Fast Reconstruction of Time-Resolved 4D MRA with Unrolled Neural Networks

Motivation: Recently an ASL-based accelerated 4D MRA was demonstrated with great potential in delineating dynamic blood flow patterns at high spatiotemporal resolution. However, the time-consuming reconstruction presents a bottleneck for wider clinical translation. Goal(s): To introduce DL-AngioNet, a ML-based framework that accelerates the reconstruction while providing improved SNR. Approach: The network was trained using historical data via a data-driven method with a physical model. The unrolled structure of the network provided a data consistency term to ensure validity of the results. Results: DL-AngioNet accelerated the reconstruction by ~30-fold while preserving good flow dynamic information. Results demonstrated superior SNR comparing to the conventional PICS method. Impact: DL-AngioNet significantly accelerates 4D MRA reconstruction by ~30-fold, which not only preserves good 4D MRA flow dynamics, but also provides improved SNR in the results. DL-AngioNet could facilitate 4D MRA into a wider clinical use.