Curvature-Regularized Graph Tractography for Reproducible Multi-Site Diffusion MRI Connectomes

This paper introduces a step-invariant minimum radius-of-curvature constraint for graph-based diffusion-MRI connectome construction, applied during beam-search path enumeration over an orientation graph. Prior graph tractography controls path smoothness with a fixed minimum-angle threshold that must be re-tuned to voxel resolution; expressing the criterion as a radius of curvature instead makes it step-length invariant and adds a single interpretable parameter with no per-dataset tuning. On 229 subjects across five independent acquisition sites, with reconstruction, atlas, and seeding held fixed, the constraint reduces pooled connectome edge-count variability by 19 percent, driven by a 35 percent reduction in the between-site component (cross-site harmonization), and it is more reproducible than both deterministic and probabilistic streamline tractography. The method is therefore a construction-time complement to post-hoc statistical harmonization: a simple, interpretable geometric prior that improves the cross-site consistency of graph-based connectomes, a first-order requirement for reliable multi-site diffusion-MRI studies.