Staged Pretraining and Implicit Gridding Enable Fast 3D MR Fingerprinting

SPUR-iG, a deep unrolled reconstruction, yielded whole-brain 3D MR fingerprinting in under 15 seconds (111× speedup) and achieved T1 accuracy from 30-second scans matching that from 2-minute conventional scans.

• SPUR-iG uses implicit gridding with learned kernels for efficient FFT-based data consistency and a staged training strategy to make 3D unrolled learning tractable.
• On cross-vendor out-of-distribution data, it outperformed both locally low-rank and hybrid 2D-3D unrolled baselines in T1/T2 accuracy.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.