Accelerated Phase-Contrast Flow MRI with Low-Rank and Generative Subspace Modeling

Phase-contrast magnetic resonance imaging (PC-MRI) has been developed into a powerful technique for measuring blood flow dynamics in the heart and great vessels. However, its practical utility has been constrained by prolonged acquisition times, in particular for flow imaging experiments with high spatial and temporal resolution, broad volume coverage, and multi-directional velocity encoding. In this paper, we present a new learning-based image reconstruction method for accelerated PC-MRI, which incorporates a low-rank model with deep generative priors. Specifically, the proposed method utilizes a low-rank model to capture the strong spatiotemporal correlation of dynamic images in both temporal and velocity encoding directions for PC-MRI. Moreover, it employs an untrained generative neural network to represent the spatial subspace of the model, which mitigates the ill-conditioning problem due to the undersampling of the temporal subspace in the low-rank and subspace reconstruction. We develop an algorithm based on variable splitting and the alternating direction method of multipliers to solve the resulting optimization problem. We evaluate the performance of the proposed method on 2D cine PC-MRI experiments.