Research On Brain MRI Reconstruction Algorithm Based On Generative Adversarial Network

Magnetic Resonance Imaging(MRI)is one of the most important methods of clinical diagnosis.It has the characteristics of non-radiation,non-invasive to the human body and high resolution to soft tissue.However,one of the main problems of MRI is the long imaging time,and movement by the patient during the long imaging process can lead to aliasing artifacts,which can also cause anxiety,discomfort and other symptoms in patients.Therefore,it is very important to accelerate the reconstruction of MRI.In recent years,deep learning has developed rapidly in various fields,and medical image processing has been applied more and more widely.The method based on Convolutional Neural Network(CNN)has become the gold standard for medical image reconstruction.However,due to the limited ability of CNN to capture long-distance information,there may be aliasing in the structure of reconstructed images.In addition,the unreal texture and aliasing artifacts in the reconstructed results make its performance difficult to meet the requirements of clinical application.To solve these problems,this paper took brain MRI as the research object,designed novel brain MRI reconstruction algorithms,and verified the reconstruction performance of the model in the open-source dataset.The main research work of this paper is as follows:(1)In view of the limited receptive field of the reconstruction model based on CNN and the problem of inadequate representation of the detailed features,a new dual-domain generative adversarial network based on Swin Transformer is proposed for brain MR image reconstruction,namely SwinGAN.Firstly,the Swin Transformer module is introduced into the generator to replace the original CNN network.And the long-range dependencies in brain MR images can be captured by using the shifted window attention mechanism of Swin Transformer,thus enhancing the global feature information of the network.Secondly,the contextual image relative position encoder(ci RPE)in Transformer is designed to improve the ability to capture contextual information.In addition,in order to make full use of the original MRI data,the frequency domain generator and image domain generator are designed to process the frequency domain information and image domain information in the data respectively.The complex information in under-sampled k-space is used for reconstruction,which can more fully capture the feature information in different domains.At the same time,the loss function of the network is redesigned and frequency domain loss is introduced to enable the network to capture more image information.The experimental results show that the SwinGAN can effectively enlarge the receptive field of the network,thus improving the reconstruction performance.(2)To solve the problem that the traditional reconstruction network has poor texture generation ability for high-resolution images,a new MRI image reconstruction algorithm named Diff GAN is proposed by combining diffusion model and GAN network.First,a Ushaped network LVTN based on Local Vision Transformer is introduced into the generator,which utilizes LVT modules and local self-attention to capture high quality local features and image details.Secondly,GAN network is used to accelerate the generation of diffusion model in the process of reverse diffusion because of the instability caused by gradient vanishing in the training process of GAN.Finally,a new hybrid loss is designed to further improve the ability of the model to extract image details.The experimental results show that Diff GAN model alleviates the problem of network gradient vanishing,stabilizes the network training process and improves the network reconstruction performance.