Abstract
Purpose :
To apply deep learning to create an algorithm for automated detection of glaucomatous optic nerve head features in fundus photographs and to compare the performance of the algorithm with manual grading by ophthalmologists.
Methods :
A deep convolutional neural network was trained using a retrospective development data set of 27590 fundus images, which were graded 1 to 4 times for risk of glaucoma based on the presence of characteristic optic nerve head features, by a panel of US licensed ophthalmologists. Majority decision of the ophthalmologist panel was used to create a reference standard for each image. The training set consisted of 27590 images from 23500 patients (mean age 54 years; 59% women, prevalence of glaucoma suspect 1547/23846 gradable images [6.4%]). The resultant algorithm was validated using a separate data set not used for training graded by at least 3 US board-certified ophthalmologists. The sensitivity and specificity of the algorithm for detecting glaucoma suspect subjects were generated based on the reference standard of the majority decision of the ophthalmologist panel. The algorithm was evaluated at 2 operating points selected from the development set, one selected for high specificity and another for high sensitivity.
Results :
The test set consisted of 3239 images from 2366 patients (mean age 57 years; 62% women, prevalence of glaucoma 608/3071 gradable images [20%]). For detecting glaucoma suspect patients the algorithm had an area under the receiver operating curve of 0.926 (95% CI, 0.915-0.936). Using the first operating cut point with high specificity, the sensitivity was 77.8% and the specificity was 90%. Using a second operating point with high sensitivity in the development set, the sensitivity was 90% and specificity was 79.4%.
Conclusions :
In this evaluation of fundus photographs, an algorithm based on deep machine learning had high sensitivity and specificity for detecting glaucoma suspect subjects. Further research is necessary to determine the feasibility of applying this algorithm in the clinical setting and to determine whether use of the algorithm could lead to improved care and outcomes compared with current ophthalmologic assessment.
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.