Research On Brain MRI Image Generation Method Based On GANs

With the continuous development of Internet technology,people's lives and solutions to problems are constantly changing.With a large number of research scholars' in-depth research on artificial intelligence algorithms,artificial intelligence algorithms are applied to various fields,such as: Image processing,speech recognition,natural language processing,etc.,and achieved a series of remarkable results.In recent years,artificial intelligence technology has begun to be applied in the medical field,and some good results have been obtained in text processing.However,in terms of medical images,medical image data is not easy to obtain due to the low prevalence of rare diseases and the personal information of patients,which correspondingly leads to the artificial intelligence algorithm can not get better results in the field.Among the methods for solving image scarcity problems,mainly include traditional image data enhancement,Variational AutoEncoders(VAE),and Generative Adversarial Networks(GANs).The common traditional image data enhancement method can increase the number of image samples to a certain extent,but the large-scale generation of image samples will cause over-fitting risk;the variable-point automatic encoder VAE method solves the over-provision caused by a single generated image.The problem is solved,but because the VAE method directly determines the quality of the generated image by directly calculating the mean square error between the image and the source image,resulting in a blurred image and the next medical image research;The generated confrontation network GANs,which generate clear and usable medical images through continuous confrontation learning between the generator and the discriminator,the learning ability of the generator and the discriminator in the network directly affects the quality of the generated image.However,the generator structure of the original GANs consists only of a simple convolutional network,and it is not possible to quickly generate clear images;and because the raw data is scarce,it is impossible to provide enough samples for the discriminator to train.In view of the above problems,this paper also proposes a new generator network structure-NedNet,and applies it to the generator part of WGANs to solve the problem that the original generator can not quickly generate clear images.Based on the advantages of traditional image data enhancement,this paper uses traditional image data to enhance the real data set to train the discriminator,which is used to solve the problem of too few discriminator training samples.The main work of this paper is summarized as follows:1、First,the paper introduces the traditional image data enhancement,residual network ResNet and the related basic knowledge of Generative Adversarial Networks GANs,and deduces the specific process of GANs generator and discriminator against learning.2、In this paper,the innovation of the Generative Adversarial Networks GANs is in two aspects,and a new model based on the traditional image data enhancement method and the improved WGANs——AUG-RGANs is proposed.Among them,firstly,based on the advantages of the traditional image data enhancement method,this paper uses it to process the real data set as the input of the new model discriminator to train the discriminator's resolving power.Then the NedNet network structure proposed in this paper is used as the generator network of WGANs to improve the quality of medical image generation.Finally,the trained discriminator competes with the generator to identify the forged image of the generator,and continues to generate a clear and usable medical image sample through continuous combat learning.3、In order to verify the image generation ability of the new model AUG-RGANs,this paper designs an image quality comparison test.By comparing the experimental results,the scores of the images generated by the proposed model in MSE,PSNR and SSIM are better than those of the traditional image enhancement method and WGANs against the network model,which proves that the proposed method can generate better Images.4.In order to verify that the image generated by the new model AUG-RGANs can be applied in the medical field,the proposed model is applied to the field of Alzheimer's image generation and classification.It is proved that the medical image generated by the new model is better and with the new model training the accuracy of the classification model is higher,which proves the feasibility of applying the new model to generate medical images.