Accelerated 4D Flow MRI Using Deep Low-Rank Plus Sparse Neural Network and Complex Difference

Yuyang Ren, Wenjian Liu, Zijian Zhou, Haiqun Qi, Peng Hu

Abstract: 4D flow MRI is well-known for providing hemodynamic information but requires long scanning and image reconstruction time. To overcome such limitation, we propose a FlowL+S-Net technique for accelerated 4D flow MRI. FlowL+S-Net is a model-based reconstruction method that combines compressed sensing-based low-rank plus sparse decomposition(CS-L+S) and convolutional neural network. The framework of CS-L+S is maintained, but the sparse subproblem in CS-L+S were solved directly via a series of (1+3)D temporal-spatial convolutions. Additionally, we leverage the sparse prior of complex difference to segregate the network into a flow-compensated branch and a flow-encoded branch, thus enhancing the network’s capacity to learn velocity-related regions within flow-encoded images.The proposed method was validated through retrospective and prospective imaging experiments, comparing with CS-L+S and FlowVariationalNetwork(FlowVN). It achieved rapid 4D flow reconstruction in 9 seconds while CS-L+S took 20 minutes and FlowVN took 12 seconds. During retrospective experiments, FlowL+S-Net exhibited the best performance at acceleration factors ranging from 8 to 24. In prospective experiments with a 12 fold acceleration, the proposed method exhibited the highest agreement with the fully-sampled reference image in terms of peak flow and net flow.