More with less: Simulation-based inference enables accurate diffusion-weighted MRI with minimal acquisition time

Abstract

Diffusion-weighted magnetic resonance imaging (dw-MRI) is an essential tool in neuroimaging, providing non-invasive insights into brain microstructure. However, obtaining reproducible and accurate maps requires lengthy acquisition due to the need to massively oversample the parameter space. This means that tensor-based dw-MRI accessibility is still relatively low in daily practice, and more advanced approaches with increased sensitivity and specificity to microstructure are seldom applied in research and clinical contexts. Motivated by recent advances in simulation-based inference (SBI) methods, this work uses neural networks to model the posterior distribution of key diffusion parameters when provided experimental data, allowing accurate estimation with fewer measurements and without the need to train on real data. We find that SBI outperforms standard non-linear least squares fitting under noisy and sparse data conditions in both diffusion tensor and kurtosis imaging, reducing imaging time by 90% while maintaining high accuracy and robustness. Demonstrated on simulated and real data in healthy and pathological brains, this approach can substantially impact radiology by: i) increasing dw-MRI access to more patients, including those unable to undergo long exams; ii) promoting advanced dw-MRI protocols for greater microstructure sensitivity; and iii) rescuing older data where noise hindered analysis. Combining SBI with dw-MRI could greatly improve clinical MRI workflows by reducing patient discomfort, enhancing scan efficiency, and enabling advanced imaging approaches in a data and privacy friendly way.