Deep Learning Is Effective for Classifying Non-referable versus Referable Eye Condition using Fundus Photographs

Abstract

Purpose : Fundus photographs is the most common imaging modality for screening eye disease. This study aimed to determine whether deep learning could be utilized to distinguish referable eye disease (RED) from normal fundus photographs for general eye screening.

Methods : This retrospective observational database study included fundus photographs and results for 98,816 anonymous eye examinations from single tertiary center between March 2013 and June 2016. Automated extraction of a fundus photographs from health screening database was performed and linked to the results of ophthalmologic examination report. Deep learning algorithm was developed randomly selected 20,000 non-RED and a total of 13,877 RED images in order to make similar ratio. RED includes any macular diseases, glaucoma suspect, or media opacity and so on. Validation was performed using a random subset of 300 fundus photographs from 4,199 health screening database between Jan 2017 to Mar 2017 was performed. The main outcome measure of our study is the accuracy and the area under the receiver operating characteristic curve.

Results : For detecting RED, we achieve an area under the ROC curve of 0.898 (95% CI, 0.863-0.933) using validation set. Using the cut point with high sensitivity of >90%, validation study showed that sensitivity/specificity was 91.0% (95% confidence interval [CI] 84.4 to 95.4) / 73.0% (95% CI 65.9 to 79.4) in validation sets.

Conclusions : Our findings suggest that an algorithm based on deep learning have a possibility to use for detecting referable eye disease with a high sensitivity setting. Further clinical based study to determine feasibility of applying this algorithm is needed.

This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.