Purchase this article with an account.
Joelle Hallak, Nooshin Mojab, Joseph Baker, Vahid Noroozi, Dimitri T Azar, Mark Rosenblatt; Detecting Glaucoma and Suspect Progression through Longitudinal Fundus Photos. Invest. Ophthalmol. Vis. Sci. 2019;60(9):1472.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To develop an automated glaucoma detection system using fundus photos, and determine whether early fundus photos can distinguish between suspects who progress to glaucoma versus suspects who do not progress.
We created a large-scale ophthalmic imaging dataset “The Illinois Ophthalmic Database Atlas” (I-ODA) with deep learning methods labeling 3,668,649 images of 33,876 patients linked to their diagnosis using billing data. We isolated 45,627 fundus photos for 1,423 glaucoma and 2,474 non-glaucoma patients. To test the usability of retrospective fundus photos in detecting glaucoma, we used a deep Convolutional Neural Network (ResNet50) to extract features for input images, and then applied a simple classification task using support vector machine (SVM) to classify glaucoma and non-glaucoma images. To test whether we are able to predict suspect progression, we divided patients into progressors, suspects who progressed to glaucoma, and non-progressors, patients who remained as suspect throughout their follow-up. Patients with at least one-year follow-up were included. SVM classification was used on first fundus photos and fundus photos preceding diagnosis change from suspect to glaucoma.
SVM revealed a sensitivity of 89.3% and specificity of 97.6% in classifying glaucoma versus non-glaucoma fundus photos. There were 2,402 patients with 11,135 longitudinal fundus photos who were diagnosed with suspect at their first visit and remained as suspect at the last visit, average follow-up was 3 years + 2.6, and 326 patients with 960 longitudinal fundus photos who were diagnosed with suspect at their first visit and then progressed to glaucoma, average follow-up was 3 years + 2.3 years. Using fundus photos at the first visit, SVM revealed a 59.2% sensitivity and 98.8% specificity. When taking fundus photos on exam dates with at least one year before progressing to glaucoma, SVM revealed a 53.3% sensitivity and a 99.1% specificity.
Our method in classifying glaucoma versus non-glaucoma using fundus photos yielded high sensitivity and specificity, however this method showed that fundus photos may not distinguish in identifying progressors in glaucoma at first photo or one year prior to progression. Additional analyses with various methods and datasets are needed to determine the time point of structural changes in fundus photos to distinguish progressors from non-progressors.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.
This PDF is available to Subscribers Only