Research On Medical Image Generation Method Based On GAN

Tumor is a major disease facing human beings,and its morbidity and mortality are increasing year by year.How to achieve personalized and precise diagnosis of tumors has become an important issue in medical research.Computer-aided diagnosis(CAD)combines the characteristics of two fields of research,medical research and computer vision,and is widely used in research tasks such as classification,segmentation and registration of medical images.However,it has been found that highquality training data are often difficult to obtain due to privacy protection and collection costs,limiting the performance of the assisted diagnosis models.Generative Adversarial Network(GAN)-based image generation methods provide possible solutions for generating training images in medical image research,where image generation quality and application methods of generated images are the key issues to be studied.This paper focuses on GAN-based medical image generation methods with the aim of solving the problem of sparse or unbalanced data in medical image datasets,and the main work of this paper includes:1)To address the problems of pattern collapse and gradient disappearance in GAN model training,this paper proposes Multi-GAN,a generative adversarial model fusing multi-scale features,which can learn structural features at low scales and texture features at high scales of medical images by improving the generator structure and loss function of Sin GAN,effectively improving the quality of generated images.The experimental results show that the malignant mammogram images generated by MultiGAN can effectively compensate for the imbalance of benign and malignant data in the INbreast dataset and improve the prediction performance of the classification model,achieving an AUC value of 0.92 and a classification accuracy of 85.5%,which improves the AUC value by 0.15 compared with the traditional data augmentation method.2)To address the structural inconsistency between the generated and real images in the GAN model,this paper proposes a generative adversarial model FP-GAN incorporating regional features.The cycle consistency mechanism of this model ensures the structural consistency between the generated and real images,and,by adding a priori regional features to the model to assist the generator learning,it can generate high-quality medical images and the corresponding region of interest images(ROI).The experimental results show that the method effectively improves the quality of the generated images and achieves better results in several image evaluation metrics such as peak signal-to-noise ratio(PSNR),mean square error(MSE),structural similarity(SSIM)and Hausdorff distance compared with other methods.Validation experiments on the BRATS 2017 dataset show that the images generated by this method can effectively assist the training of machine learning models,improve the generalization performance of prediction models,and improve the classification accuracy by 4% compared to the images generated by the classical model cycle GAN.The results of this work show that the GAN-based image generation method can effectively expand the data volume of medical image datasets,alleviate the problems of small data size,unbalanced datasets and poor data quality in medical research to a certain extent,and improve the performance of medical image models.