Deep-Learning Cortical Registration Guided by Structural and Diffusion MRI and Connectivity

Abstract

Accurate cortical surface registration is crucial for group-level neuroimaging analyses, yet geometry-based methods often yield suboptimal functional alignment across subjects due to substantial inter-individual variability. We present a novel deep-learning approach that incorporates white matter structural connectivity features from diffusion MRI (dMRI) tractography into the Joint Surface-based Registration and Atlas Construction (JOSA) framework. Our method generates vertex-wise connectivity maps by detecting streamline-surface intersections, followed by heat diffusion smoothing on the cortical manifold. We combine these connectivity features with scalar diffusion metrics (fractional anisotropy and apparent diffusion coefficient) and structural features as input to JOSA (which we call "JOSAConn"). Evaluated on HCP-YA subjects across 15 task contrasts, our method significantly outperforms FreeSurfer in functional alignment (Bonferroni-corrected p < 0.05 for 12 of 15 contrasts). This multimodal approach demonstrates that structural connectivity effectively bridges the gap between cortical geometry and functional organization while maintaining clinical applicability.