C+Mag: New AI method sharpens dynamic MRI without extra scan time
Researchers harness k-space magnitude data to improve MRI reconstruction with no additional acquisition time.
Dynamic MRI captures moving anatomy like the heart, but fast imaging requires undersampling k-space, which causes reconstruction artifacts. Standard deep learning methods rely solely on complex-valued measurements, but this team found that k-space magnitude information remains highly consistent across time frames in steady-state dynamic scans. That observation led to C+Mag, a magnitude-informed, physics-driven reconstruction framework that combines complex measurements with auxiliary magnitude data without any extra scan time.
The method uses an ADMM-based unrolling architecture, where a novel magnitude-aware data-fidelity term replaces the usual Euclidean loss. Since magnitude constraints are non-convex and non-differentiable, the authors applied quadratic smoothing and momentum-based updates to ensure stable optimization. They validated C+Mag on retrospectively undersampled cine MRI and phase-contrast flow MRI datasets, plus prospectively undersampled real-time cine acquisitions. Results show better artifact suppression, sharper anatomical recovery, and more accurate phase information than conventional physics-driven deep learning baselines, with blinded expert radiologists favoring the C+Mag outputs.
- C+Mag exploits k-space magnitude consistency across time frames to guide reconstruction without extra scan time
- Uses ADMM-unrolled deep learning with a magnitude-aware data-fidelity term, quadratic smoothing, and momentum updates
- Outperforms conventional physics-driven DL on cine, phase-contrast, and real-time MRI, confirmed by blinded expert evaluation
Why It Matters
Faster, sharper dynamic MRI could improve cardiac and flow imaging in clinical practice without prolonging patient scans.