Deformable image registration relies on accurate anatomical correspondences. Existing learning‑based methods usually treat matching scores alone and ignore how reliability interacts with spatial context selection and deformation update strength. This omission becomes critical in homogeneous, weakly textured, or heavily deformed regions where ambiguous matches can produce erroneous displacements. We introduce ReliMorph, a symmetric framework for 3D deformable registration. The core idea is to characterize correspondence ambiguity with local correlation distributions built from shared anatomical features and adapted to each imaging modality. These reliability cues serve two purposes: they weight the fusion of spatial experts and they gate residual velocities, progressively limiting update magnitude across a coarse‑to‑fine pyramid. Inverse‑inconsistency feedback at each level is fed back to re‑adjust routing and gating decisions. To our knowledge ReliMorph is the first 3D deformable registration system that jointly conditions expert routing and velocity gating on correspondence reliability. Extensive experiments on both unimodal and multimodal tasks show consistent superiority in anatomical overlap metrics, confirming the benefit of reliability‑guided control.
Review