TC2-Res: a structured fusion of tract-level and connectome-level brain imaging in small-sample cohorts of athletes

Combining diffusion tract descriptors with structural connectome descriptors may help characterize athlete-related brain imaging patterns; however, this approach is challenging in small-sample studies, where flexible learned models can easily overfit. In such settings, simple linear classifiers often serve as strong baselines, although they do not explicitly encode the anatomical correspondence between tract-level microstructural descriptors and connectome-level organization. To address this gap, we introduce Tensor Connectome Consistency Residual (TC2-Res), a lightweight structured fusion framework that combines pair-aware tract representations, global modality branches, family-level encoders, and a tract-connectome consistency regularizer that encourages matched anatomical families to align in a shared latent space. In a small cohort of collegiate athletes, TC2-Res achieved a slightly higher mean balanced accuracy compared to a matched naive learned fusion baseline on the primary football-vs.-others task, whereas a classical linear support vector machine (SVM) remained the strongest overall classifier. The observed performance gain was modest and inconsistent across folds, and evidence of improvement specifically attributable to the consistency term was limited. These results suggest that anatomically structured fusion represents a plausible lightweight design direction for learned multimodal classification in limited-data settings while also highlighting the continued strength of classical linear baselines in this regime.