Solving a Nonlinear Blind Inverse Problem for Tagged MRI with Physics and Deep Generative Priors
New CVPR 2026 framework solves three key cardiac imaging problems simultaneously for the first time.
A research team from Johns Hopkins University and Yale has introduced a groundbreaking AI framework for tagged MRI analysis, accepted at CVPR 2026. The paper 'Solving a Nonlinear Blind Inverse Problem for Tagged MRI with Physics and Deep Generative Priors' addresses long-standing challenges where periodic tags used to track tissue motion become entangled with anatomy and degrade over time. For decades, clinicians had to handle tag removal, motion estimation, and anatomical recovery as separate, suboptimal tasks. This new approach represents the first unified solution that tackles all three problems simultaneously through a blind inverse framework.
The core innovation lies in synergizing MR physics models with deep generative priors, allowing the system to estimate unknown forward imaging parameters while reconstructing high-resolution underlying anatomy. The model tracks 3D diffeomorphic Lagrangian motion over time, overcoming issues like tag fading due to T1-relaxation that previously disrupted brightness constancy assumptions. Experiments on tagged brain MRI demonstrate superior results: the framework produces clearer anatomical images, higher-quality cine sequences, and more accurate motion tracking than specialized methods. This represents a significant advance for cardiac imaging where precise motion analysis is critical for diagnosing conditions like heart failure and cardiomyopathies.
- Unifies three previously separate tasks: anatomical recovery, cine synthesis, and 3D motion estimation
- Combines MR physics models with deep generative priors to solve blind inverse problems
- Demonstrated on tagged brain MRI with results surpassing specialized methods in accuracy
Why It Matters
Enables more precise cardiac motion analysis from single MRI scans, potentially improving diagnosis of heart conditions.