Title: A Physics-Guided Foundation Model Enhances Quantitative Diffusion MRI

To truly comprehend the human brain, researchers must gain insight into its microscopic tissue architecture. While diffusion magnetic resonance imaging (MRI) offers the sole noninvasive method for visualizing whole-brain microstructure in living subjects, the process of generating reliable quantitative maps has historically been restricted to specialized research environments. These settings typically demand dense data sampling and highly optimized acquisition protocols.

To bridge this divide, we introduce the Physics-Informed Generative Microstructure Network (PIGMENT). This approach learns a universal generative prior of human brain microstructure and adapts it in a zero-shot manner to individual participants’ measured data, thereby recovering subject-specific maps. The model was trained on a comprehensive dataset comprising 11,375 scans collected across various sites, vendors, and magnetic field strengths.

In evaluations using external datasets from five independent centers, PIGMENT successfully enabled reliable quantitative mapping for tensor, kurtosis, and NODDI models. The network demonstrated robust performance in scenarios where conventional fitting methods failed, capable of extracting meaningful maps from extremely sparse acquisitions. Additionally, it supports downstream applications such as structural connectivity mapping and tractography.

PIGMENT’s estimates exhibited strong biological validity. It preserved submillimeter cortical microarchitectural patterns and captured early-childhood white matter developmental trajectories, even when derived from scans accelerated tenfold. Furthermore, the model facilitates reliable quantitative tensor mapping on cost-effective low-field systems and allows for the extraction of tumor-related biomarkers using ultra-fast clinical protocols.

Collectively, these findings position PIGMENT as a physics-informed foundation model that expands the reach of quantitative diffusion MRI. It makes reliable analysis possible in regimes that were previously considered too sparse, heterogeneous, or constrained by clinical limitations.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC