ReMiDi: Reconstruction of Microstructure from Diffusion MRI Signal

Motivation: Current dMRI techniques are limited to estimating simple 3D diffusion tensors indicating whether diffusion is restricted, and provides minimal voxel level microstructural detail. Goal(s): To estimate the tissue microstructure at voxel level from dMRI signal. Approach: We implement a differentiable dMRI simulator based on matrix formalism solution of BTPDE and use backpropagation guided by a decoder to reconstruct microstructure for a given signal. Results: We successfully optimized arbitrary 3D meshes using a differentiable dMRI simulator combined with a spectral autoencoder. We also performed comprehensive ablation studies on 3D mesh shapes, and diffusion sequence parameters and directions. Impact: Our work achieves a next-generation dMRI reconstruction, opening the ability to reconstruct brain microstructure with arbitrary meshes. This will enable in-vivo mesoscopic mapping of the human brain, and lead to improved biomarkers for Multiple Sclerosis and Traumatic Brain Injury.