Application Evaluation Of Artificial Intelligence Reading Diagnosis System In Diabetic Retinopathy Screening In Shanxi Province

Objective:Screening of diabetic retinopathy(DR)in adult diabetic population in some areas of Shanxi.And applying a new method,called the artificial Intelligence(AI)film reading diagnostic system based on deep learning(DL)is used to diagnose DR.And the consistency evaluation with expert diagnosis is carried out to evaluate the application value of AI film reading diagnostic system in DR screening.Methods:From June 2019 to August 2019,4 sites(2 districts,2 counties)in Shanxi Province,Xinghualing District of Taiyuan city,Shanxi Province,Yuci District of Jinzhong City,Huguan County of Changzhi City,Jiangxian County of Yuncheng City,were randomly selected to carry out an epidemiological survey of diabetes complications,based on the China Diabetes Complications Research Project,at the monitoring sites of chronic diseases and their risk factors in China.Four streets(or towns)were randomly selected from each monitoring site,and 120 patients were randomly selected from each street(or towns)according to gender and age,with a total of 1920 patients.Basic information of patients,color fundus images of both eyes and other disease information were collected.A trained ophthalmologist used a mydriasis-free fundus camera to take fundus photos of DM patient(two 45°Color fundus image were taken in each eye,one centered on the optic disc and the other centered on the macula).The image reliability of two Color fundus image in each eye was evaluated,and the samples of fundus color films with poor credibility were discarded.Two deputy chief ophthalmologist made a comprehensive diagnosis of two fundus films of monocular in double-blind condition.When the diagnosis was consistent,the final diagnosis of monocular was directly obtained;when the diagnosis was different,the third chief ophthalmologist participated in the joint diagnosis to obtain the final diagnosis of monocular and calculate the prevalence of DR in each region.Upload color fundus images to the AI film reading diagnostic system,the two monocular overed color fundus image diagnosis and comparison respectively,take more serious read system final diagnosis.The consistency test was conducted on the diagnosis results of the AI film reading diagnosis system and expert diagnosis in three aspects of DR classification,whether there is DR and whether there is a need for referral of DR,and the specificity and sensitivity of the AI film reading diagnosis system were calculated respectively.All the above data were processed by statistical method.We use SPSS 25.0statistical software to analyze data.Counting data were expressed as eye number and percentage,and measurement data were expressed as mean±standard deviation(`x±s).Taking expert diagnosis as the gold standard,the consistency test was carried out between the AI film reading diagnosis system and expert diagnosis in three aspects of DR classification,whether there is DR and whether there is a need for referral of DR,and the specificity and sensitivity of the AI diagnosis system were calculated respectively.P< 0.05 was statistically significant.Results:1.According to the reliability evaluation standard of fundus color film,7444 color films of 3722 eyes were included,including 1868 color films of 934 eyes in Huguan County of Changzhi,1860 color films of 930 eyes in Jiangxian County of Yuncheng,1848 color films of 924 eyes of Yuci District of Jinzhong,and 1868 color films of 934 eyes of Xinghualing District of Taiyuan.2.In this study,the prevalence of DR in adult DM patients in four regions of Shanxi Province was about 20.6%,the prevalence rates of mild non-proliferative diabetic retinopathy(NPDR),moderate NPDR,severe NPDR and proliferative diabetic retinopathy(PDR)were 10.7%,5.6%,3.3% and 1.1%,respectively,indicating that the prevalence rates of DR in rural areas and urban DR were 22.3% and 18.9% respectively.The prevalence of DR in county and township areas(22.3%)was higher than that in district and street areas(18.9%).3.According to the expert diagnosis,2955 eyes had no DR state,and 767 eyes had different degrees of DR.In the diagnosis results of AI film reading diagnosis system,the overall diagnosis coincidence rate was 81.0% and the specificity was 86.4%,the diagnosis coincidence rate of DR was 60.4% and the Kappa coefficient was 0.541(P<0.001)in the consistency test with the expert diagnosis.AI film reading and diagnosis system has medium consistency with expert diagnosis in DR grading.In terms of DR or not,the specificity and sensitivity of the AI film reading diagnosis system were 86.4%and 85.7% compared with the expert diagnosis results.In the consistency test,the Kappa coefficient was 0.632(P<0.001),showing a high degree of consistency.In terms of DR or not,the specificity and sensitivity of the AI film reading diagnosis system were 90.6%and 91.4% compared with the expert diagnosis results.In the consistency test,the Kappa coefficient was 0.612(P<0.001),showing a high degree of consistency.Conclusions:The prevalence rate of DR in adult DM population in Parts of Shanxi Province is20.6%,which is higher in Townships than in Street areas.Therefore,DR is an ophthalmic disease that ophthalmologists in Shanxi Province,especially in rural areas,need to focus on prevention and treatment.Deep learning-based AI film reading diagnosis system has a high degree of consistency with expert diagnosis results in the presence of DR and the need for referral of DR,which can be used for screening of DR in primary medical institutions.