Mingguang He, Zhixi Li, Stuart Keel, Yifan He, Wei Meng
Purpose: To access the performance of a custom deep learning algorithm (DLA) on detecting refer- able diabetic retinopathy (DR), glaucoma suspect, late age-related macular degeneration (AMD) andretinal arteriosclerosis.
Method: A cloud-sourcing clinical labeling system was established to grade approximately 150,000 reti- nal images, including a training set (56,095 images for DR, 32,000 for glaucoma suspect, 24,000 for AMD and 44,815 for atherosclerosis) and an independent validation set (9,372 for DR, 15,150 for glaucoma sus- pect, 18,128 for AMD and 7,334 for arteriosclerosis). These datasets were used to develop a custom DLA. A sample of 4,000 images was randomly selected from an independent validation set for each disease condi- tion respectively. Referable DR was deftned as R2 or worse using the English NHS Classiftcation; glaucoma suspect was deftned as vertical CDR ? 0.7 and other changes of glaucomatous optic neuropathy; late AMD was deftned including late wet and late dry AMD; arteriosclerosis was deftned as AVR less than 0.5 and arteriovenous nicking, focal arteriolar narrowing.
Results: In this independent validation assessment, the sensitivity, speciftcity and the area under curve (AUC) were 92.6%, 94.3% and 0.934, respectively among 2817 negative and 1183 positive referable DR images. These numbers were 94.5%, 94.1% and 0.943 among 3279 negative and 721 positive glau-
coma suspect; 90.1%, 88.2% and 0.891 among 3732 negative and 735 positive late AMD images; and 87.0%, 88.2% and 0.876 among 3839 negative and 161 positive retinal arteriosclerosis images.
Conclusion: This custom DLA achieves clinically- sound accuracy on classifying common sight- threatening disorders that would potentially improve the efftciency and accessibility of screening service.