摘要目的:比较基于彩色眼底像阅片的人工智能（AI）系统分别在社区和医院筛查和（或）诊断糖尿病视网膜病变（DR）的效率和差异，初步评价其应用价值。方法:回顾性和前瞻性相结合研究。回顾性收集2018年7月至2021年3月于河南省眼科研究所连续就诊的老年糖尿病患者1 608例的临床资料。其中，男性659例，女性949例；年龄中位数64岁。前瞻性收集2018年12月至2019年4月以社区为来源主动招募的老年糖尿病患者496例的临床资料。其中，男性202例，女性294例；年龄中位数62岁。由眼科或经培训的内分泌科医生对患者行双眼免散瞳眼底彩色照相检查，拍摄以黄斑中心凹为中心后极部45°正位片。AI系统基于深度学习YOLO源码开发，采用"AI+人工复核"方式最终确定DR诊断并分为0~ 4期，其中1期为无需转诊DR，2~ 4期为需转诊DR。结果:AI总读片1 989例（94.5%，1 989/2 104 ），其中社区、医院来源患者分别为437 （88.1%，437/496）、1 552 （96.5%，1 552/1 608）例。社区来源AI读片率低于医院来源，差异有统计学意义（ χ2=51.612， P<0.001 ）。社区图像质量差的主要原因为小瞳孔（47.1%，24/51）、白内障（19.6%，10/51）、白内障合并小瞳孔（21.6%，11/51）。AI诊断DR阴性62.4% （1 241/1 989 ）；其中，社区、医院来源分别为84.2%、56.3%，社区来源AI诊断DR阴性率高于医院，差异有统计学意义（ χ2=113.108， P<0.001）。AI诊断需转诊DR 20.2% （401/1 989 ）。其中，社区、医院来源分别为6.4%、24.0%，社区来源AI诊断需转诊DR率低于医院，差异均有统计学意义（ χ2=65.655， P<0.001）。不同来源AI诊断DR不同分期患者构成比比较，差异有统计学意义（ χ2=13.435， P=0.001）。其中，社区来源患者以无需转诊DR为主（52.2%，36/69）；医院来源患者以需转诊DR为主（54.9%，373/679 ），且已治疗DR检出率更高（14.3%）。AI自动识别眼底病灶数顺位中，社区、医院来源排首位的分别为玻璃膜疣（68.4%）和视网膜内出血（48.5%）。 结论:AI诊断DR，社区以无需转诊DR为主，更适合筛查早期DR；医院以需转诊DR为主。

abstractsObjective:To study the efficiency and difference of the artificial intelligence (AI) system based on fundus-reading in community and hospital scenarios in screening/diagnosing diabetic retinopathy (DR) among aged population, and further evaluate its application value.Methods:A combination of retrospective and prospective study. The clinical data of 1 608 elderly patients with diabetes were continuously treated in Henan Eye Hospital & Henan Eye Institute from July 2018 to March 2021, were collected. Among them, there were 659 males and 949 females; median age was 64 years old. From December 2018 to April 2019, 496 elderly diabetes patients were prospectively recruited in the community. Among them, there were 202 males and 294 female; median age was 62 years old. An ophthalmologist or a trained endocrinologist performed a non-mydriatic fundus color photographic examination in both eyes, and a 45° frontal radiograph was taken with the central fovea as the central posterior pole. The AI system was developed based on the deep learning YOLO source code, AI system based on the deep learning algorithm was applied in final diagnosis reporting by the"AI+manual-check" method. The diagnosis of DR were classified into 0-4 stage. The 2-4 stage patients were classified into referral DR group.Results:A total of 1 989 cases (94.5%, 1 989/2 104) were read by AI, of which 437 (88.1%, 437/496) and 1 552 (96.5%, 1 552/1 608) from the community and hospital, respectively. The reading rate of AI films from community sources was lower than that from hospital sources, and the difference was statistically significant ( χ2=51.612, P<0.001). The main reasons for poor image quality in the community were small pupil (47.1%, 24/51), cataract (19.6%, 10/51), and cataract combined with small pupil (21.6%, 11/51). The total negative rate of DR was 62.4% (1 241/1 989); among them, the community and hospital sources were 84.2% and 56.3%, respectively, and the AI diagnosis negative rate of community source was higher than that of hospital, and the difference was statistically significant ( χ2=113.108, P<0.001). AI diagnosis required referral to DR 20.2% (401/1 989). Among them, community and hospital sources were 6.4% and 24.0%, respectively. The rate of referral for DR for AI diagnosis from community sources was lower than that of hospitals, and the difference was statistically significant ( χ2=65.655, P<0.001). There was a statistically significant difference in the composition ratio of patients with different stages of DR diagnosed by AI from different sources ( χ2=13.435, P=0.001). Among them, community-derived patients were mainly DR without referral (52.2%, 36/69); hospital-derived patients were mainly DR requiring referral (54.9%, 373/679), and the detection rate of treated DR was higher (14.3%). The first rank of the order of the fundus lesions number automatically identified by AI was drusen (68.4%) and intraretinal hemorrhage (48.5%) in the communities and hospitals respectively. Conclusions:It is more suitable for early and negative DR screening for its high non-referral DR detection rate in the community. Whilst referral DR were mainly found in hospital scenario.