Medical Image Super-resolution Reconstruction Based On Generative Adversarial Network

The medical images that have been obtained may be blurred and ghosted due to various reasons,which may affect the doctor’s judgment of the correct condition of the patient.Medical image super-resolution reconstruction uses digital processing methods to process the medical images that have been obtained,super-reconstructing high-resolution images from low-resolution images,making the images clearer.Due to the special of medical images,it is hard to obtain them again,so the super-resolution reconstruction of medical images is very meaningful.This thesis mainly studies the application of Generative AdversarialNetworks in the field of medical image super-resolution reconstruction,and mainly completes the following tasks:By analyzing the basic model and algorithm core of the Generative AdversarialNetworks,it can be known that the Generative AdversarialNetworks has a strong generation ability,but the traditional Generative AdversarialNetworks has the problems of gradient disappearance and unstable training.So,this thesis designs a network model SRWGAN that can be used for super-resolution reconstruction of medical images.On the basis of the SRGAN network model,the network’s adversarial loss function is replaced with Wasserstein distance as the theoretical basis,which effectively stabilizes the gradient of the generated adversarial network training.The network model uses 16 residual blocks as the main part of the generator,and the multi-layer convolutional neural network is the discriminator.At the same time,content loss,perceptual loss and total variation loss are introduced as the loss function.This thesis also designs a network model DNGAN that the generator combines the ResNet structure and the DenseNet structure.In this network,the WGAN-GP theory is used as the adversarial loss to stably the training of the Generative AdversarialNetworks.In addition,the content loss function and the perceptual loss function are also used as the loss function of the network.To make better use of the rich frequency domain information of the MRI,the frequency domain information of the MRI is added to the network as a frequency domain loss function.In order to verify the effectiveness of the SRWGAN model and DNGAN model proposed in this thesis,a large number of experiments were carried out on the knee joint MRI dataset provided by Southwest Hospital.Experiments are conducted on whether to introduce the total variation loss function in the SRWGAN model.Experiments are conducted on whether the SRWGAN model uses Wasserstein distance as the adversarial loss function.Finally,the experimental results of the SRWGAN model are compared with the experimental results of the bicubic interpolation method.The experimental results show that the SRWGAN model introduces the total variation loss function and the super-resolution generated image has the best effect,the PSNR is 4.24 dB higher than the bicubic interpolation method,and the SSIM index is 0.106 higher than the bicubic interpolation method.Experiments are conducted on whether the DNGAN model is effective.At the same time,the DNGAN model is compared with the bicubic interpolation method and the SRGAN model.The experimental results show that the DNGAN model introduces the frequency domain loss function to generate the best image.The PSNR is 4.184 dB higher than the worst effect of the bicubic interpolation method,and the SSIM is 0.0901 higher than the worst effect of the bicubic interpolation method.