Abstract
Purpose :
To develop a deep learning system (DLS) capable of distinguishing “normal optic nerves”, “papilledema” (optic disc edema from proven intracranial hypertension), and “other optic nerve abnormalities” on standard digital fundus photographs from a large, multiethnic, worldwide, patient population. This classification was chosen to provide a non-invasive test that facilitates identification of patients with optic disc abnormalites (papilledema) associated with intracranial abnormalities potentially responsible for sight- and/or life threatening consequences.
Methods :
We developed and validated a DLS to automatically classify optic discs as “normal” or “abnormal”, and specifically detect “papilledema”, using 15,846 digital ocular fundus photographs (14,341 images for DLS training and validation; 1,505 for external testing) from adult patients as part of an international consortium (BONSAI, Brain and Optic nerve Study with Artificial Intelligence). The DLS performance to classify the optic disc appearance was evaluated by calculating the area under the receiver operating curve (AUC), sensitivity and specificity, with reference to reference standards, based on clinical examination and ancullary investigations.
Results :
We included 9,156 images of “normal” discs, 2,148 images with “papilledema”, and 3,037 images with “other” optic disc abnormalities. In the primary validation dataset, the DLS successfully discriminated “normal” from “abnormal” optic discs (AUC 0.99 [0.99-0.99]), and “papilledema” from “other” (AUC 0.98 [0.98-0.98]. Similar performance was observed on external datasets, with AUC 0.98 (0.97-0.98), sensitivity 95.3 (93.8-96.6) and specificity 86.6 (83.8-89.3) for the detection of “normal”, and AUC 0.96 (0.95-0.97), sensitivity 96.4 (94.2-98.1) and specificity 84.7 (82.6-86.7) for the detection of “papilledema”.
Conclusions :
A fundus photograph-based DLS can automatically discriminate normal optic discs, papilledema and other optic disc abnormalities in a multi-country, multi-ethnic patient population, with potential applications for the management of headache and neurologic patients in various clinical settings.
This is a 2020 ARVO Annual Meeting abstract.