Development of an Automated Screening System for Retinopathy of Prematurity Using a Deep Neural Network for Wide-Angle Retinal Images

Abstract

Retinopathy of prematurity (ROP) is one of the main causes of childhood blindness. However, insufficient ophthalmologists are qualified for ROP screening. The objective of this paper is to evaluate the performance of a deep neural network (DNN) for the automated screening of ROP. The training and test sets came from 420 365 wide-angle retina images from ROP screening. A transfer learning scheme was designed to train the DNN classifier. First, a pre-processing classifier separated unqualified images. Then, pediatric ophthalmologists labeled each image as either ROP or negative. The labeled training set (8090 positive images and 9711 negative ones) was used to fine-tune three candidate DNN classifiers (AlexNet, VGG-16, and GoogLeNet) with the transfer learning approach. The resultant classifiers were evaluated on a test dataset of 1742 samples and compared with five independent pediatric retinal ophthalmologists. The receiver operating characteristic (ROC) curve, ROC area under the curve, and precision-recall (P-R) curve on the test dataset were analyzed. Accuracy, precision, sensitivity (recall), specificity, F1 score, the Youden index, and the Matthews correlation coefficient were evaluated at different sensitivity cutoffs. The data from the five pediatric ophthalmologists were plotted in the ROC and P-R curves to visualize their performances. VGG-16 achieved the best performance. At the cutoff point that maximized F1 score in the P-R curve, the final DNN model achieved 98.8% accuracy, 94.1% sensitivity, 99.3% specificity, and 93.0% precision. This was comparable to the pediatric ophthalmologists (98.8% accuracy, 93.5% sensitivity, 99.5% specificity, and 96.7% precision). In the screening of ROP using the evaluation of wide-angle retinal images, DNNs had high accuracy, sensitivity, specificity, and precision, comparable to that of pediatric ophthalmologists.