Study On Automated Diagnosis Of Aggressive Posterior Retinopathy Of Prematurity Based On Deep Learning

Objective: An automatic diagnostic system based on deep learning was designed to diagnose aggressive posterior retinopathy of prematurity(AP-ROP)and evaluate its value in clinical application.Methods: A total of 5474 fundus images of 242 premature infants undergoing screening and treatment for retinopathy of prematurity(ROP)in Shenzhen eye hospital from January 2009 to May 2018 were collected as the data set of this study.The images were taken by the wide-field digital pediatric retinal imaging system(Ret Cam).Six ophthalmologists diagnosed images in the training set as normal,regular ROP and AP-ROP according to the international classification of ROP(ICROP).The convolutional neural network(CNN)learned the diagnosis from experts,so as to establish an algorithm model that can identify the disease,and eventually diagnosed AP-ROP.This study included two CNNs,and the identification of diseases were divided into two steps.First,all images were classified as normal and ROP using network-1.Second,all the AP-ROP images were identified from the ROP images using network-2.The network performance was evaluated using the following methods: The test set,the comparison between dual network association and single network,and the man-machine comparison.Results: In the network-1,we achieved an accuracy of 96.53%,sensitivity of 96.74% and specificity of 96.39%.In the network-2,and the results(98.46% accuracy,100.00% sensitivity and 96.90% specificity)were comparable with the network-1.The results of the comparison between double network and single network showed that the accuracy of the double network was 95.93%,which was higher than that of the single network.The results of man-machine comparison showed that the accuracy of ophthalmologists and the algorithm model were all higher than 90% in identifying ROP,but in recognition of AP-ROP,only experts and the algorithm model achieved the accuracy of 90%.Conclusion: The automatic diagnostic system based on deep learning has high accuracy,sensitivity and specificity in diagnosing AP-ROP.Applying this technique to Ret Cam system or telemedicine system is expected to help ophthalmologists improve their efficiency to identify diseases.