UNeXtMorph: Multi-scale UNet with Cross Attention for Deformable Image Registrations
Zhaoyang Liu, Sen Pan, Xiang‐Yu Kong, Xiuyang Zhao · 2025
Deformable registration is a key technique in medical image analysis, which can align a pair of images by establishing precise correspondence between them. However, insufficiently exploring the correspondence between moving and fixed images during the registration process challenges the precise registration. Besides, incomprehensive features caused by poor feature representation and inherent information loss in forward stage lead to relative rough deformations, failing to get fine registration. To address above limitations, we propose a mutual attention based framework with multi-scale perception for unsupervised deformable medical image registration, called UNeXtMorph. In UNeXtMorph, we introduce a mutual attention with multi-scale perception (MAMS) module to fully explore the correspondence between moving and fixed images. Meanwhile, this module facilitates non-local information capturing in a low computation cost. To facilitate the feature representation and effectively deliver multi-scale voxel correspondence discovered by MAMS to the deformations decoder stage, a mutual-attention features integration (MAFI) based on Gaussian modelings is proposed by readjusting spatial and channel features, and then the integrated features are injected into correspondent decoder layers via skip connection. The qualitative and quantitative results on human brain datasets demonstrate the superiority and feasibility of the proposed UNeXtMorph compared to other advanced image registration algorithms.