Cortical surfaces stop crossing each other: zero collisions in 92.14% of 560 cases, accuracy still on par with the strongest baseline

All four cortical surfaces can be reconstructed together in one shared deformation space instead of each being fitted to its target boundary on its own. Earlier deep learning methods either reconstructed the white matter first and expanded to the pial surface, or predicted a separate displacement for each surface on the mesh, leaving the relative arrangement of surfaces unconstrained. Across 560 cases from 14 cohorts, thirteen unseen during training, the method showed no detected inter-surface collision in 92.14% (516/560) of cases, whereas every baseline produced at least one collision in every case.

It first builds topologically correct initial surfaces from a segmentation, then a network that reads both the T1 image and a cortical ribbon probability map predicts multi-scale velocity fields and moves all four surfaces together. The initial surfaces come from the subject's own segmentation: white matter surfaces are derived directly from the segmentation with topology correction, and pial surfaces are taken as adaptive isolevels of the white matter signed distance field, so the starting configuration is already collision-free. Evaluation covers 560 cases from 14 cohorts spanning ages 6 to 89, healthy and clinical populations, and scanners from three vendors, with every method scored against the same FreeSurfer reference surfaces under one evaluation implementation.

Researchers running population morphometry can apply the pipeline to existing T1-weighted cohorts and check whether collision and self-intersection rates are similarly near zero on their own data. Group analyses that need vertex-wise correspondence still require an added surface registration or template resampling step. Reproduction and comparison can use the released code, pretrained weights, preprocessed derivatives, and fixed splits.

The reference surfaces come from FreeSurfer without manual editing and serve as a silver standard, so all distance and thickness metrics are defined relative to it. Thickness is estimated as a nearest-surface white-to-pial distance rather than vertex-wise anatomical correspondence. Same-hemisphere white-pial contacts near the medial wall may be anatomically expected yet counted as collisions, which can slightly inflate collision values. Training saw only high-quality research-grade T1 images, so behavior on routine clinical scans remains to be observed.