Adaptive correlation matching network with confidence-guided B-spline Free-Form deformation for unsupervised deformable brain MRI registration

Deformable image registration is a fundamental task in medical image analysis, aiming to establish accurate spatial correspondences between medical images acquired from different subjects or at different time points. Although learning-based unsupervised registration has advanced rapidly in recent years, most existing methods focus mainly on improving feature extraction or feature interaction. In contrast, the joint treatment of correspondence reliability and deformation estimation remains underexplored, which limits final registration performance. To address this issue, we propose an Adaptive Correlation Matching Network (ACM-Net) for unsupervised brain MRI registration. The proposed method uses a weight-sharing dual-stream encoder to extract hierarchical features and a progressive decoder to refine registration in a coarse-to-fine manner. This decoder comprises an Adaptive Correlation Matching (ACM) module and a confidence-guided B-spline Free-Form Deformation (BFFD) solver. The former converts local matching distributions into soft correspondence fields and confidence maps, while the latter uses them as reliability-weighted observations to solve a regularized continuous deformation field. Experimental results demonstrate that the proposed method achieves superior or comparable registration accuracy to representative baseline methods on three public brain MRI datasets, namely OASIS, LPBA40, and Mindboggle, while maintaining acceptable deformation regularity. Additional evaluation on a single abdominal CT benchmark provides preliminary evidence of the applicability of ACM-Net under this non-brain benchmark setting; however, it does not establish general cross-anatomical transferability. The source code is publicly available at https://github.com/wan722123/ACM-Net .