Fast CS-MRI Reconstruction Based On Deep Residual Generative Adversarial Network

During the process of Magnetic Resonance Imaging(MRI),due to the influence of its own imaging mechanism and following the Nyquist sampling theorem,it will consume a lot of acquisition time.The prolonged scanning operation will easily cause discomfort to the patient and may introduce motion artifacts.In recent years,compressive sensing(CS)MRI reconstruction algorithms based on generative countermeasure networks have greatly improved the reconstruction time,but due to the original generative countermeasure networks,there are poor training stability,difficulty,pattern collapse,and feature extraction insufficient,so the reconstruction quality is not significantly improved compared to non-deep learning reconstruction methods.To solve the above problems,this article proposes a fast CS-MRI reconstruction model based on the least squares adversarial loss based on the deep residual Generation Adversarial Network(GAN).The main content of this article includes:(1)The design of the generative model of the generative adversarial network: A U-shaped network model is designed to form a generative model combined with the removal of residual blocks from batch standardized operations.First,the proposed generative model refers to the idea of the medical image segmentation U-net model,The convolutional layer and the deconvolutional layer are symmetrically connected by jumping connections to form a U-shaped full convolution model structure to solve the traditional full convolutional neural network information transmission loss and loss;secondly,it will be removed The residual blocks of the batch normalization operation are added to the U-shaped structure,which increases the depth of the generated model and improves the quality of the reconstructed image;and the reconstruction time of the image does not increase much.(2)Introduce least squares confrontation loss: Introduce least squares confrontation loss to replace the original GAN’s cross-entropy confrontation loss.Using least squares confrontation loss can solve the problems of poor training stability of the original GAN,poor quality of generated images,and mode collapse.(3)MRI T2 image reconstruction network migration model with lesion information: After two model migrations,the MRI T2 image reconstruction model with lesion information is obtained.Specifically,(a)a large number of MRI T1 images without lesion information are reconstructed Pre-training;(b)Model migration of the feature extraction part of the pre-training model,and fine-tuning with a small number of MRI T1 images with lesion information;(c)Using a small number of MRI T2 images with lesion information for micro-reconstruction of the MRI T2 model test.The MRI T1 normal brain data set in the Medical Image Computing and Computer Assisted Intervention Society(MICCAI)2013 competition and the MRI T1 head data in the IXI data set of Imperial College London are used to verify the MRI T1 reconstruction algorithm without lesion information in this paper.The experimental results show that the reconstruction algorithm in this article has a greater improvement in both objective evaluation indicators and subjective visual effects than the GAN reconstruction algorithm based on pixels,frequency domain and perceptual loss(PFPGR).The MRI T2 lesion image reconstruction network migration model was verified by using the MICCAI 2018 competition MRI T1 and MRI T2 brain data sets with lesion information.The experimental results show that under the same data size,the reconstructed image quality of the MRI T2 image reconstruction model with lesion information obtained through two model transfers is higher than that of the reconstructed image of the one-transfer reconstruction model.