PIGMENT: a physics-informed AI model for clinical brain MRI microstructure

A team led by Zihan Li et al. (multiple Chinese and international institutions) introduced PIGMENT (physics-informed generative microstructure network) to solve a key limitation of quantitative diffusion MRI: reliable microstructural mapping currently requires dense sampling and optimized protocols, confining it to research. PIGMENT learns a generative prior of human brain microstructure from a vast training set, then adapts zero-shot to each patient’s measured data — recovering subject-specific maps from scans that are far too sparse for conventional fitting. This physics-informed approach embeds the known signal models (tensor, kurtosis, NODDI) into a deep generative framework, ensuring biological plausibility even with minimal data.

The model was trained on 11,375 scans collected across multiple sites, vendors, and field strengths (including low-field systems). It was validated on external datasets from five independent centers, demonstrating reliable mapping of tensor, kurtosis, and NODDI parameters. Crucially, PIGMENT remains effective where conventional fitting breaks down: it produces meaningful maps from 10-fold accelerated acquisitions, preserves submillimeter cortical patterns, and tracks early-childhood white matter development. It also enables quantitative tensor mapping on cost-effective low-field scanners and extracts tumor biomarkers from ultra-fast clinical protocols. These results establish PIGMENT as a foundation model that extends quantitative diffusion MRI into regimes traditionally too sparse, heterogeneous, or clinically constrained for reliable analysis.