Automatic classification and triage of diabetic retinopathy from retinal images based on a convolutional neural networks (CNN) method

• Published in: Acta ophthalmologica (Oxford, England) England), 2019-12, Vol.97 (S263), p.n/a
• Main Authors: Galdran, Adrian, Chakor, Hadi, Alrushood, Abdulaziz A., Kobbi, Ryad, Christodoulidis, Argyrios, Chelbi, Jihed, Racine, Marc‐André, Benayed, Ismail
• Format: Article
• Language: English
• Subjects: Aging; Classification; Computer vision; Deep learning; Demographics; Developed countries; Diabetes; Diabetes mellitus; Diabetic retinopathy; Epidemiology; Learning algorithms; Neural networks; Ophthalmology; Pattern recognition; Primary care; Retina; Retinal images; Retinopathy
• ISSN: 1755-375X; 1755-3768
• DOI: 10.1111/j.1755-3768.2019.5391

Purpose Diabetic retinopathy (DR) is one of the leading causes of adult vision loss in the developed countries. Epidemiological and demographic factors, including the rising rates of diabetes related to obesity and an aging population, are driving the incidence of diabetic eye complication inexorably higher. Method Deep learning emerges as a powerful tool for analyzing and classifying retinal images in an automatic way, but the classification results depend greatly on the availability of large datasets. As the number of categories and the imbalance ratio increase, the performance of deep learning models diminishes. In the context of DR grading, minority classes (mild and severe DR) are critical to diagnose. Experiments were performed on a real dataset developed at local hospital and at different hospitals around the world. Overall, 42 179 retinal images were obtained from Diagnos database. All images were graded by 3 retinal experts using the early treatment diabetic retinopathy study severity scale (ETDRS). The dataset was built by expanding on 4 categories: R0 or normal, R1 or mild DR, R2 or moderate DR, and R3&R4 or severe and proliferative DR. The data was split 90/10 for training and testing respectively, and an ensemble of Convolutional Neural Networks was trained to perform DR grading. Results The proposed method achieves high accuracy in predicting DR grades, with the R1 class showing lower performance, in line with recently proposed methods. An area under the ROC curve of 0.96 (0.95–0.96) for R0, 0.70 (0.65–0.75) for R1, 0.95 (0.94–0.95) for R2 and 0.92 (0.89–0.96) for R34. Conclusion Comparable to the score of human experts, the deep learning techniques in this study were effective to be applied in clinical use as primary care setting and could be a valuable tool to help primary care triage. Improvement in detection of R1 subjects is needed for further progressing in this area. References 1. He K, Zhang X, Ren S & Sun J (2016): Deep Residual Learning for Image Recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 770–778. 2. Krause J et al. (2018): Grader variability and the importance of reference standards for evaluating machine learning models for diabetic retinopathy. Ophthalmology 125: 1264–1272.