对抗学习下的眼底图像视盘视杯分割算法  

OPTIC DISC AND CUP SEGMENTATION ALGORITHM FOR FUNDUS IMAGES UNDER ADVERSARIAL LEARNING

作　　者：王彩云 玄祖兴[2] 周建平 胡晰远[4] 程钢炜[5] 宋禄琴 Wang Caiyun;Xuan Zuxing;Zhou Jianping;Hu Xiyuan;Cheng Gangwei;Song Luqin(Beijing Key Laboratory of Information Service Engineering,Beijing Union University,Beijing 100101,China;Institute of Fundamental and Interdisciplinary Sciences,Beijing Union University,Beijing 100101,China;School of Computer Science and Technology,Anhui University of Technology,Maanshan 243032,Anhui,China;School of Computer Science and Engineering,Nanjing University of Science and Technology,Nanjing 210094,Jiangsu,China;Peking Union Medical College Hospital,Beijing 100005,China)

机构地区：[1]北京联合大学北京市信息服务工程重点实验室,北京100101 [2]北京联合大学基础与交叉科学研究所,北京100101 [3]安徽工业大学计算机科学与技术学院,安徽马鞍山243032 [4]南京理工大学计算机科学与工程学院,江苏南京210094 [5]北京协和医院眼科中心,北京100005

出　　处：《计算机应用与软件》2024年第2期229-237,共9页Computer Applications and Software

基　　金：北京市属高校高水平教师队伍建设支持计划项目(IDHT20170511);北京联合大学人才强校优选计划(BPHR2020EZ01);北京联合大学研究生科研创新资助项目(YZ2020K001)。

摘　　要：由于不同数据集的质量、拍摄条件和采集状态的差异,导致模型图像分割效果参差不齐。针对这种情况,提出对抗学习下的眼底图像视盘视杯分割算法,从不同眼底图像数据集中分割视盘(OD)和视杯(OC),在生成对抗网络的基础上,改进生成器网络,加入密集连接块,使网络在更低计算成本、更短训练时间的情况下,获得更优的性能,提高了模型在不同数据集中的泛化能力。实验结果表明,在REFUGE数据集中验证了该算法在分割性能方面的稳定性,同时将算法推广到无须进一步训练就能测试来自不同设备的眼底数据集中均取得了较好的效果。The differences in quality,shooting conditions and acquisition status of different datasets lead to uneven results in model image segmentation.To address this situation,the fundus image optic disc and cup segmentation algorithm under adversarial learning is proposed to segment optic disc(OD)and optic cup(OC)from different fundus image datasets.It improved the generator network based on generating adversarial network by adding densely connected blocks,so that the network could obtain better performance with lower computational cost and shorter training time,and improve the generalization ability of the model in different datasets.Experimental results show that the stability of the algorithm in terms of segmentation performance is verified in REFUGE dataset,and the extension of the algorithm to test fundus datasets from different devices without further training achieves better results.

关 键 词：深度学习 图像分割 生成对抗网络(GAN) 眼底图像 青光眼诊断 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]