Automated identification of retinopathy of prematurity by image-based deep learning

• Published in: Eye and vision (Novato, Calif.) Vol. 7; no. 1; pp. 40 - 12
• Main Authors: Tong, Yan, Lu, Wei, Deng, Qin-Qin, Chen, Changzheng, Shen, Yin
• Format: Journal Article
• Language: English
• Published: England BioMed Central Ltd 01-08-2020; BioMed Central; BMC
• Subjects: Accuracy; Algorithms; Analysis; Artificial intelligence; Artificial neural networks; Automation; Blindness; Cable television broadcasting industry; Classification; Datasets; Deep learning; Diabetic retinopathy; Disease; Experts; Fundus image; Hospitals; Identification; Labeling; Medical personnel; Neural networks; Premature babies; Premature birth; Retinal detachment; Retinopathy of prematurity; Studies; Technology application; United States; Deep learning; Fundus image; Retinopathy of prematurity; Artificial intelligence
• ISSN: 2326-0254; 2326-0246; 2326-0254
• DOI: 10.1186/s40662-020-00206-2

Retinopathy of prematurity (ROP) is a leading cause of childhood blindness worldwide but can be a treatable retinal disease with appropriate and timely diagnosis. This study was performed to develop a robust intelligent system based on deep learning to automatically classify the severity of ROP from fundus images and detect the stage of ROP and presence of plus disease to enable automated diagnosis and further treatment.
A total of 36,231 fundus images were labeled by 13 licensed retinal experts. A 101-layer convolutional neural network (ResNet) and a faster region-based convolutional neural network (Faster-RCNN) were trained for image classification and identification. We applied a 10-fold cross-validation method to train and optimize our algorithms. The accuracy, sensitivity, and specificity were assessed in a four-degree classification task to evaluate the performance of the intelligent system. The performance of the system was compared with results obtained by two retinal experts. Moreover, the system was designed to detect the stage of ROP and presence of plus disease as well as to highlight lesion regions based on an object detection network using Faster-RCNN.
The system achieved an accuracy of 0.903 for the ROP severity classification. Specifically, the accuracies in discriminating normal, mild, semi-urgent, and urgent were 0.883, 0.900, 0.957, and 0.870, respectively; the corresponding accuracies of the two experts were 0.902 and 0.898. Furthermore, our model achieved an accuracy of 0.957 for detecting the stage of ROP and 0.896 for detecting plus disease; the accuracies in discriminating stage I to stage V were 0.876, 0.942, 0.968, 0.998 and 0.999, respectively.
Our system was able to detect ROP and differentiate four-level classification fundus images with high accuracy and specificity. The performance of the system was comparable to or better than that of human experts, demonstrating that this system could be used to support clinical decisions.