Compressed Sensing Mri Reconstruction Based On Deep Learning

Magnetic Resonance Imaging(MRI)technology has irreplaceable significance in the field of medical imaging due to its advantages of no contact and no radiation in scanning,high image resolution,multi-directional imaging and so on.However,its slow imaging speed is easy to cause discomfort to patients,and it greatly affects the utilization of its equipment,resulting in difficult for patients to seek medical treatment.Compressed Sensing(CS)can realize image reconstruction while ensuring the sparsity of magnetic resonance K space data,thus reducing the scanning time of MRI.In recent years,compressed sensing magnetic resonance imaging reconstruction(CS-MRI)based on deep learning has become the mainstream method to solve this problem.In this paper,two reconstruction algorithms are proposed based on generative adversarial network and Transformer to achieve high quality reconstruction of magnetic resonance images at lower sampling rates.The main contents of this paper include:(1)Aiming at the problems of complex reconstruction model and difficult training,a CS-MRI reconstruction algorithm based on multi-feature aggregation module generation adversarial network(MFAGAN)was proposed.Firstly,a dual-branch multi-feature aggregation module is employed in the generator to extract and fuse multi-feature information,optimize the calculation process,reduce the redundancy of the model,and enhance the feature information of the reconstructed image.Secondly,the edge loss function was utilized to participate in the training to highlight the different tissue information of the reconstructed image,enhance the contrast of the reconstructed image and reduce the residue of aliasing artifacts.Experiments show that MFAGAN can enhance the reconstruction quality of the image,and the PSNR can be improved by1.7%～2.8%.The number of model parameters is reduced to 30.6% of the optimal structure,which reduces the training difficulty of the network.(2)In order to further reduce the number of model parameters and improve the reconstruction speed,a CS-MRI reconstruction algorithm Trans CNN(Transformer and CNN feature extraction layer combined network)combining Transformer and CNN feature extraction is proposed.First of all,the modeling ability of Transformer for long-distance dependency relationship is used to effectively reduce the number of model parameters.Secondly,the proposed CNN feature extraction layer is used to provide reliable high-resolution feature information for Transformer encoding,and meanwhile,the complicated pre-training process of Transformer is saved and the computation of the model is greatly reduced.Compared with MFAGAN,PSNR is increased by 1.49% and the number of model parameters is decreased by 37%.