Fast Imaging Method Based On Magnetic Resonance Physical Model And Deep Learning

Magnetic resonance imaging（MRI）with non-invasive,non-radiation and high soft tissue contrast,is a kind of imaging technology widely used in clinical diagnosis and treatment.However,its inherent and slow imaging process limits its wider application.A common method to realize fast MRI is to sample magnetic resonance（MR）data with a high undersampling rate,and then reconstruct the high-quality image by image processing technology.Recently,deep learning has achieved inspired results in MR image reconstruction tasks.However,most reconstruction methods based on deep learning are directly training a deep neural network to fit the nonlinear mapping function from undersampled data to fully-sampled data,without considering the use of MR physical model for image reconstruction.Therefore,this paper studies the fast imaging method based on MR physical model and deep learning,and mainly completes the following projects:1.In this paper,a reconstruction method combining generalized alternating projection（GAP）algorithm with deep learning is proposed.The combination of the algorithm and deep learning is realized by replacing the nonlinear image sparse transform domain with convolutional neural network（CNN）and replacing the iterative step and shrinkage-thresholding with learnable parameters.Validation is performed on an internal MR dataset,the experimental results show that the proposed method can achieve better image reconstruction quality than the comparison methods.2.In this paper,a MR image reconstruction framework based on self-supervised learning is proposed utilizing the properties of MR data.The ability of network learning to recover frequency information is promoted by recovering undersampled k-space data from the subsets of the data.The correctness of information recovery at the unscanned frequency points is indirectly constrained by the construction of parallel network framework and the definition of the difference loss term.Experimental results on an open source dataset show that the proposed self-supervised reconstruction method can achieve the same reconstruction performance as the corresponding reconstruction method based on supervised learning.