Alternating Unrolling Network Of Jointly Low-Rank And Sparse Tensor For Accelerating Dynamic MRI

Deep unrolling networks have shown remarkable improvement over traditional optimization-based methods in accelerating dynamic MRI. The unrolling networks that combine low-rank (LR) and sparse priors also have the potential to further enhance the reconstruction performance. However, the iterative algorithm for solving the model of joint LR and sparse constraint is complicated, resulting in redundant network structure and computational burden if strictly unfolded into a deep unrolling network. In this paper, we propose a novel alternating unrolling network of jointly low-rank and sparse tensor for accelerating dynamic MRI, coined AlteRS-Net. Specifically, we propose a novel "DC-LR-DC-S" structure for each iteration phase of the network. DC is the data consistency layer and can be implemented by the gradient descent step. The tensor low-rank and sparse priors in the CNN-learned domain are exploited in LR and S layers, respectively. Experimental results demonstrate that the alternating architecture can achieve superior performance over the strictly unrolled one, and the proposed AlteRS-Net outperforms the SOTA unrolling networks.