Research On MR Image Reconstruction Method Using Wasserstein Generative Adversarial Network

Magnetic resonance imaging(MRI)is a widely used medical image technology and can provide non-invasive diagnostic imaging of the tissue structures in the human body.For MR image reconstruction,the k-space dates in the time domain are acquired firstly and inverse fast Fourier transformation(FFT)is used to implement the image reconstruction,so as to generate real-space images in the frequency domain.Moreover,MRI does not involve exposure to ionizing radiation,and thus avoids the associated carcinogenic risk.However,the scanning time to acquire high-resolution images can be too long for some patients to endure the discomfort of keeping the same posture.In addition,imaging quality is susceptible to physiological movements and motion artifacts.In the traditional magnetic resonance image reconstruction algorithm,Compressed Sensing is usually used to recover the under sampled magnetic resonance data,but the reconstruction effect needs to be improved at low sampling rate.With the development of Deep Learning technology,neural network-based magnetic resonance image reconstruction technology has become the hot topic of rapid magnetic resonance imaging research.This paper focuses on reconstructing magnetic resonance images by using the Wasserstein Generative Adversarial Network(WGAN)based deep learning method.The specific research contents are as follows:(1)A magnetic resonance image reconstruction algorithm based on De-aliasing Fine-tuningWasserstein Generative Adversarial Network(DA-FWGAN)is proposed.In the original GAN method,the discriminator with better training performances,the generator gradient disappears more serious,and the network training becomes unstable.WGAN,as a variant of GAN,can effectively avoid the problem of gradient disappearance and improve training stability by using its Wasserstein distance.At the same time,fine-tuning methods are used to accurately train neural network parameters.In addition,in order to better reconstruct the texture structure in the image,we add perceptual loss,image loss and frequency loss to the loss function for network training.Compared with other deep learning methods,the experimental results show that the DA-FWGAN-based magnetic resonance image reconstruction algorithm has improved image quality and stability.(2)De-aliasing Wasserstein Generative Adversarial Network with Gradient Penalty(DAWGAN-GP)method is proposed to implement MR image reconstruction.As an alternative to traditional generative adversarial network,Wasserstein is used to generative adversarial network.In order to overcome the slow convergence of WGAN based MR image reconstruction,the gradient penalty method is proposed to improve the training speed and model stability.The experimental results show that,compared with several GAN-based magnetic resonance image reconstruction algorithms,the DAWGAN-GP based magnetic resonance image reconstruction method can reconstruct better organizational structure and local details.The analysis of the experimental results shows that,compared with the compressed sensing MRI reconstruction and GAN reconstruction algorithm,the WGAN based image reconstruction algorithm can improve the MRI reconstruction quality and model stability.