More with less: Simulation-based inference enables accurate diffusion-weighted MRI with minimal acquisition time
Diffusion-weighted magnetic resonance imaging (dMRI) is an essential tool in neuro-science, providing non-invasive insights into brain microstructure. However, obtaining accurate maps demands long acquisitions, due to oversampling of the parameter space. Leveraging simulation-based inference (SBI) with neural networks, we directly approximate posterior distributions of diffusion parameters from experimental measurements without requiring real-data training. SBI achieves accurate parameter estimation using up to 90% fewer acquisitions and outperforms standard nonlinear least squares under noisy, sparse sampling. We demonstrate improvements across diffusion tensor imaging, diffusion kurtosis imaging, and biophysical models estimating axonal density and calibre. Validation on simulated and real datasets from healthy and pathological brains shows SBIs robustness and generalizability. This approach can expand dMRI access, e.g. for paediatric and other time-sensitive patients, enable advanced microstructure-sensitive protocols, and rescue legacy data with suboptimal quality. By reducing scan times and preserving privacy, SBI-integrated dMRI workflows promise faster, more comfortable exams, MRI-based virtual tissue biopsy, and have potential to substantially impact radiology workflows.