Synthesizing brain-originated realistic diffusion-weighted MRI signal for in silico experiments

Motivation: Collecting training datasets for machine learning applications in diffusion MRI is challenging due to the variability in acquisition protocols and difficulties in adequately labelling the data. Goal(s): To introduce a method to generate in silico diffusion-weighted data under realistic scenarios based on in vivo acquisitions. Approach: We derive an analytical approach to obtain a proxy for the extra-cellular perpendicular diffusivity, model the variabilities in brain characteristics using gamma distribution, and synthesize realistic diffusion-weighted signals from in vivo measurements. Results: The proposal enables the generation of diffusion-weighted training data under realistic brain-originated variability for a wide range of b-values. Impact: The proposal models brain-originated characteristics and enables diffusion-weighted data synthesis reflecting the observed MRI signal. Compared to previous solutions, which fix brain characteristics or draw them randomly, our approach realistically varies signal properties based on what is observed.