Purchase this article with an account.
Emily Seo, Nicolas Jaccard, Sameer Trikha, Louis R. Pasquale, Brian J Song; Automated Evaluation of Optic Disc Images for Manifest Glaucoma Detection Using a Deep-Learning, Neural Network-Based Algorithm. Invest. Ophthalmol. Vis. Sci. 2018;59(9):2080. doi: https://doi.org/.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To compare the diagnostic accuracy of optic disc image evaluation by glaucoma specialist evaluation versus an automated, deep-learning, decision support tool (Pegasus; Visulytix Ltd, London, United Kingdom) for detection of manifest glaucoma.
Two glaucoma specialists (BJS and LRP), masked to clinical data, independently reviewed monoscopic color optic disc photographs for neuroretinal rim thinning (yes/no), retinal nerve fiber layer (RNFL) defect (yes/no), and likelihood for manifest visual field loss (normal, borderline, abnormal). Disparities were adjudicated by consensus. We deposited optic disc images into Pegasus, which generated a disc anomaly score ranging from 0.00 to 1.00. Sensitivity and specificity were determined through comparison with a reference standard for manifest glaucoma defined by a red or yellow color code on the quadrant scan of high quality (signal strength ≥6/10) spectral-domain RNFL optical coherence tomography in combination with a repeatable abnormal glaucoma hemifield test on a reliable (<33% fixation losses; <20% false positive and negative errors) 24-2 Humphrey visual field, both of which were within 6 months of test optic disc photographs. Diagnostic accuracy was compared by calculating area under the receiver operating characteristic curve (AUROC).
Of 186 optic disc images from 186 patients, 81 (43.5%) met reference standard criteria for manifest glaucoma (Table). All images were deemed gradable by Pegasus and both glaucoma specialists. Sensitivity and specificity of glaucoma specialist evaluation of optic disc images for manifest glaucoma were 77.8% and 93.3%, respectively. Sensitivity and specificity of Pegasus for manifest glaucoma were 77.8% and 81.9%, respectively. AUROC of glaucoma specialist evaluation and Pegasus were 0.87±0.03 and 0.84±0.03, respectively (Figure), and this difference was not statistically significant (p=0.27; DeLong test).
Automated deep-learning evaluation of morphologic optic disc features performs similarly to adjudicated expert evaluation of monoscopic optic disc photographs for prediction of manifest glaucoma. Such algorithms may be helpful to distinguish glaucoma suspect and glaucomatous optic discs in the absence of ancillary functional testing, such as in ocular telemedicine programs.
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.
Table. Baseline Demographics
Figure. ROC Curves for Manifest Glaucoma Detection
This PDF is available to Subscribers Only