Purchase this article with an account.
Ailin Song, Kyung-Min Roh, Jay Lusk, Nita Valikodath, Eleonora M Lad, Alexander T Limkakeng, Joseph A. Izatt, Ryan P McNabb, Anthony N Kuo; Assessment of deep learning-based triage for robotically acquired retinal optical coherence tomography (OCT) images in an emergency department population. Invest. Ophthalmol. Vis. Sci. 2022;63(7):3018 – F0288.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
It is difficult for emergency department (ED) providers to evaluate the posterior segment of the eye with their current standard of care tools. We have previously described the use of robotically aligned OCT (RAOCT) to semi-autonomously acquire high-resolution retinal images. As a next step, we utilize deep learning to triage retinal RAOCT images in a general ED population.
A deep learning model was developed to classify retinal OCT images as referable vs. non-referable for ophthalmology consultation. Using TensorFlow, the model was trained and internally validated on two publicly available datasets (Kermany DS, et al. Cell 2018; Srinivasan PP, et al. BOEx 2014) and images previously captured with our system. For external testing, adult Duke ED patients presenting with suspected posterior eye conditions were consented and enrolled under an IRB-approved protocol and imaged with the RAOCT. For model evaluation, a reference standard was established using a combination of ophthalmology consult diagnosis and OCT image interpretation by two retina specialists. For interpretability, the integrated gradients method was used to generate heatmaps showing areas contributing most to the model classification.
The training and internal validation datasets included 91739 images, of which 60928 were abnormal. The model had a training accuracy of 96% and a validation accuracy of 98%. For external testing, our ED population (RAOCT volumes for 72 eyes of 38 patients; 51% with referable pathology) included a broad range of posterior eye pathologies such as retinal artery occlusion, papilledema, retinal detachment, macular degeneration, and acute retinal necrosis. In this set, the model had an AUC for the detection of referable posterior eye pathology of 0.88, an accuracy of 82%, a sensitivity of 95%, and a specificity of 69% (Fig 1). Areas contributing most to the model classification matched pathologic regions (Fig 2).
An automated OCT-based approach combining robotic OCT imaging and deep learning triage of the images shows promise to help ED providers evaluate patients with potential posterior eye disease.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.
Fig 1. A) Receiver operating characteristic curve and B) confusion table for the external testing set.
Fig 2. Example RAOCT images from our external testing ED population with heatmap overlay.
This PDF is available to Subscribers Only