The deep learning system reported in this article was designed and trained to detect abnormal optic discs using digital colour ocular fundus photographs collected from the Brain and Optic Nerve Study with Artificial Intelligence (BONSAI) international consortium of neuro-ophthalmologists with expertise in optic nerve disorders; the optic disc captured by fundus photographs carries nerve fibres that send visual messages from the eye to the brain. Abnormalities in the optic disc could be a sign of a number of diseased conditions including possible brain conditions, and accurate assessments of ocular fundus photographs are crucial to timely medical interventions to avert potential visual loss and neurologic complications. This work focuses on the detection of papilledema, which is a serious medical condition where the optic nerve becomes swollen due to pressure in or around the brain possibly arising from infection, bleeding or blockage of blood in the brain.
The deep learning system classified images into three categories: (i) those with papilledema; (ii) those with optic disc abnormalities other than papilledema; and (iii) normal optic discs. The deep learning system was trained using 14,341 fundus photographs, and achieved AUC for the detection of papilledema of 0.96 (1 means 100% correct) on a separate testing data set comprising 1,505 photographs from 5 countries (i.e. Thailand, Denmark, Germany, Iran and USA). Interestingly, for 10 of the 177 eyes in the testing data set for which the deep learning system provided a different classification from the original neuro-ophthalmologist, four expert neuro-ophthalmologists not involved in the original analyses reviewed these images, and agreed with the deep learning system; these discrepancies were subsequently verified to be the result of labelling errors by site investigators. To the best of our knowledge, this is the first study showing that artificial intelligence could reliably detect papilledema from fundus photographs images taken at many international centres using a variety of commercially available digital fundus cameras, and independent of the ethnicity and age of the patients.
While this deep learning system was calibrated primarily to identify the papilledema optic disc abnormality, the current success suggests that deep learning systems could be developed for various ophthalmic conditions to augment limited neuro-ophthalmologic expertise for better ophthalmologic care. The IHPC team has secured grant funding from a NMRC Clinician Scientist Award (CSA) to carry out further studies, and is also in discussions with clinicians from Singapore Eye Research Institute on options for translating this technology to real world clinical applications, such as trialling in emergency rooms to detect papilledema in patients with headache and other neurologic complaints.
The article titled "
Artificial Intelligence to Detect Papilledema from Ocular Fundus Photographs" arising from a collaboration with SingHealth Singapore Eye Research Institute (SERI) has been accepted for publication in New England Journal of Medicine (impact factor of 70.670). The current IHPC researchers involved in this work are
Dr Xu Xinxing and
Dr. Liu Yong.
#IHPC #ArtificialIntelligence #AI #AIDeepLearning #Papilledema