Multi-scale dual-attention frequency fusion for joint segmentation and deformable medical image registration

Abstract

Deformable medical image registration is a crucial aspect of medical image analysis. Improving the accuracy and plausibility of registration by information fusion is still a problem that needs to be addressed. To solve this problem, we propose DAFF-Net, a novel framework that systematically unifies three kind of information fusion (low-level fusion, high-level fusion, and loss fusion) to enhance registration precision and plausibility: (i) low-level fusion: DAFF-Net employs a shared global encoder to extract common anatomical features from both moving and fixed images in two tasks, reducing redundancy and ensuring foundational consistency across tasks; (ii) high-level fusion: through the dual attention frequency fusion (DAFF) module, DAFF-Net dynamically combines multi-scale registration and segmentation features, leverages features of low-frequency structural coherences and high-frequency boundary details, and adaptively reweighting them to enhance registration via global and local attention mechanisms; (iii) loss fusion: a unified loss function enforces bidirectional consistency, i.e., segmentation supervises registration through anatomical constraints, while registration refines segmentation via deformation-correct anatomical consistency. Extensive experiments on three public 3D brain magnetic resonance imaging (MRI) datasets demonstrate that the proposed DAFF-Net and its unsupervised variant outperform state-of-the-art registration methods across several evaluation metrics, demonstrating the effectiveness of our approach in deformable medical image registration. The proposed framework holds promise for practical clinical applications such as preoperative planning, longitudinal disease tracking, and structural analysis in neurological disorders.