Spinverse: Differentiable Physics for Permeability-Aware Microstructure Reconstruction from Diffusion MRI

A research team led by Prathamesh Pradeep Khole and seven collaborators has introduced Spinverse, a groundbreaking approach to microstructure reconstruction from diffusion MRI (dMRI) that addresses a fundamental limitation in medical imaging. Traditional methods either assume impermeable tissue boundaries or estimate only voxel-level parameters without recovering explicit interfaces, limiting their ability to reveal actual biological barriers. Spinverse represents a paradigm shift by employing a fully differentiable Bloch-Torrey simulator that inverts dMRI measurements through a physics-aware optimization process, allowing the system to learn where diffusion barriers actually exist rather than assuming them.

The technical innovation centers on representing tissue on a fixed tetrahedral grid where each interior face's permeability becomes a learnable parameter. This approach enables microstructural boundaries to emerge naturally during optimization without changing mesh connectivity or vertex positions. To overcome the ill-posed nature of permeability inversion, the team implemented mesh-based geometric priors and a staged multi-sequence optimization curriculum that prevents local minima and avoids outline-only solutions. Across synthetic voxel meshes, Spinverse successfully reconstructed diverse geometries while demonstrating that sequence scheduling and regularization are critical for improving both boundary accuracy and structural validity, potentially revolutionizing how researchers analyze tissue microstructure from non-invasive imaging.