DW-MLSR: Unsupervised Deformable Medical Image Registration Based on Dual-Window Attention and Multi-Latent Space

Abstract

(1) Background: In recent years, the application of Transformers and Vision Transformers (ViTs) in medical image registration has been constrained by sliding attention mechanisms, which struggle to effectively capture non-adjacent but critical structures, such as the hippocampus and ventricles in the brain. Additionally, the lack of labels in unsupervised registration often leads to overfitting. (2) To address these issues, we propose a novel method, DW-MLSR, based on dual-window attention and multi-latent space. The dual-window attention mechanism enhances the transmission of information across non-adjacent structures, while the multi-latent space improves the model’s generalization by learning latent image representations. (3) Experimental results demonstrate that DW-MLSR outperforms mainstream registration models, showcasing significant potential in medical image registration. (4) The DW-MLSR method addresses the limitations of sliding attention in transmitting information between non-adjacent windows, improves the performance of unsupervised registration, and demonstrates broad application prospects in medical image registration.