基于注意力机制的深度学习网络模型在糖尿病视网膜病变筛查中的价值评估  

Evaluation of deep learning network model based on attention mechanism in diabetic retinopathy screening in China

作　　者：杜紫薇 柳江 胡浩 杨楠 温良 王凤华 王蓓[5] 袁扬[1] 孙子林[1] Du Ziwei;Liu Jiang;Hu Hao;Yang Nan;Wen Liang;Wang Fenghua;Wang Bei;Yuan Yang;Sun Zilin(Department of Ophthalmology,Zhongda Hospital,School of Medicine,Southeast University,Nanjing 210009,China;Fushun Eye Hospital,Fushun 113006,China;Beijing Tongren Eye Center,Beijing Tongren Hospital,Capital Medical University,Beijing Ophthalmology&Visual Science Key Lab,Beijing 100176,China;Department of Epidemiology and Statistics,School of Public Health,Southeast University,Nanjing 210009,China;Department of Endocrinology,Zhongda Hospital,Institute of Diabetes,School of Medicine,Southeast University,Nanjing 210009,China)

机构地区：[1]东南大学附属中大医院内分泌科东南大学糖尿病研究所,南京210009 [2]东南大学附属中大医院眼科,南京210009 [3]辽宁省抚顺市眼病医院眼底病科,113006 [4]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学与视觉科学重点实验室,100176 [5]东南大学公共卫生学院流行病与卫生统计学系,南京210009

出　　处：《中华糖尿病杂志》2021年第12期1148-1154,共7页CHINESE JOURNAL OF DIABETES MELLITUS

基　　金：国家重点研发计划(2016YFC1305700)。

摘　　要：目的评估基于注意力机制的深度学习网络模型——"慧眼糖网"模型在自然人群和糖尿病人群中筛查糖尿病视网膜病变(DR)的效能, 以及分别应用单方位及两方位眼底照相的筛查效能。方法为横断面调查。自2016年12月至2017年6月, 采取分层多阶段整群抽样的方法从全国8省共10个地区6个不同民族的18至70岁常住居民中选择代表性样本作为研究对象, 共纳入8 948名参与者的17 118张眼底图像至"慧眼糖网"系统进行评分。以DR早期治疗研究(ETDRS)评分系统作为诊断DR的金标准对眼底图像进行分级。以"需转诊的DR(ETDRS>31)"作为参考变量, 绘制受试者工作特征(ROC)曲线, 评价"慧眼糖网"的曲线下面积(AUC)、灵敏度、特异度, 确定该系统的筛查效能。结果基于每位受试者, 在自然人群中, 使用"慧眼糖网"单方位眼底照相筛查"需转诊的DR"的AUC为0.941, 灵敏度和特异度分别为98.15%、90.08%;两方位眼底照相筛查灵敏度为100%, 特异度为86.91%。在糖尿病人群中, "慧眼糖网"单方位眼底照相筛查"需转诊的DR"时, AUC、灵敏度和特异度分别为0.901、98.08%和82.10%。结论在自然人群和糖尿病人群中, "慧眼糖网"系统均表现出较高的灵敏度和特异度, 可以作为DR筛查的一种辅助手段。Objective To evaluate the diagnostic accuracy of the SG-DR screening system(the ophthalmic image intelligent recognition software designed and produced by Zhuhai Shang Gong Co.,China.Guangdong Registration Certificate for Medical Device No.20172700901)in detecting diabetic retinopathy(DR)in the general and diabetes population,using single-or double-field fundus photography.Methods The cross-sectional database was used to analyze among a randomized cluster stratification sample of 18-70 years old residents in six different ethnic groups in eight provinces of China from December 2016 to June 2017.For 8948 enrolled participants,two retinal fundi(one disc centered 45°and one macula centered 45°)images per eye were collected for the evaluation of DR.All 17118 images were graded by the Early Treatment Diabetic Retinopathy Study(ETDRS)scale(gold standard for DR grading),and the SG-DR screening system.Area under the receiver operating characteristic curve(AUC),sensitivity and specificity of the SG-DR screening system were performed by the receiver operating characteristic(ROC)curve based on the DR in need of referral(ETDRS>31).Results In the general population,the AUC for screening DR in need of referral using SG-DR single-field fundus photography screening system was 0.941 with sensitivity of 98.15%and specificity of 90.08%respectively.Using SG-DR double-field fundus photography screening,the sensitivity and specificity were 100%and 86.91%,respectively.In the diabetes group,for DR in need of referral,the AUC,sensitivity and specificity were 0.901,98.08%,and 82.10%,respectively.Conclusions The SG-DR screening system has shown high sensitivity and specificity in detecting DR both in the general population and diabetic group and could be as an adjunct in the screening of DR.

关 键 词：糖尿病视网膜病变 人工智能 筛查 

分 类 号：R587.2[医药卫生—内分泌] R774.1[医药卫生—内科学]