The sensitivity and sepecificity of fundus disease specialist doctors non-fundus disease specialist doctors,and AI software in fundus image interpretation
Objective:This study aims to analyze and compare the consistency of fundus image interpretation among professional fundus disease doctors,non-professional fundus disease doctors,and artificial intelligence-assisted diagnostic software.Methods:A total of 750 fundus images from 523 patients with fundus diseases were collected.These images were diagnosed by a professional group of fundus disease doctors,a non-professional group of fundus disease doctors,and AI-assisted software.The expert group's diagnosis results were used as the gold standard.The sensitivity and specificity of the different diagnostic groups were analyzed and compared.Results:Among the images,the left and right eye compositions were 51.54％and 48.46％,respectively.According to the gold standard,there were 75 cases of diabetic retinopathy,14 cases of macular edema,56 cases of age-related macular degeneration,71 cases of retinal vein occlusion,12 cases of macular epiretinal membrane,317 cases of pathological myopia,and 147 other cases.Negative evaluation criteria were defined as no detection of fundus diseases requiring referral or diagnosis,and positive criteria were defined as the detection of fundus diseases requiring referral or other diseases.The diagnostic performance results for various fundus diseases are as follows:The diagnostic sensitivity(Se)of the fundus disease specialist group was 93.80％(91.62％,95.55％),and the specificity(Sp)was 55.26％(45.66％,64.58％).The diagnostic sensitivity(Se)of the non-ophthalmologists group was 86.33％(83.39％,88.92％),and the specificity(Sp)was 85.96％(78.21％,91.76％).The diagnostic sensitivity(Se)of the AI group was 83.31％(80.17％,86.13％),and the specificity(Sp)was 79.24％(70.28％,86.51％).The diagnostic sensitivity of the professional group of fundus disease doctors was higher than that of the non-fundus disease professional group and the AI group,while the diagnostic specificity of the non-fundus disease professional group was higher than that of the fundus disease professional group and the AI group.The sensitivity of AI-assisted software in the overall diagnosis of multiple ophthalmic diseases is high,meeting the design requirements when compared to a single target value and non-ophthalmologists.However,the specificity does not meet the design criteria in each indicator design.Conclusion:The working years and outpatient volume of professional doctors in the field of fundus disease and non-fundus disease in the study are basically the same,as are the clinical studies they participate in.However,the doctors in the professional group of fundus disease have received specialized training in rapid and extensive reading of fundus images,likely the main reason for their high diagnostic sensitivity.AI software demonstrates good specificity and sensitivity in the overall diagnosis of various fundus diseases and can assist in training the fundus disease reading ability of professional doctors in fundus diseases,providing new ideas for cultivating young doctors.
万方数据
