SCMR/ISMRM Workshop: Data-driven image reconstruction and motion compensation-Low-Rank Deep Image Prior Reconstruction for Real-Time Cine Imaging at 1.5 and 0.55 Tesla

Low-Rank Deep Image Prior Reconstruction for Real-Time Cine Imaging at 1.5 and 0.55 Tesla

Video Transcription

Video Summary

Researchers developed a self-supervised low-rank deep image prior for real-time spiral functional cardiac MRI at 1.5T and 0.55T. The method uses two untrained U-Nets to learn spatial and temporal basis functions from a single undersampled scan, avoiding fully sampled training data. GROG preprocessing speeds training by enabling FFT-based reconstruction. In simulations and in vivo studies of healthy subjects, the approach outperformed L1 ESPIRiT and GRASP Pro, especially at higher acceleration, with strong noise and aliasing suppression. It achieved good agreement with standard gated cine imaging for ejection fraction, though reconstruction remains time-consuming.

Keywords

self-supervised MRI reconstruction

cardiac MRI

low-rank deep image prior

spiral functional imaging

GROG preprocessing