Clinical value of artificial intelligence in large-scale screening of diabetic retinopathy in grassroots hospitals
Objective To explore the clinical value of artificial intelligence(AI)system in screening diabetic retinopathy(DR)in grassroots hospitals.Methods A total of 28 300 diabetic patients(56 600 eyes)underwent DR screening in Haining People's Hospital from July 2020 to December 2021.Among them 3 512 cases(7 024 eyes)were selected for the study by stratified random sampling method.Using fundus pictures from AFC-330 automatic mydriasis fundus camera as a picture library,the AI system(Zhejiang Ruiyi Medical AI System Retina Reading System)provided diagnosis of DR,DR grading and referable diabetic retinopathy(RDR).DR diagnosis,DR grading diagnosis and RDR diagnosis were given by ophthalmologists based on the results of fundus photography,optical coherence tomography(OCT)and fluorescein fundus angiography(FFA).DR diagnosis,DR grading diagnosis,RDR diagnosis and the average time consuming were compared between the efficacy of AI system and ophthalmologist.Kappa test was used to evaluate the consistency of AI system with ophthalmologist's diagnosis of DR,DR grading and RDR.Results With ophthalmologists'manual diagnosis as the gold standard,the sensitivity and specificity of the AI system in diagnosing DR were 91.69％and 98.63％,respectively;the Kappa value of the consistency analysis of the two diagnostic results was 0.831(P＜0.001).The sensitivity and specificity of AI system in diagnosing mild non-proliferative diabetic retinopathy(NPDR),moderate NPDR,severe NPDR,and proliferative diabetic retinopathy(PDR)were 82.72％,84.36％,86.11％,86.95％and 98.93％,99.27％,99.92％,99.97％,respectively;the Kappa value of the consistency analysis between the two methods for DR grading diagnosis was 0.847(P＜0.001).The sensitivity and specificity of the AI system for diagnosing RDR are 88.16％and 99.32％,respectively.The Kappa value of the consistency analysis between the two for RDR diagnosis is 0.828(P＜0.001).The average time taken for a single film reading in the AI system was(1.58±0.22)s,which is shorter than that of the manual reading(5.83±2.11)s(P＜0.001).Conclusion AI system has the same sensitivity and specificity as manual reading in DR diagnosis,DR grading and RDR diagnosis of diabetes patients with high consistency,indicating that AI system may be used for large-scale DR screening in grassroots hospitals.
万方数据
维普
