ProMA-Net: MR-TRUS prostate registration via a dual-stream Swin Transformer-based network with mixed attention.

Multimodal registration between magnetic resonance (MR) and transrectal ultrasound (TRUS) prostate images is crucial for targeted biopsy and therapy, yet this task remains extremely challenging due to complex cross-modality anatomical correspondences, fusion conflicts between modality-specific and modality-common features, and difficulty in balancing computational efficiency and registration accuracy. We address the challenging problem of deformable registration between preoperative MR and intraoperative TRUS prostate images by proposing a novel dual-stream Swin Transformer-based network. Our model processes MR and TRUS volumes with separate Swin Transformer-based encoding streams, and employs the Window Self-Cross Mixed Attention (WSCMA) mechanism with alternating window shifts to enable efficient multimodal feature fusion without the computational overhead of explicit attention. In addition, we introduce a lightweight Gating Cross-Feature Fusion (GCFF) module that adaptively reweights cross-modal features and a Multi-Scale Patch-Neighborhood Modality-Independent Neighborhood Descriptor (MIND) Contrastive Loss to reinforce both global and local anatomical alignment. We evaluate our approach on the public μProReg prostate dataset. Our method achieves superior registration accuracy, with the highest prostate Dice overlap and the lowest target registration error among recent advanced methods and shows the potential to improve image-guided prostate interventions by more accurately fusing MR and ultrasound data. The code is publicly available at https://github.com/Hongzunhei/ProMA-NET.