DRT: Deformable Region-Based Transformer for Nonrigid Medical Image Registration With a Constraint of Orientation

Abstract

Nonrigid medical image registration is a key technology for clinical disease diagnosis by integrating the vital information taken from input images acquired using multiple imaging sensors. Recently, transformer networks have attracted attention in image registration, but splitting images into patches remains a problem. Although there are usually multiple structural regions with different sizes and shapes in medical images, current transformer methods usually use an equal-sized or rectangular patch embedding that cannot represent the structural differences and directionality of different regions. In addition, it may also destroy the semantic consistency of objects. To address this problem, we propose a new deformable region-based structural relevance embedding (DRSRE) module which learns to adaptively split the images into deformable structural regions of different sizes and shapes with orientation constraints rather than using equal-sized or rectangular patches. In this way, our method can represent the structural differences and directionality between regions and effectively preserve the semantic consistency in regions to calculate accurate transformation relationships for nonrigid image registration. We term the DRSRE module with a transformer part as a deformable region-based transformer (DRT) and conduct extensive evaluations of DRT in the context of nonrigid medical image registration, and the results validate that the proposed DRT outperforms the state-of-the-art nonrigid registration methods, achieving 0.748 and 0.886 Dice scores with 0.217% and 0.308% foldings for deformation field on the LPBA40 and ADNI datasets.