Recursive cross-stage modulated network for medical image registration

Medical image registration is a key task in medical image analysis, as it enables precise alignment of complex deformations between images. Current methods, however, face limitations in feature refinement, cross-stage information interaction, and multiscale deformation calibration. To address these issues, we propose a recursive cross-stage modulated registration framework. Our method recursively optimizes features across stages through three key components: (1) a dense feature refinement fusion (DFRF) module that preserves underlying details by fusing multilevel features; (2) a gated cross-stage feature modulation (GCFM) mechanism to dynamically enhance information transfer; and (3) a multifield alignment module (MFAM) for cross-stage feature calibration via multiperceptual fields. Experiments conducted on four public datasets [LONI Probabilistic Brain Atlas (LPBA), Mindboggle, Imperial College Institute of Neurology and Psychiatry (IXI), and Fast and Low-GPU-memory Abdomen CT Organ Segmentation Challenge (FLARE)] demonstrate that the proposed method achieves superior medical image registration performance compared with mainstream approaches, exhibiting enhanced accuracy and robustness.