Purchase this article with an account.
Dan Milea, Zhubo Jiang, Caroline Vasseneix, Clare Fraser, Philippe Goher, Raymond Najjar, Shweta SInghal, Daniel Ting, Ambika Selvakumar, Jost B Jonas, Nancy Newman, Neil Miller, Yong Liu, Valerie Biousse, Tien Yin Wong; Brain and Optic Nerve Study with Artificial Intelligence (BONSAI). Invest. Ophthalmol. Vis. Sci. 2019;60(9):1476.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To assess the performance of a newly developed deep learning system for automatic detection of papilledema and classification of differential diagnosis, on fundus images.
International study in 12 centres in 8 countries, using supervised deep convolutional neural networks (CNN) applied to fundus photographs, in controls and in patients with swollen optic discs related to 1/ raised intracranial pressure (papilledema related to space-occupying lesions or idiopathic intracranial hypertension) 2/ other confirmed optic neuropathies (ischemic, inflammatory, compressive, etc). A group of patients with pseudo-papilledema (optic disc drusen, hyperopia) was also included, as a differential diagnosis of truly swollen discs. Curation of the included images was followed by pre-processing, normalization and bootstrap training with a transfer learning strategy, using a deep CNN, pre-trained on more than a million images.
We have included 2623 participants (including 1776 patients with optic disc abnormalities) and analyzed in total 6443 fundus images (including 3563 abnormal discs images). Among those, 5066 images were used for training and 1377 images for testing. The algorithm achieved excellent performance in the testing dataset, for discrimination of papilledema from normal optic discs (AUC, 0.98) and papilledema from ischemic optic neuropathies (AUC=0.95). More interestingly, the system could distinguish between true papilledema and false papilledema (pseudo-papilledema), (AUC, 0.94).
Deep learning algorithms, appllied to a large, international dataset of patients with different ethnicities, can accurately distinguish on fundus images between papilledema (caused by intracranial abnormalities) and normal discs. Moreover, our system could accurately distinguish between true papilledema and pseudo-papilledema, as well as between papilledema and anterior ischemic optic neuropathy. Artificial Intelligence-based fundus photography may enhance neuro-ophthalmologic diagnoses, and potentially impact the acute management of vision- or life-threatening conditions in patients who may initially present to non-ophthalmic care providers.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.
Receiver operating characteristics curve analysis of our DLS for discrimination on retinal images (from left to right), of: normal discs vs. papilledema, non-arteritic ischemic optic neuropathy (NAION) vs. papilledema and pseudo-papilledema vs. true papilledema.
This PDF is available to Subscribers Only