PRISM: Differentiable Analysis-by-Synthesis for Fixel Recovery in Diffusion MRI

A research team led by Mohamed Abouagour, Atharva Shah, and Eleftherios Garyfallidis has introduced PRISM (Differentiable Analysis-by-Synthesis for Fixel Recovery in Diffusion MRI), a novel framework that addresses the nonconvex limitations of traditional voxelwise microstructure fitting. The model employs an explicit multi-compartment forward model that combines cerebrospinal fluid (CSF), gray matter, and up to K white-matter fiber compartments using a stick-and-zeppelin approach, with explicit fiber directions and soft model selection via repulsion and sparsity priors. PRISM supports both a fast mean squared error (MSE) objective and a Rician negative log-likelihood (NLL) that jointly learns sigma without requiring oracle information, and includes a lightweight nuisance calibration module for robustness.

In testing, PRISM demonstrated significant improvements over existing methods. On synthetic crossing-fiber data with SNR=30, PRISM achieved a best-match angular error of 3.5 degrees with 95% recall—1.9 times lower than the best baseline (MSMT-CSD at 6.8 degrees with 83% recall). When operating in NLL mode with learned sigma, the error dropped further to 2.3 degrees with 99% recall, successfully resolving fiber crossings as narrow as 20 degrees. On the DiSCo1 phantom using NLL mode, PRISM improved connectivity correlation over CSD baselines across all four tracking angles, achieving a best correlation of r=.934 at 25 degrees compared to .920 for MSMT-CSD. The framework also proved computationally efficient, completing whole-brain HCP fitting of approximately 741,000 voxels in just 12 minutes on a single GPU while maintaining near-identical results across random seeds.