Denoising Vastly Undersampled Radial Portal Venous 4D Flow Data via Deep Learning

Radially undersampled 4D flow MRI is a promising method for non-invasive mapping of blood flow in the portal venous system. However, collecting sufficient projections to produce clinically viable images can lead to long scan times (10+ minutes) as fewer projections cause undersampling artifacts that appear as structured noise. In this study, we propose a data-driven, deep learning method to denoise vastly undersampled (<10% of full Nyquist sampling) radial 4D flow MRI data in the portal vein. We train a network on a heterogeneous, time-averaged dataset with two levels of undersampling and perform a quantitative hemodynamic analysis to compare results.