论文部分内容阅读
目的:探索人工智能（AI）初筛分流在大规模糖尿病视网膜病变（DR）筛查中的应用。方法:2018年5至7月在新疆维吾尔自治区喀什市和克孜勒苏柯尔克孜自治州，8 005例糖尿病患者参加了DR筛查，所有患者均行免散瞳的眼底彩照检查，每眼采集2张眼底彩照（分别以视盘和黄斑为中心）。拍照完成后，首先使用AI系统对每例患者的4张眼底彩照进行判定，如均为可分级且无DR时，直接产生AI报告，并随机抽取1/3至读片中心由眼科医师组分级以作质量控制（组A）。此外，如4张眼底彩照中，任意一张图片判别为无法分级或存在DR时，该患者的所有眼底彩照均需行人工分级（组B）。从组A和组B中分别随机抽取300例患者的眼底照片，由眼科医师组作出最终判别，确定AI和人工分级判定需转诊DR（增殖前或增殖期DR，或糖尿病黄斑水肿）的准确性。结果:在8 005例参加筛查的糖尿病患者中，男3 220例（40.2%），女4 785例（59.8%），年龄（58.3±10.6）岁。经AI初筛后，AI直接产生报告5 267例（65.8%），另外2 738例（34.2%）需要进行人工分级。在组A中，AI和人工判定需转诊DR的准确性和特异度均为100.0%。在组B中，AI和人工判定需转诊DR的准确性分别为75.8%和90.3%，灵敏度分别为100%和79.1%。结论:在大规模的DR筛查中，使用AI作为DR的初筛手段，可在不遗漏需转诊DR病例的情况下，减少约60%的图片分级工作量。“,”Objective:To investigate the diagnostic accuracy and efficiency of an artificial intelligence (AI) triaging model in a diabetic retinopathy (DR) screening program.Methods:A DR screening program was conducted in Kashi City and Kizilsu Kirghiz Autonomous Prefecture of the Xinjiang Uyur Autonomous Region from May to July 2018, and 8 005 patients with diabetes mellitus were included. Fundus images, one centered at optic disc and one centered at macula, were taken for both eyes. A previously validated AI algorithm was applied as the first step to identify the patients with all 4 images. If the images were classified as gradable and negative DR, an AI-generated report was immediately provided without sending to manual grading, and 1/3 of these patients were randomly sampled for manual grading and quality control (group A). For the patients with at least one image classified as ungradable or positive for any DR, all images were sent for manual grading (group B). Finally, 300 patients were randomly selected from group A and group B respectively for accuracy assessment, where the patients and their images were classified by a specialist panel for referral DR (pre-proliferative DR, or proliferative DR, and/or diabetic macular edema).Results:Among 8 005 patients for DR screening [including 3 220 males and 4 785 females, aged (58.3±10.6) years], after AI triaging, 5 267 (65.8%) potentially received reports from AI system and 2 738 (34.2%) required manual grading. In group A, the accuracy and specificity of AI classification and manual grading on referral DR were all 100%. In group B, the accuracy of AI and manual grading were 75.8% and 90.3%, respectively, while the sensitivity of AI and manual grading was 100% and 79.1%, respectively.Conclusion:AI alleviates 60% of the workload of manual grading without missing any referral patients with the aid of the current AI triaging model.