SynchroMorph: unsupervised synchronous multi-scale feature fusion for high-precision brain MRI registration

Structural brain MRI registration underpins group-wise neuroimaging and longitudinal studies, yet accurate deformable alignment remains difficult because of complex neuroanatomy, strong inter-subject variability, and large non-rigid deformations. We propose SynchroMorph, an unsupervised Transformer–CNN registration network that emphasizes synchronous fixed–moving feature modeling and topology-aware deformation regularization. In the encoder, a Multi-Scale Fusion Attention Block (MS-FAB) combines window self-attention with parallel 3³ and 5³ convolutions to integrate contextual correspondence cues and local anatomical boundaries. A bottleneck Multi-Scale Fusion (MSF) module further aggregates 3³/5³/7³ receptive fields, while a lightweight Feature Fusion Block (FFB) refines decoder features through upsampling, 3D convolutional refinement, and skip fusion. A Jacobian penalty is used to discourage non-positive local volume changes rather than to theoretically guarantee diffeomorphism. On OASIS, SynchroMorph attains DSC = 81.75% versus TransMorph’s 80.29% (+ 1.46 pp), improves boundary accuracy with HD95 = 1.86 mm (vs. 1.92 mm), and runs in 0.07 s/volume (vs. 0.12 s for TransMorph). On IXI, SynchroMorph reaches DSC = 78.66% and HD95 = 1.89 mm, with lower deformation variability (SDlog J = 0.12 vs. 0.19) and fewer non-positive Jacobian voxels than strong baselines. These results suggest that synchronous multi-scale feature fusion can improve the accuracy–regularity trade-off for T1-weighted brain MRI registration, particularly for small and high-curvature anatomical structures.