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 significant advances in deep learning based registration methods, accurately aligning images with large deformations while preserving anatomical plausibility remains a challenging task. In this paper, we propose a novel Attention-Driven Framework for Non-Rigid Medical Image Registration (AD-RegNet) that employs attention mechanisms to guide the registration process. Our approach combines a 3D UNet backbone with bidirectional cross-attention, which establishes correspondences between moving and fixed images at multiple scales. We introduce a regional adaptive attention mechanism that focuses on anatomically relevant structures, along with a multi-resolution deformation field synthesis approach for accurate alignment. The method is evaluated on two distinct datasets: DIRLab for thoracic 4D CT scans and IXI for brain MRI scans, demonstrating its versatility across different anatomical structures and imaging modalities. Experimental results demonstrate that our approach achieves performance competitive with state-of-the-art methods on the IXI and DIRLab datasets. The proposed method maintains a favorable balance between registration accuracy and computational efficiency, making it suitable for clinical applications. A comprehensive evaluation using normalized cross-correlation (NCC), mean squared error (MSE), structural similarity (SSIM), Jacobian determinant, and target registration error (TRE) indicates that attention-guided registration improves alignment accuracy while ensuring anatomically plausible deformations.

💡 Research Summary

The paper introduces AD‑RegNet, a novel deep‑learning framework for non‑rigid medical image registration that leverages attention mechanisms to improve correspondence estimation and anatomical plausibility. The architecture consists of four main components: (1) a 3‑dimensional U‑Net backbone that extracts multi‑scale feature maps from both the fixed and moving volumes; (2) a Bidirectional Cross‑Attention Module (BCAM) that explicitly models relationships between the two images at each resolution level. BCAM projects the moving and fixed features into query, key, and value spaces, computes attention weights in both moving‑to‑fixed and fixed‑to‑moving directions, and combines them with a learnable balance parameter α, thereby capturing both global and local correspondences even under large deformations. (3) A Regional Adaptive Attention (RAA) block that operates on high‑level encoded feature maps rather than raw voxels. The feature maps are partitioned into non‑overlapping patches (16³ for DIR‑Lab CT and 24³ for IXI MRI). For each patch a descriptor is obtained via average pooling, and a self‑attention across all descriptors yields importance weights w_i. The weighted sum reconstructs a refined feature map that emphasizes anatomically salient regions such as lung lobes or cortical structures. (4) A Multi‑Resolution Deformation Field Synthesis module that predicts a candidate displacement field φ_l at each level from the attention‑enhanced features, upsamples the coarser field, and blends them using learned blending coefficients α_l. The final deformation Φ_0 is applied once to avoid cumulative interpolation blur.

The loss function combines a similarity term (a weighted sum of Normalized Cross‑Correlation and Mean Squared Error) with a regularization term that enforces smoothness and penalizes non‑positive Jacobian determinants, encouraging diffeomorphic transformations.

Experiments were conducted on two publicly available datasets: DIR‑Lab (4‑D thoracic CT) and IXI (3‑D brain MRI). The authors evaluated registration quality using five metrics: NCC, MSE, Structural Similarity Index (SSIM), Jacobian determinant positivity, and Target Registration Error (TRE). Compared with state‑of‑the‑art learning‑based methods such as VoxelMorph, TransMorph, and VTN, AD‑RegNet achieved comparable or superior performance. Notably, TRE decreased from an average of 1.8 mm (baseline) to 1.2 mm, and the proportion of voxels with positive Jacobian determinants rose from 95.3 % to 98.7 %, indicating more anatomically plausible deformations. SSIM and NCC also showed modest improvements. Computationally, AD‑RegNet processed a full 3‑D volume in approximately 0.12 seconds on a modern GPU, making it suitable for real‑time clinical scenarios.

The authors discuss limitations, including the current focus on single‑pair registration rather than groupwise or multimodal settings, and the sensitivity of patch size and blending parameters to dataset characteristics. Future work is suggested to extend the framework to multimodal registration, incorporate automatic hyper‑parameter tuning, and integrate the method into clinical workflows.

In summary, AD‑RegNet demonstrates that integrating bidirectional cross‑attention with region‑level adaptive attention and hierarchical deformation synthesis yields a powerful, efficient, and anatomically consistent solution for non‑rigid medical image registration across different modalities and anatomical regions.