Integrating Local Precision With Global Consistency for Unsupervised Magnetic Resonance Image Registration

To address the issues of interpretability and learning efficiency in traditional single-stream frameworks, as well as the limitations in capturing long-range dependencies and semantic relationships caused by the dependence of two-stream architectures.

This paper proposes GLACF, a triple-stream global-to-local attention framework for deformable MR image registration, integrating CNNs, Transformers, and coarse-to-fine strategies to enhance alignment accuracy. Its key innovations include: a triple-stream feature extraction based on Transformers for enriched feature representation; multi-scale decoupling blocks (MSDBs) that leverage low-resolution features to generate coarse displacement fields; and a voxel-wise local attention module (VLFM) that refines high-resolution features to produce precise displacement fields.

Experimental results on three public 3D brain MRI datasets (e.g., LONI LBPA40, IXI, and OASIS) demonstrate that GLACF consistently outperforms state-of-the-art methods in registration similarity, smoothness, and invertibility. Specifically, on LONI LBPA40, GLACF achieves a Dice similarity coefficient (DSC) of 72.7%, structural similarity index measure (SSIM) of 0.973, percentage of negative Jacobian determinants (% of |J_ϕ| ≤ 0) of 0.15, and 95th percentile Hausdorff distance (HD95) of 6.088. On IXI, it attains 76.8% DSC, 0.952 SSIM, 0.48% of |J_ϕ| ≤ 0, and 3.284 HD95. On OASIS, it reaches 89.1% DSC, 0.973 SSIM, 0.55% of |J_ϕ| ≤ 0, and 1.328 HD95.

These results confirm the robustness and accuracy of GLACF across diverse datasets, highlighting its strong potential for clinical translation in high-fidelity image registration scenarios.