Research On Undersampled MR Image Reconstruction

Magnetic Resonance Imaging(MRI)is a reproducible and quantitative tissue measurement method,which plays an important role in the diagnosis and treatment of major diseases.However,MRI is acquired sequentially in K-space,and the acquisition speed is limited by hardware and physiology.Therefore,a significant drawback for MRI is the long acquisition time.This defect will not only bring discomfort to the patients,but also the low inspection efficiency will bring expensive inspection costs,which limits its further promotion.An effective solution is compressed sensing,which can use a small amount of raw data to reconstruct images.However,the existing reconstruction methods generally have the problems of unclear details,slow convergence and long reconstruction time,which are not suitable for clinical application.In addition,these methods may suffer from the hallucination of details and artifact when reconstructing highly undersampled MR image,which usually not be able to reconstruct the fine textures and structural details of images faithfully.In order to solve the above problems,this article adopts generative adversarial network and self-attention to conduct in-depth research.The main work are listed as follows:(1)A novel MR image reconstruction method based on residual self-attention generative adversarial network is proposed.Aiming at the problem of unclear image details,slow convergence and long reconstruction time of existing MR image reconstruction methods,in this paper,a novel crucial feature enhancement module is designed and embedded at the bottom of the U-NET compression path in the generator,which uses the correlation between pixels in the image and the average information of high-level features.Then,the undersampled MR image is input into the generator,and the reconstructed image is obtained by confrontation training between the generator and the discriminator.This design can effectively constrain the network to optimize in the right direction,accelerate the model convergence and improve the reconstruction speed while enhancing the image fine details.The experimental results show that the details of the reconstructed image obtained by residual self-attention generative adversarial network are more clear and coherent.In addition,it is superior to other methods in terms of convergence speed and reconstruction time.(2)An effective MR image reconstruction method based on multi-scale fusion generative adversarial network is proposed.Aiming at the problem that the real textures and details of the images can not be restored effectively when the existing methods reconstruct the highly undersampled MR image with acceleration factor of 5 times or more,a multi-scale feature optimization decoder and a multi-scale fusion unit are constructed in generator.They can adaptively integrate the crucial features and context information of different scales,so as to improve the accuracy of network inference.Comprehensive comparison studies show that the performance of the proposed method is better than the existing methods,and it can reconstruct the complex textures and structural details while retaining 20% K-space raw data.(3)An undersampled MR image reconstruction system is designed and implemented.This paper adopts PHP,Mysql and Python to engineer the MR image reconstruction algorithm and design the MR image reconstruction system.It is divided into user administration module,MR image undersampling module,MR image reconstruction module and MR image reconstruction quality evaluation module to realize the undersampled MR image reconstruction.