目的：评估基于深度学习的糖尿病视网膜病变(diabetic retinopathy，DR)筛查智能诊断系统的应用价值。

方法：收集2017-01/06在我院就诊的糖尿病患者186例372眼，比较专家诊断及基于深度学习的人工智能诊断的应用情况，并比较其特异性和敏感性。

结果：专家诊断组显示42眼(11.3%)为无DR，330眼(88.7%)患有不同程度DR； 其中轻度非增殖型糖尿病视网膜病变(non-proliferative diabetic retinopathy，NPDR)者62眼(16.7%)，中度NPDR者55眼(14.8%)，重度NPDR者155眼(41.7%)，PDR者58眼(15.6%)。而智能诊断结果显示38眼(10.2%)为无DR，44眼为PDR(11.8%)，其他为不同分期NPDR。智能诊断系统与专家诊断结果DR一致性分析结果显示，高度一致性为309眼(83.1%)，Kappa值为0.78。智能诊断灵敏度为0.82，特异性为0.91，Kappa为0.77(χ²=20.39，P<0.05)。

结论：基于深度学习的DR人工智能诊断系统能较好显示眼底病变的严重程度，有望为DR提供一种新的筛查工具。

METHODS:A total of 186 patients(372 eyes)with diabetes were recruited from January to July 2017. Discrepancies between manual grades and artificial intelligence results were sent to a reading center for arbitration. The sensitivity and specificity in the detection of DR were determined by comparison with artificial intelligence diagnostic system and experts human grading.

RESULTS:Based on manual grades, the results as follows: non DR(NDR)in 42 eyes(11.3%), 330 eyes(88.7%)in different stages of DR. Among 330 DR eyes, there were mild non proliferative DR(NPDR)in 62 eyes(16.7%), moderate NPDR in 55 eyes(14.8%), severe NPDR in 155 eyes(41.7%), and proliferative DR(PDR)in 58 eyes(15.6%). Based on artificial intelligence diagnostic system, the results were as follows: NDR in 38 eyes(10.2%), PDR in 44 eyes(11.8%), others were NPDR. The sensitivity and specificity of artificial intelligence diagnostic system, compared with human expert grading, for the detection of any DR were 0.82 and 0.91, and the kappa value was 0.77(χ²=20.39, P<0.05).

CONCLUSION: This study shows that a deep learning-assisted diagnostic system with an artificial intelligence for grading diabetic retinopathy is a reliable alternative to diabetic retinopathy assessment, thus the use of this system may be a valuable tool in evaluating the DR.