Abstract
Purpose :
To assess the performance of a deep learning algorithm for detecting referable glaucomatous optic neuropathy (GON) and late wet age-related macular degeneration (LW-AMD) based on color fundus photographs.
Methods :
A deep learning system for automated classification of GON and LW-AMD was developed based on more than 50,000 fundus photographs. A total of 21 trained ophthalmologists were invited to label all these images using an online cloud source system. Referable GON was defined as vertical cup/disc ratio ≥ 0.7 and other typical changes of GON. AMD was classified as absent, early, intermediate, late dry and late wet subtypes. The reference standard was made when ≥3 graders had achieved agreement. A separate validation dataset of 8,000 fully gradable fundus photographs was used to assess the performance of this algorithm.
Results :
In the validation dataset, this deep learning system achieved an AUC of 0.986 with sensitivity of 95.6% and specificity of 92.0% for GON; these results were 0.995, 98.0% and 94.0% for LW-AMD, respectively. The most common reasons for false negative grading (n=87) for GON were co-existing eye conditions (n=44, 50.6%) while the reason for false positive cases (n=480) was having other eye conditions (n=458, 95.4%), mainly including physiologic cupping (n=267, 55.6%). For the misclassification of LW-AMD, serous detachment of the sensory retina or retinal pigment epithelium (n=8, 57.2%) were identified as the most common features in the false negative classification. Images with other eye diseases (n=145, 76.3%), such as diabetic retinopathy (n=38, 20%), myopic maculopathy (n=22, 11.5%), early or intermediate AMD (n=21, 11.1%) and choroiditis (n=12, 6.3%), were the commonest features for false positive classification.
Conclusions :
A deep learning system can detect referable GON and LW-AMD with high sensitivity and specificity. The algorithm should be further validated among the images collected in real-world clinical practice.
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.