FlexMorpher: Deformable Attention-Enhanced Multi-Scale Displacement Field Refinement with Spatial Diffusion Model

Deformable image registration is essential in medical imaging, but existing approaches often struggle to jointly model fine-grained deformations and maintain topological plausibility. We propose FlexMorpher, a unified probabilistic framework that reformulates registration as a spatially guided denoising task. FlexMorpher integrates a score-based diffusion model with a multi-scale displacement-to-deformation refinement (DDR) network enhanced by deformable attention. The diffusion module estimates a structure-aware score field conditioned on anatomical priors from the fixed image, while the DDR network translates this into dense deformation fields through coarse-to-fine refinement. This combination enables FlexMorpher to capture both global structural alignment and localized non-rigid deformations. Experimental results on 2D and 3D registration benchmarks demonstrate that FlexMorpher outperforms existing methods in accuracy, deformation smoothness, and topological stability. These results highlight the effectiveness of integrating generative diffusion modeling with spatial attention mechanisms for robust and anatomically consistent medical image registration.