BIDMIR: Bi-Directional Medical Image Registration with Symmetric Attention and Cyclic Consistency Regularization

The past years witness the remarkable progress in developing deep learning-based image registration methods, which leverage convolutional neural networks (CNNs) for efficient and end-to-end regression of deformation fields from an image pair. Identified limitations of existing methods include (a) ignorance of intra- and inter-image long-range spatial relevance, leading to failure in finding semantically meaningful correspondences of anatomical structures; and (b) single direction of image registration with no enforcement of topology preservation, resulting in loss of structural information. To address these issues, we propose a novel bi-directional medical image registration method, referred as BIDMIR, integrating symmetric attention with cyclic consistency regularization. The proposed method consists of a learnable volumetric embedding module, a symmetric attention module for feature enhancement, and a bi-directional registration field inference module. The symmetric attention module explicitly models the intra- and inter-image long-range relevance in the embedding, facilitating bi-directional correspondence of semantically meaningful structures. Cyclic consistency regularization is additionally proposed to encourage topology preservation. Results demonstrate the efficacy of our approach.