Simulation-based inference at the theoretical limit for fast, robust microstructural MRI with minimal diffusion data

Diffusion-weighted magnetic resonance imaging provides a non-invasive way to probe brain tissue microstructure and is widely used in neuroscience and clinical research. Reliable microstructural maps usually require long scan times because many measurements are needed to sample the underlying parameter space. This limits clinical feasibility and accessibility. The aim of this study is to determine whether simulation-based inference can reduce the amount of diffusion data required while preserving estimation fidelity across commonly used diffusion models.

We apply simulation-based inference using neural posterior estimation to infer diffusion parameters directly from measured signals. The approach is tested on diffusion tensor imaging, diffusion kurtosis imaging, and biophysical models of axonal density and size. Models are trained entirely on simulated data and evaluated using both simulated datasets and experimental brain data from healthy and pathological individuals. Performance is compared with standard non-linear least squares fitting under noisy and sparsely sampled conditions.

Here we show that simulation-based inference achieves reliable parameter estimates using up to 90% fewer measurements than conventional approaches. The method consistently outperforms standard fitting when data are noisy or limited and remains robust across models, sampling schemes, and both healthy and pathological brain data.

This study demonstrates that simulation-based inference enables fast and robust microstructural imaging with substantially reduced scan times. The approach supports privacy-preserving workflows, could expand dMRI access, e.g., for pediatric and other time-sensitive patients, enable advanced microstructure-sensitive protocols, and rescue legacy data with suboptimal quality.