Abstract:

Compressed sensing (CS) is a novel and important technique in MRI reconstruction, which can be used to reconstruct magnetic resonance images with few sampled data while simultaneously speeding up imaging. The conventional CS-based MRI is implemented from a global view, which leads to some disadvantages: it not only loses many local structures but also fails to preserve detail information. To obtain better reconstruction quality, we propose a novel CS-based reconstruction model, which is incorporated with nonlocal technology to allow for the preservation of extra details. The proposed model groups similar patches within the nonlocal area and stacks them to form a 3D array unit. Then, to process the array in a realistic 3D manner, a tensor-based sparsity constraint is developed as the regularization on the reconstructed image. Experimental results show that the performance of the proposed method is superior to those of conventional methods.