible enough to be combined with robust penalty designs that would further improve the registration performance. In a preliminary test, two thoracic CT volumes at end inhale and end exhale were registered under extremely limited partial prior conditions, and the proposed method demonstrated its promise even in this pessimistic scenario.