Spatial masked-set learning slashes diffusion MRI error by 22.4%

Diffusion MRI (dMRI) typically requires dense multi-shell sampling to capture rich microstructural information, but that means long acquisition times that are impractical in clinical settings. Standard sparse reconstruction methods often rely on analytical q-space models that can't fully capture complex white-matter geometry. Now, a team led by Yousef Sadegheih, Pratibha Kumari, and Dorit Merhof at the University of Regensburg has proposed a new deep learning approach that treats measured diffusion-weighted signals as an unordered set and learns to synthesize missing signals directly, bypassing the limitations of fixed sampling grids.

Their spatial masked-set framework combines a local 3×3×3 neighborhood for spatial context with radial-order-6 SHORE (SHell-ORiented) coefficient prediction for each center voxel. These coefficients can be decoded analytically to generate signals at arbitrary q-space locations, enabling flexible reconstruction. The training regime includes shell-wise gradient dropping, dense signal supervision, and rotation-consistent targets, ensuring the sparse input signals stay aligned with the coefficient supervision under data augmentation. Evaluated on held-out HCP100 white-matter voxels, the method outperformed both analytical q-space models and a continuous dMRI synthesis model. In the b=1000 setting with 10 input gradients, it hit 2.70% NMSE, a 22.4% relative reduction over the continuous baseline. The code is available online, and the work has been accepted at the MICCAI CDMRI Workshop 2026.