BrainCSD: a hierarchical consistency-driven MoE framework for connectome synthesis and multitask brain trait prediction

Functional and structural connectomes (FC/SC) are informative biomarkers for brain disorders, yet their clinical use remains limited by acquisition cost, lengthy connectome construction pipelines, and missing modalities in real-world neuroimaging studies. To address these challenges, we introduce BrainCSD (Brain Connectivity Synthesis and Decode), a hierarchical Mixture-of-Experts (MoE) framework for connectome-oriented representation learning and missing-modality connectome synthesis. BrainCSD does not require precomputed FC/SC matrices as direct inputs, but learns neuroanatomically constrained representations from fMRI or dMRI inputs with template- and atlas-based anatomical correspondence. Its architecture incorporates three consistency constraints: ROI-level alignment, encoding-level consistency, and network-level refinement. We evaluate BrainCSD across 7248 participants, 22 imaging sites, 9 datasets, 5 neurological and neurodevelopmental disorders, and 15 downstream tasks, including connectome synthesis, disease classification, brain-age prediction, and MMSE estimation. BrainCSD-generated connectomes provide useful complementary information for downstream diagnostic and trait prediction tasks. Additional balanced metrics, graph-topological analyses, and subgroup evaluations further support the potential utility of BrainCSD for connectome-based analysis. Nevertheless, the current results are based primarily on within-dataset validation, and larger prospective external validation, uncertainty calibration, and clinical expert assessment are required before any clinical use.