A foundation model for efficient and assumption-free characterization of brain microstructure from diffusion MRI

Diffusion MRI (dMRI) is the primary tool for in vivo investigation of brain microstructure. However, dominant analytical methods rely on biophysical models with specific assumptions and large computational requirements, limiting their biological accuracy and generalisability. Here, we introduce SSDiff, a self-supervised transformer-based foundation model that summarises the entire dMRI scan into a compact set of biologically interpretable latent features, without a priori model assumptions. Trained on ~80,000 scans from diverse populations, protocols, and ages, these features capture individual variability in both local microstructural properties (reflecting aspects of conventional metrics like fractional anisotropy and neurite density) and tract-level architecture. Across three large population datasets (N=65,316), these features significantly improved phenome-wide prediction of over 3,000 non-imaging phenotypes compared to standard dMRI biomarkers, such as diffusion tensor imaging, NODDI and structural connectomes. Crucially, SSDiff demonstrated robust zero-shot generalizability, accurately predicting clinical phenotypes across 7 independent cohorts (N=3,671) with different diseases, including data acquired with simplified clinical-grade protocols. A genome-wide association study on SSDiff features in 53,276 participants revealed 363 independent novel genetic loci for brain microstructure. Mendelian randomization analyses further revealed causal relationships between these features and brain disorders; for instance, linking the microstructural properties of the cerebellum to Parkinson's disease and of the hippocampus to Alzheimer's disease. SSDiff is a powerful, biologically interpretable foundation model for dMRI analysis, enabling more accurate clinical prediction and revealing novel genetic insights into brain health and disease.