Fluid-Physics Informed Deep-Learning Enabled 2-point Velocity Encoded 4D flow MRI

Motivation: 4D Flow MRI enables comprehensive hemodynamic assessments, but its clinical usage is hindered by long scan times. Goal(s): To enable a 2-point velocity encoded 4D flow MRI by using deep-learning to estimate the missing two velocity vector-components, reducing scan time by 50%. Approach: Convolutional neural networks (CNNs) were trained with a single velocity vector-component as input data to generate a 3D velocity vector field (complete 4D flow dataset). CNN performance was evaluated in peak velocity, net and peak flow, and Qp/Qs compared to standard 4D flow MRI. Results: AI-derived 4D flow MRI showed strong-to-excellent agreement to standard 4D flow MRI across all comparisons. Impact: This technique enables reduction of 4D flow MRI scan time and data acquired by 50%. Future work will focus on coupling our method with conventional imaging acceleration techniques to achieve greater scan time reductions and/or improvements in temporal resolution.