DisMorph AI separates MRI scanner distortion from true brain atrophy

Longitudinal MRI is essential for tracking brain changes in aging and neurodegenerative disease, but scanner-induced geometric distortions—especially gradient non-linearity (GNL)—vary across systems and protocols. Conventional deformable registration conflates these technical artifacts with true biological change, biasing morphometric measurements. To solve this, MGH/Harvard scientists created DisMorph, a deep learning registration framework trained entirely on synthetic data. It explicitly predicts two separate dense deformation fields: one encoding technical distortion, the other anatomical change. A novel generative model synthesizes both effects independently during training, providing disentanglement supervision, while domain randomization enables generalization across acquisition protocols.

DisMorph was validated in three settings. On simulated data with known ground truth, it detected anatomical changes more accurately and consistently than conventional registration. On real image pairs differing only by GNL distortion, it correctly assigned geometric change to the distortion field, demonstrating specificity. On longitudinal Alzheimer's disease (AD) pairs, DisMorph isolated AD-related structural changes while also identifying residual distortion remaining after standard correction. Accepted at the SASHIMI Workshop at MICCAI 2026, DisMorph paves the way for more reliable longitudinal morphometry in clinical settings where maintaining acquisition consistency is challenging—potentially improving early detection and monitoring of neurodegenerative disease.