Attention-Driven Framework for Non-Rigid Medical Image Registration

Deformable medical image registration is a fundamental task in medical image analysis, with applications in disease diagnosis, treatment planning, and image-guided interventions. Despite recent advances in deep learning based registration, accurate alignment under large anatomical deformations remains challenging, particularly when deformation plausibility must be preserved. In this paper, we propose an Attention-Driven Framework for Non-Rigid Medical Image Registration (AD-RegNet). The proposed method combines a 3D U-Net backbone with bidirectional cross-attention to establish multi-scale correspondences between moving and fixed images. It further introduces regional adaptive attention to emphasize anatomically relevant structures and a multi-resolution deformation field synthesis strategy to improve alignment while reducing unrealistic warping. AD-RegNet is evaluated on two datasets: DIRLab 4D CT for thoracic registration and IXI MRI for brain registration. On DIRLab, AD-RegNet reduces the mean target registration error (TRE) from $8.46 \pm 5.48$ mm before registration to $1.51 \pm 1.39$ mm after registration. On IXI, it achieves a Dice Similarity Coefficient (DSC) of $0.759 \pm 0.121$ and a low percentage of negative Jacobian determinants of $0.137\% \pm 0.293\%$, indicating competitive overlap accuracy and anatomically plausible deformations. Additional evaluation using normalized cross-correlation (NCC), mean squared error (MSE), structural similarity (SSIM), Jacobian determinant analysis, and ablation studies further confirms the contribution of the proposed attention and multi-resolution components. The results show that AD-RegNet provides a favorable balance between registration accuracy, deformation plausibility, and computational efficiency for 3D medical image registration.