Enhancing fMRI Registration Accuracy via Multi-Scale Cycle-Consistent Neural Networks

In recent years, deep learning-based registration methods have revolutionized medical image registration. However, existing approaches for fMRI registration often neglect the rationality of predicted deformation fields and fail to adequately consider detailed functional region information. This paper introduces a multi-scale cycle-consistent network registration model that extracts spatially aggregated functional characteristics derived from temporal BOLD signals at multiple scales, ensuring deformation field plausibility and registration accuracy. Validated using the 1000 Functional Connectomes Project dataset, the method enhances coverage of co-activated functional regions and significantly improves suprathreshold voxel counts (average increases of 16.8%, 7.6% and 6.4% compared to SPM, VoxelMorph and TransMorph, respectively). Furthermore, it achieves optimal peak t-values in group-level statistical maps, confirming its superior capability in maintaining functional consistency across registered images. The dataset comes from the publicly available 1000 Functional Connectomes Project dataset.