关键词: 糖尿病视网膜病变/诊断, 眼底图像, 人工智能, 一致性

Abstract: Objective To compare the consistency of artificial intelligence (AI) diagnosis models and ophthalmologist in grading diabetic retinopathy (DR). Design Diagnosis tests. Participants 236 cases (468 eyes) of diabetic patients in Beijing Tongren Hospital were selected. Methods Two-field non-mydriatic macula- or optic disc-centered fundus photographs were performed. Using ophthalmologist diagnosis as the golden standard, three kinds of AI diagnosis model(model 1, 2, 3) based on Convolutional Neural Network (CNN) were evaluated, including DR yes/no, referable DR and DR stagingtests. Main Outcome Measures Sensitivity (Se), specificity (Sp), area under curve (AUC). Results The Se, Sp and AUC of DR yes/no were 95.8%, 90.1% and 0.930 on model 1, 96.5%, 85.2% and 0.908 on model 2, and 96.2%，83.5% and 0.917 on model 3, respectively. The Se, Sp and AUC of referable DR were 93.9%, 90.1% and 0.933 on model 1, 97.7%，89.3% and 0.935 on model 2, and 95.4%, 89.8% and 0.926 on model 3, respectively. The Se, Sp and AUC of stages of DR were 72.9%~90.1%, 93.9%~97.8% and 0.854~0.930 on model 1, 68.8%~92.1%， 90.6%~98.2% and 0.831~0.914 on model 2, and 75.0%~83.5%, 89.2%~96.8% and 0.849~0.917 on model 3, respectively. Conclusion Three kinds of AI diagnosis model have good sensitivity and specificity for the determination of DR and referable DR, and it could be used as an auxiliary tool for DR screening of diabetic patients. (Ophthalmol CHN, 2021, 30: 355-359)

Key words: diabetic retinopathy/diagnosis, fundus photography, artificial intelligence, consistency