A physics-informed generative foundation model for quantitative diffusion MRI that maps brain microstructure (tensor, kurtosis, NODDI) and adapts zero-shot to each participant's data.
Diffusion MRI is the only noninvasive technique that can probe the microscopic tissue architecture of the living human brain. Translating raw diffusion-weighted images into reliable quantitative microstructure maps — diffusion tensor (DTI), diffusion kurtosis (DKI), and Neurite Orientation Dispersion and Density Imaging (NODDI) parameters — has traditionally required dense q-space sampling and carefully optimized acquisition protocols, confining quantitative mapping largely to specialized research settings. PIGMENT (Physics-Informed Generative Microstructure Network) addresses this bottleneck with a pretrained generative foundation model that learns a universal prior of human brain microstructure and then adapts to each new participant's measured data without any retraining.
Introduced in a May 2026 arXiv preprint by Zihan Li, Jialan Zheng, and colleagues — work consistent with the Tsinghua University diffusion-MRI group behind related physics-informed methods such as DIMOND and Diff5T — PIGMENT reframes microstructure estimation as a generative inference problem rather than a per-scan model fit. (The arXiv record does not list author affiliations; the Tsinghua attribution is inferred from the senior author's research group and should be confirmed against the published paper.) The network is trained once across a large, heterogeneous corpus of brain scans, after which it conditions its generative prior on whatever measurements a given subject provides — even extremely sparse or accelerated acquisitions — to recover full quantitative maps in a zero-shot fashion.
This positions PIGMENT alongside a growing class of physics-informed and self-supervised deep-learning methods for diffusion MRI (such as DIMOND), but it is distinguished by its foundation-model framing: a single pretrained prior that generalizes across sites, vendors, field strengths, and signal models rather than a method retrained for each dataset or protocol.
PIGMENT is a physics-informed generative network pretrained on 11,375 brain diffusion-MRI scans aggregated across multiple imaging centers, scanner vendors, and magnetic field strengths. The model learns a universal generative prior over human brain microstructure; at inference, it adapts this prior zero-shot to a new participant by conditioning on that subject's measured diffusion signal, eliminating the per-scan optimization or retraining that conventional model fitting and many deep-learning approaches require. The same pretrained prior produces DTI, DKI, and NODDI parameter maps. The authors report external validation across five independent centers, including evaluation under heavily accelerated and sparse acquisitions and on low-field hardware, with downstream maps preserving cortical patterns and pediatric white-matter developmental trends. As a preprint, PIGMENT does not yet have publicly released pretrained weights, code, or a model hub deployment; the arXiv paper is currently the only available resource.
PIGMENT targets neuroimaging researchers and clinicians who need quantitative microstructure maps but cannot run lengthy, densely sampled research protocols. By recovering reliable DTI, DKI, and NODDI parameters from sparse or accelerated scans — and even from low-field scanners — it could broaden access to quantitative diffusion MRI in clinical neuroimaging, pediatric and developmental studies, and resource-limited or point-of-care settings. Its cross-vendor, cross-field generalization also makes it attractive for harmonizing multi-site cohort studies where acquisition protocols differ.
PIGMENT illustrates how the foundation-model paradigm — pretrain a broad generative prior once, then adapt zero-shot — can transfer from sequence and structure modeling into quantitative medical imaging. If its five-center external validation and accelerated-acquisition results hold up under peer review and independent reproduction, the approach could lower the barrier to quantitative diffusion MRI, particularly on low-field and abbreviated protocols that are otherwise unsuitable for microstructure estimation. As a recent preprint without released weights or code, its real-world adoption remains to be established, and claims of biological fidelity and generalization will benefit from external replication.