Dynamic MRI Reconstruction Using Tensor-SVD

In this paper we propose to reconstruct dynamic magnetic resonance images from highly sparse sampling k-t space data by enhancing the low rankness and sparsity simultaneously. We introduce Tensor Singular Value Decomposition (t-SVD) instead of matrix SVD to maintain the structure of dynamic MR images. The reconstruction is casted into an optimization framework where the tensor nuclear norm (TNN) minimization is used to enhance the low rankness and the l₁ norm minimization of tensor gradient along each mode is applied to enhance the sparsity. In addition, we utilize alternating direction method of multipliers (ADMM) algorithm to efficiently solve the proposed optimization problem. Experimental results demonstrate the superior performance of the proposed method.