Application Of Artificial Intelligence Diagnosis For Fundus Disease Based On Color Fundus Images

Objective:To develop and validate a deep transfer learning(DTL)algorithm for detecting abnormalities in fundus images from nonmydriatic fundus photography examinations.To evaluate the diagnostic performance and feasibility of an offline artificial intelligence(AI)-based referable diabetic retinopathy(RDR)screening model in the physical examination center.Methods:A total of 1295 fundus images were collected to develop and validate a DTL algorithm for detecting abnormal fundus images.After removing 366 poor images,the DTL model was developed using 929(370 normal and 559 abnormal)fundus images.Data preprocessing was performed to normalize the images.The Inception-Res Net-v2 architecture was applied to achieve transfer learning.We tested our model using a subset of the publicly available Messidor dataset and evaluated the testing performance of the DTL model for detecting abnormal fundus images.Heatmaps were introduced in true positive results to identify important areas that DTL model pays attention to when classifying abnormal fundus images.Physical examination people who were initiatives choosing DR screening in the physical examination center from November 2019 to January 2020 were enrolled.The AI system displayed a binary RDR or non-RDR result.The accuracy,sensitivity,specificity,F1 score,positive predictive value,and negative predictive value of the offline AI-based system for screening RDR was evaluated regarding the standard classification criteria.Results:In the internal validation dataset(n = 273 images),the area under the curve(AUC),sensitivity,accuracy,and specificity of DTL for correctly classified fundus images were0.997,97.41%,97.07%,and 96.82%,respectively.For the test dataset(n = 273 images),the AUC,sensitivity,accuracy,and specificity of the DTL for correctly classifying fundus images were 0.926,88.17%,87.18%,and 86.67%,respectively.Of the 163 participants 31(19%)patients with a history of diabetes mellitus.The mean(SD)age of the participants was 51(10)years which ranged from 22 to 83 years.The AI system correctly classified 10(3.3%)images as having RDR,while 2 images(0.7%)were missed diagnosed,and 7 images(2.3%)were misdiagnosed.The performance of the AI system for diagnosing RDR referencing to ophthalmologist grading standard: the sensitivity,accuracy,F1 score,and specificity were: 83.3%(95%CI,50.9%-97.1%),97.1%,70%,and97.6%(95%CI,95.0%-99.0%),respectively.Conclusion:DTL showed high sensitivity and specificity for detecting abnormal fundus-related diseases.Further research is necessary to improve this method and evaluate the applicability of DTL in community health-care centers.This pilot study results exhibit a promising future for using an offline AI-based system screening referable diabetic retinopathy in the physical examination center.