Objective To evaluate the performance of an artificial intelligence ( AI ) assisted diagnosis system for diabetic retinopathy ( DR) based on deep learning theory. Methods Diagnostic performance of a robot assisted diagnosis system called SongYue for DR was trained by using 25297 retinal images tagged by fundus doctors from multiple hospitals in China. Four types of DR detection model consisting of abnormal DR,referable DR,severe non-proliferative and proliferative DR as well as proliferative DR according to fundus leisions identification were established. The ability of the system to distinguish DR was determined by using receiver operator characteristic (ROC) analysis,sensitivity and specificity of the system. Results SongYue system achieved an area under the ROC curve ( AUC) of 0. 920 for successfully distinguishing normal images from those DR with a sensitivity of 96. 0％at a specificity of 87. 9％. The AUC of SongYue for referable DR was 0. 925,sensitivity was 90. 4％,and specificity was 95. 2％. For severe non-proliferative and proliferative DR,AUC was 0. 845,sensitivity was 72. 7％,and specificity was 96. 2％. For proliferative DR, AUC was 0. 855, sensitivity was 73. 5％, and specificity was 97. 3％. Conclusions SongYue robot assisted diagnosis system has high AUC,sensitivity and specificity for identifying DR, showing good clinical applicable benefits.%目的 建立和评估基于深度学习理论的糖尿病视网膜病变(DR)人工智能(AI)机器人辅助诊断系统.方法 联合包括北京协和医院等国内8家医院眼底专业医师对25297张糖尿病患者眼底图像病变进行标记,训练和建立一个深度学习框架同时识别DR病变的AI机器人辅助诊断系统,即"嵩岳"机器人系统.依据DR分级和标识眼底病变,构建是否罹患DR、是否需要转诊DR、是否重度非增生性和增生性DR以及是否增生性DR 4种模型,建立一个基于病变识别技术的DR筛选系统,应用受试者工作特征曲线(ROC曲线)、敏感度、特异度进行AI诊断性能的数据分析.结果 "嵩岳"系统中是否罹患DR模型的敏感度为96.0％,特异度为87.9％,曲线下面积(AUC)为0.920;需要转诊DR模型的敏感度为90.4％,特异度为95.2％,AUC为0.925;是否为重度非增生性和增生性DR模型的敏感度为72.7％,特异度为96.2％,AUC为0.845;增生性DR模型的敏感度为73.5％,特异度为97.3％,AUC为0.855.结论 "嵩岳"AI机器人辅助诊断系统具有精确高效的DR诊断性能,具有良好的临床应用价值.
展开▼
