【摘要】 提出一种带出血病症的眼底图像生成方法，该方法可以丰富眼底图像样本，提升眼底出血检测系统的准确率。该方法用图像分割技术从现有图像中分割出血管树和出血块，利用GAN生成大量血管树和出血块，并经过预处理合并，把合并后的图片和真实眼底图片一起输入到改进的CycleGAN中，生成大量眼底图片。其中对CycleGAN进行改进：改进模型结构，引入Wassertein距离，并加入同一映射损失和感知损失。实验表明，用该方法生成图像的PSNR值比现有技术提高9.82%,SSIM值提高4.17%且收敛速度更快。把生成图像添加到出血检测系统的训练集中，系统的AUC值提升3.51%,证明该方法优于现有技术。更多还原

【Abstract】 This paper proposes a method for generating fundus images with bleeding disorders, which can enrich the fundus image samples and improve the accuracy of the fundus bleeding detection system. We used the image segmentation technique to segment the vascular tree and bleeding blocks from the existing picture. We adopted GAN to generate a large number of vascular trees and bleeding blocks, and then preprocessed and merged them. The merged picture and the real fundus picture were input to the improved CycleGAN and a large number of fundus pictures were generated. We improved CycleGAN. The model structure was improved, the Wassertein distance was introduced, and the same mapping loss and perceptual loss were added. The results show that the PSNR value of the image generated by the proposed method increases by 9.82%, the SSIM value increases by 4.17%, and the convergence speed is faster. Adding the generated image to the training set of the bleeding detection system, the AUC value of the system increases by 3.51%, which proves that this method is superior to the current ones.更多还原