人工智能深度学习技术在辅助青光眼性眼底病变图像标注中的应用研究  

Application of artificial intelligence deep learning technology to assist annotation of glaucomatous fundus

作　　者：刘含若 白玮玲 张悦 杜一帆 王宁利 Liu Hanruo;Bai Weiling;Zhang Yue;Du Yifan(Wang Ningli Beijing Tongren Eye Cener,Beijing Tongren Hospital,Capital Medical University,Beijing Institute of Ophthalmology,Beijing Ophthalmology&Visual Sciences Key Lab.,Beijing 100730,China)

机构地区：[1]首都医科大学附属北京同仁医院,北京同仁眼科中心,北京市眼科研究所,北京市眼科学与视觉科学重点实验室,100730

出　　处：《中华眼科医学杂志（电子版）》2020年第4期234-238,共5页Chinese Journal of Ophthalmologic Medicine(Electronic Edition)

基　　金：国家自然科学基金项目(81700813);北京市医院管理局“青苗”计划专项经费项目(QML20180205);北京市科技新星项目(Z191100001119072);首都医科大学附属北京同仁医院拔尖人才培养计划;医药协同科研创新研究专项(Z181100001918035)。

摘　　要：目的探索人工智能(AI)深度学习技术用于青光眼性眼底病变(GON)筛查检出的效能。方法采用诊断试验的研究方法,收集自2013年8月至2019年7月于首都医科大学附属北京同仁医院北京同仁眼科中心青光眼门诊就诊患者的眼底图像200张(200只眼)。根据阅片方式的不同,将采用AI系统阅片的定义为AI系统组,将医师人工阅片的定义为医师组。其中,医师组包括高年资眼科医师(眼底照片诊断经验为10年以上)、低年资眼科医师(眼底照片诊断经验为5年以上,10年以内)及全科医师。各组均对所有眼底图像进行阅片和标注。年龄和性别的数据经S-W检验证实呈正态分布者,以均数±标准差表示。经Levene检验证实方差齐,AI系统与不同级别医师单张平均阅片时间的比较采用单因素方差分析,两两比较采用LSD-t检验。采用敏感度、特异度及工作特性曲线下面积(AUC)对AI模型进行预测和性能评价。诊断符合率以百分数表示,AI系统与眼科高级医师阅片诊断的符合率、AI系统与不同级别医师阅片诊断符合率的比较,均采用卡方检验。结果在医师组中,高年资眼科医师、低年资眼科医师及全科医师共6位医师共完成2次阅片。其诊断符合率的比较,差异均有统计学意义(χ2=4.324,3.562,4.213,5.786,10.546,11.431;P<0.05)。AI系统阅片测试的敏感度为100.0%,特异度为88.6%,诊断符合率为90.5%,AUC为0.934。在AI系统辅助下,标注的符合率提升最大的为2位全科医师,分别达到了87.3%及82.5%。在AI系统辅助下,高年资眼科医师、低年资眼科医师及全科医师的平均阅片时间较独立阅片标注有明显缩短,差异均有统计学意义(t=4.175、3.189、6.160;P<0.05)。结论标准化智能AI标注辅助系统的应用有利于不同资质医师的准确诊断和提高效率。尤其,可以快速提升全科医师及低年资眼科医师标注的符合率。Objective To explore the effectiveness of artificial intelligence(AI)deep learning technology in the screening and detecting glaucomatous optic neuropathy(GON).Methods Using diagnostic test research method,200 fundus images(200 eyes)were collected from the glaucoma clinic of Beijing Tongren Eye Center from August 2013 to July 2019.According to the different ways of image reading,they were divided into two groups:AI system group and doctor group.Among of them,the doctor group included senior ophthalmologists(with more than 10 years′experience in fundus image diagnosis),junior ophthalmologists(more than 5 years and less than 10 years)and general practitioners.All fundus images were read and labeled by AI system group and doctor group.The data of age and sex were normal distribution by S-W test,expressed as■.As the Levene test showed that the real variance was the same,the average image reading time of AI system was compared with that of doctors of different levels by one-way ANOVA,and the LSD-t test was used for pairwise comparison.Sensitivity,specificity and area under curve(AUC)were used to predict and evaluate the performance of AI model.The diagnostic coincidence rate was expressed as a percentage.Chi square test was used to compare the diagnostic coincidence rate between AI system and senior ophthalmologists,between AI system and doctors at different levels.Results In the doctor group,6 ophthalmologists(senior ophthalmologists,junior ophthalmologists and general practitioners)read images twice.The difference of diagnostic accuracy was statistically significant(χ2=4.324,3.562,4.213,5.786,10.546,11.431;P<0.05).The sensitivity,specificity,diagnostic coincidence rate and AUC of AI image reading test were 100.0%,88.6%,90.5%and 0.934,respectively.With the help of AI system,the accuracy of labeling was improved significantly in 2 general practitioners,which reached 87.3%and 82.5%,respectively.With the aid of AI system,the average reading time of senior ophthalmologists,junior ophthalmologists and general practit

关 键 词：人工智能 青光眼 眼底图像 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] R775[自动化与计算机技术—控制科学与工程]