基于生成对抗网络的肾小球病理图像合成  

Glomerular Pathological Image Synthesis Based on GenerativeAdversarial Network

作　　者：蔡健 孔祥勇[1] 吴滢[2] 尹梓名 王平[2] 卢严砖 彭瑞阳 孙晓晗 王钰泽 CAI Jian;KONG Xiang-yong;WU Ying;YING Zi-ming;WANG Ping;LU Yan-zhuan;PENG Rui-yang;SUN Xiao-han;WANGYu-ze(School of Health Science and Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China;Children's Hospital of Shanghai Jiaotong University,Shanghai 200040,China;School of Food Science,Shihezi University,Shihezi 832099,China)

机构地区：[1]上海理工大学健康科学与工程学院,上海200093 [2]上海交通大学附属儿童医院,上海200040 [3]石河子大学食品学院,新疆石河子832099

出　　处：《软件导刊》2022年第4期13-18,共6页Software Guide

基　　金：国家自然科学基金项目(82074581,81801797);国家重点研发计划项目(2019YFC2005801,2019YFC2005800)。

摘　　要：肾脏病理图像中肾小球的自动分割提取对于快速诊断肾脏疾病至关重要。针对高质量标注的肾小球病理图像数据不足的问题,提出基于pix2pixHD的肾小球病理图像合成方法。首先,对原始病理图像中肾小球及其对应掩膜进行提取,构建肾小球病理图像数据集;其次,以肾小球掩膜作为约束条件,利用生成对抗网络模型合成高质量肾小球病理图像;最后,将生成的肾小球图像数据集并入原始数据集,使用U-Net分割模型对原始肾脏病理图像进行再分割。实验结果表明,与常用几种图像生成算法比较,新建方法表现最佳,其IS±std为1.50±0.03,FID为33.13;在测试集相同的情况下,该方法的总体分割准确率提升了4%。The automatic segmentation and extraction of glomeruli in kidney pathological images is very important for the rapid diagnosis of kidney disease.Aiming at the problem of insufficient high-quality labeled glomerular pathological image data,a glomerular pathological image synthesis method based on pix2pixHD is proposed.First,extract the glomerulus and its corresponding mask in the original pathological image to construct a glomerular pathological image data set;secondly,use the glomerular mask as a constraint condition to synthesize a high-quality kidney using the constructed generative adversarial network model The glomerular pathological image;finally,the generated glomerular image data set is merged into the original data set,and the U-Net segmentation model is used to re-segment the original kidney pathological image.Experimental results show that comparing different generation algorithms,the method in this paper performs best,in which IS±std is 1.50±0.03 and FID is 33.13;in the case of the same test set,the overall segmentation accuracy of this method is improved to 4%.

关 键 词：生成对抗网络 U-Net网络 肾小球 图像分割 病理图像 

分 类 号：TP317.4[自动化与计算机技术—计算机软件与理论]