Research on screening system of myopic maculopathy based on deep convolution neural network / 中华实验眼科杂志

ABSTRACT

Objective:To develop a fully automatic detection system based on the deep convolution neural network (DCNN) for screening myopic maculopathy (MMD) and identifying its severity.Methods:Six thousand and sixty-eight fundus images were collected from Anhui No.2 Provincial People's Hospital to construct the training set, and the public fundus images data set was selected to construct the test set.The fundus images were preprocessed and amplified, and the grade of MMD lesions was labeled and the data was cleaned.The automatic MMD detection system proposed was composed of two-level network.The first level network structure was used to identify the presence of MMD, and the second level network structure was used to diagnose the severity of MMD lesions.The accuracy, specificity, sensitivity, precision, F1 value, area under curve (AUC) and Kappa coefficient of four commonly used DCNN network methods, VGG-16, ResNet50, Inception-V3 and Densenet, in MMD screening and severity recognition tasks were compared and analyzed.The study protocol adhered to the Declaration of Helsinki and was approved by a Medical Ethics Committee of Anhui No.2 Provincial People's Hospital ([L]2019-013).Results:The performance of Densenet network model was the best in the MMD screening task, with the sensitivity, specificity, accuracy, F1 value and AUC of 0.898, 0.918, 0.919, 0.908 and 0.962, respectively.The Inception-v3 network model was the best in MMD severity recognition task, with sensitivity, specificity, accuracy, F1 value and AUC of 0.839, 0.952, 0.952, 0.892, and 0.965, respectively.The visualization results showed that the network structure model used in this study could automatically learn the clinical characteristics of MMD severity, and accurately identify diffuse and focal chorioretinal atrophy areas.Conclusions:The MMD screening method using fundus images based on DCNN can automatically extract the effective features of MMD, and accurately screen MMD and judge its severity, which can provide effective assistance in clinical practice.