Sparse Tensor Based Magnetic Resonance Imaging Algorithm

Dynamic magnetic resonance imaging is an important medical imaging technique in current medical imaging which is used in clinical diagnostics widely.However,current imaging technology can hardly guarantee the both of reconstruction accuracy and reconstruction efficiency.To achieve good image reconstruction performance on both speed and accuracy,we do some work on it.Propose a periodic time-variant subsampling and omnidirectional total variation based reconstruction method called hybrid CS-DMRI.It uses periodic time-variant subsampling for different frames.In each period,there is one reference frame which is sampled at a higher subsampling ratio and a lower subsampling ratio for others.The nearby two reference frames with good reconstruction quality can be used to give the predictions of the other frames between them.Using periodic bidirectional reference frames method can make the most of the correlation of sequences.We propose a novel omnidirectional total variation regularization for exploiting the sparsity of all the possible directions of the data.The proposed model constrains the residual frame and image frame in the finite difference domain of all directions joinly.The formulated optimization model can be solved by the iterative reweighted least squares with the preconditioned conjugate gradient method.The proposed method is compared with current popular methods,numerical experiments show that the proposed HCS-DMRI method can provide better reconstruction quality than the current offline model,and the reconstruction efficiency is greatly improved.Propose a multiple low rank plus sparse tensor model for the reconstruction of DMRI.We represent the image sequence as a linear sum of the low-rank component? and the sparse component .We minimizing the CP rank and Tucker rank for the low rank component jointly,which can exploit multi-dimensional data structures adequately.Minimizing the tensor total variation to constrait the sparse component.We reform the reconstruction model into two subproblems to iteratively caculate the ? and respectively.For the subsproblem of ?,the rank-one tensor updating and sum of nuclear norm minimization methods are uesed to solve it,primal dual method is used to caculate the.We compare the proposed method with state-of-the-art methods,experimental results show that the proposed method can achieve superior reconstruction quality than the comparison methods.