A Deep Learning Model For DME Classification Based On Fundus Photographs

Purpose: A Convolutional Neural Network(CNN)was used to establish a classification model for diabetic macular edema(DME)based on fundus photography,the sensitivity and specificity of the model for DME diagnosis were tested by 3D-OCT.Methods: In this study,fundus photographs and 3D-OCT images of diabetic patients aged ≥18years who were screened for diabetic retinopathy in the department of Endocrinology,The Second Affiliated Hospital of Shantou University Medical School from January 2016 to December 2021 were collected.After image quality evaluation,the fundus images were labeled by 3D-OCT classification results and divided into two groups,normal patients without diabetic macular edema and patients with diabetic macular edema,which were used as data sets to construct a DME classification deep learning model based on fundus photography,and the classification performance of the model was evaluated by the OCT results.Results: A total of 1041 patients,1723 eyes,3406 fundus images and the corresponding3D-OCT images were included in this study.The mean age of patients was 55.0±13.3 years.There were 488 males(46.9%)and 553 females(53.1%),and the image ratio of normal group(2089 images)to diabetic macular edema(1317 images)was 1.6:1.780 people(2525 images)were assigned to the training set,104 people(354 images)to the validation set,and 157 people(527 images)to the test set(population ratio 7.5:1:1.5).The deep learning model based on fundus images performed well in DME classification on the test set,with accuracy of 0.861,sensitivity of0.837,specificity of 0.877,F1 score of 0.856 and AUC of 0.913(95%CI: 0.885-0.941).Conclusions: In this study,we used CNN to establish a classification model of DME based on fundus photographs,and the model has achieved good performance.