Efficient Large-Deformation Medical Image Registration via Recurrent Dynamic Correlation

TL;DR

We address large-deformation medical image registration by introducing ReCorr, a recurrent correlation-based framework that performs dynamic, voxel-to-region local searches with iterative search-center relocation. The architecture decouples motion and texture information in a lightweight recurrent updater and leverages a multi-scale pyramid to progressively bridge large spatial gaps with low per-step cost. Across brain MRI and abdominal CT datasets, ReCorr achieves competitive or superior accuracy while substantially reducing FLOPs and inference time, with ReCorr-S offering an even faster option for mild deformations and a diffeomorphic variant providing trajectory-preserving transformations. The model demonstrates strong performance in both affine-regular and unregistered settings, substantial efficiency gains, and robustness to synthetic perturbations, making it suitable for practical, real-time clinical deployment.

Abstract

Deformable image registration estimates voxel-wise correspondences between images through spatial transformations, and plays a key role in medical imaging. While deep learning methods have significantly reduced runtime, efficiently handling large deformations remains a challenging task. Convolutional networks aggregate local features but lack direct modeling of voxel correspondences, promoting recent works to explore explicit feature matching. Among them, voxel-to-region matching is more efficient for direct correspondence modeling by computing local correlation features whithin neighbourhoods, while region-to-region matching incurs higher redundancy due to excessive correlation pairs across large regions. However, the inherent locality of voxel-to-region matching hinders the capture of long-range correspondences required for large deformations. To address this, we propose a Recurrent Correlation-based framework that dynamically relocates the matching region toward more promising positions. At each step, local matching is performed with low cost, and the estimated offset guides the next search region, supporting efficient convergence toward large deformations. In addition, we uses a lightweight recurrent update module with memory capacity and decouples motion-related and texture features to suppress semantic redundancy. We conduct extensive experiments on brain MRI and abdominal CT datasets under two settings: with and without affine pre-registration. Results show that our method exibits a strong accuracy-computation trade-off, surpassing or matching the state-of-the-art performance. For example, it achieves comparable performance on the non-affine OASIS dataset, while using only 9.5% of the FLOPs and running 96% faster than RDP, a representative high-performing method.

Figures (8)

• Figure 1: Illustration of the recurrent local search strategy of ReCorr. A voxel from the fixed image (left) is progressively matched to its corresponding location in the moving image via recurrent local search and dynamic updates. At each iteration, the search is performed within a local neighborhood (blue), and the deformation field is incrementally updated (black arrows), guiding the search region toward more accurate correspondences.
• Figure 2: Architecture overview of the proposed ReCorr. Given a pair of images, ReCorr first extracts multi-scale features using a same encoder, then performs iterative refinement from scale $0$ to scale $3$. At each scale $i$, the $t$-th iteration begins by conducting local search and then updates the residual deformation field $\Delta \phi_t^{(i)}$, which is added to the deformation field from the previous iteration. Except for the zero-initialized scale 0, each scale is initialized by upsampling the final output from the preceding scale. The final deformation field $\phi_{\rm final}$ is refined at the original resolution using a Conv3D block.
• Figure 3: The structure of the local search module. Note that we omit the scale index '$i$' for simplicity.
• Figure 4: The structure of the recurrent updater for each scale.
• Figure 5: The two scatter plots compare registration accuracy (Dice) against inference time (left), and FLOPs (right) on the OASIS dataset without affine pre-registration. Circle size represents model parameters, where a larger circle indicates more parameters. Arrows ($\uparrow\!\!\! \ \downarrow$) indicate the preferable direction of metrics. The proposed ReCorr achieves a favorable trade-off between accuracy and efficiency in both views.