AI-based Detection of Central Retinal Artery Occlusion within 4.5 hours on Standard Fundus Photographs

BackgroundPrompt diagnosis of acute central retinal artery occlusion (CRAO) is crucial for therapeutic management and secondary prevention of associated neurological co-morbidities. However, most stroke centers lack on-site ophthalmic expertise prior to considering fibrinolytic treatments. We aimed to develop, train, and test a deep learning system (DLS) able to accurately detect hyper-acute CRAO on retinal color fundus photographs, during the critical treatment window of 4.5 hours after visual loss. We also evaluated the diagnostic performance of the DLS within 24 hours after visual loss, aiming to improve secondary prevention of stroke after CRAO. MethodsOur retrospective, multicenter, multiethnic study included 1,322 color fundus photographs from 771 patients with various causes of acute visual loss, including CRAO, central retinal vein occlusion, non-arteritic anterior ischemic optic neuropathy, and healthy controls. Photographs were collected from 9 expert neuro-ophthalmology centers in 6 countries, including 3 randomized clinical trials. Training was performed on 1,039 photographs (517 patients), followed by testing on two datasets to discriminate CRAO cases at (i) hyper-acute stage (54 photographs, 54 patients) and (ii) within 24 hours after visual loss (110 photographs, 109 patients). ResultsThe DLS achieved an area under the receiver operating characteristic curve (AUC) of 0.96 (95% confidence interval [CI], 0.95-0.98), a sensitivity of 92.6% (95% CI, 87.0-98.0), and a specificity of 85.0% (95% CI, 81.8-92.8) for detecting CRAO at hyper-acute stage, with similar results for CRAO diagnosis within 24 hours. The DLS outperformed neurologists on a subset of testing dataset at hyper-acute stage (120 photographs from 120 patients). ConclusionsA DLS can accurately detect hyper-acute CRAO on retinal photographs within a time-window compatible with urgent fibrinolysis. Delayed diagnosis (24h) did not alter the ability of the DLS to accurately identify CRAO. If further validated, such systems could improve patient selection for fibrinolytic trials and optimize secondary stroke prevention. Clinical Trial Registration: URL: https://www.clinicaltrials.gov; Unique identifier: NCT06390579.