Purchase this article with an account.
Ashish Bora, Jonathan Krause, Anita Misra, Carter Dunn, Ali Zaidi, Oscar Kuruvilla, Joshua Carlson, Siva Balasubramanian, Christopher Semturs, Lily Peng, Dale R Webster; Deep Learning for Identifying Retinal Vein Occlusion Features in Fundus Images. Invest. Ophthalmol. Vis. Sci. 2019;60(9):1448.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Retinal vein occlusion (RVO) is a rare eye disease. Due to its low prevalence, it’s challenging to collect datasets for deep learning algorithm development. In this work, we develop and evaluate a deep learning algorithm for automated detection of retinal vein occlusion (RVO) features in fundus photographs.
A deep learning algorithm was trained using a retrospective development data set of fundus images (n=878,917) from subjects (n=198,582) with a mean age of 57 years (45% women). The images were independently graded 1 to 49 times (mean = 1.35 times) for the presence of RVO by a group of 122 ophthalmologists. The prevalence of RVO was 0.22% in the training set. The resultant algorithm was evaluated using a separate validation data set enriched for RVO presence. The validation set consisted of 100 images from 86 eyes of 90 patients with a mean age 58 years (31% women). The reference standard for the validation set was decided by the majority decision of 3 US board-certified retina specialists. In this set, 12 images were ungradable, and RVO was present in 59 of 88 gradable images (67%). The algorithm was then evaluated against the reference standard on gradable images (n=88).
For detecting RVO in fundus images, the algorithm had an area under the receiver operating characteristic curve of 0.90 (95% CI, 0.83-0.96; Figure 1). At the chosen operating point (Figure 1), the sensitivity was 81.36% and specificity was 82.76%.
In this study, a deep learning algorithm shows high sensitivity and specificity for detecting RVO features in fundus photographs. Further research is necessary to determine the feasibility of applying this algorithm in the real world clinical setting for improved diagnosis compared with current ophthalmologic assessment.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.
Figure 1. Receiver operating characteristic curve of the deep learning model for RVO presence, with the chosen operating point marked.
This PDF is available to Subscribers Only