A Deep Learning Approach to Accurately Discriminate Between Optic Disc Drusen and Papilledema on Fundus Photographs

ObjectiveTo assess the performance of a deep learning system (DLS) to discriminate between optic disc drusen (ODD) and papilledema caused by intracranial hypertension, using standard color ocular fundus photographs collected in a large international multi-ethnic population. DesignRetrospective study. ParticipantsThe study included 4,508 color fundus images in 2,180 patients from 30 neuro-ophthalmology centers (19 countries) participating in the Brain and Optic Nerve Study with Artificial Intelligence (BONSAI) Group. MethodsWe trained, validated, and tested a dedicated DLS for binary classification of ODD vs. papilledema (including various subgroups within each category), on conventional mydriatic digital ocular fundus photographs. For training and internal validation, we used 857 ODD images and 3,230 papilledema images, in 1,959 patients. External-testing was subsequently performed on an independent dataset (221 patients) including 207 images with ODD (96 visible and 111 buried), provided by 3 centers of the Optic Disc Drusen Studies Consortium, and 214 images of papilledema (92 mild-to-moderate and 122 severe) from a previously validated study. Main outcome measuresArea under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity were used to discriminate between ODD and papilledema. ResultsOverall, the DLS could accurately distinguish between all ODD and papilledema (all severities included): AUC 0.97 (95% confidence interval [CI], 0.96 to 0.98), accuracy 90.5% (95% CI, 88.0% to 92.9%), sensitivity 86.0% (95% CI, 82.1% to 90.1%), and specificity 94.9% (95% CI, 92.3% to 97.6%). The performance of the DLS remained high for discrimination of buried ODD from mild-to-moderate papilledema: AUC 0.93 (95% CI, 0.90 to 0.96), accuracy 84.2% (95% CI, 80.2%-88.6%), sensitivity 78.4% (95% CI, 72.2% to 84.7%), and specificity 91.3% (95% CI, 87.0% to 96.4%). ConclusionsA dedicated DLS can accurately distinguish between ODD and papilledema caused by elevated intracranial pressure, even when considering buried ODD vs mild-to-moderate papilledema. Future studies are required to validate the utility of this DLS in clinical practice.