Hamann, Steffen

Background Prompt diagnosis of acute central retinal artery occlusion (CRAO) is crucial for therapeutic management and stroke prevention. However, most stroke centers lack onsite ophthalmic expertise before considering fibrinolytic treatment. This study aimed to develop, train, and test a deep learning system to detect hyperacute CRAO on retinal fundus photographs within the critical 4.5‐hour treatment window and up to 24 hours after visual loss to aid in secondary stroke prevention. Methods Our retrospective, cross‐sectional study included 1322 color fundus photographs from 771 patients with acute visual loss due to CRAO, central retinal vein occlusion, nonarteritic 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 included 1039 photographs (517 patients), followed by testing on 2 data sets: (1) hyperacute CRAO (54 photographs, 54 patients) and (2) CRAO within 24 hours after visual loss (110 photographs, 109 patients). Results The deep learning system achieved an area under the receiver operating characteristic curve 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 hyperacute stage, with similar results within 24 hours. The deep learning system outperformed stroke neurologists on a subset of hyperacute testing data set (120 photographs, 120 patients). Conclusions A deep learning system can accurately detect hyperacute CRAO on retinal photographs within a time window compatible with urgent fibrinolysis. If further validated, such systems could improve patient selection for fibrinolytic trials and optimize secondary stroke prevention. Registration URL: https://www.clinicaltrials.gov ; Unique identifier: NCT06390579.

ObjectiveTo compare the diagnostic performance of an artificial intelligence deep learning system with that of expert neuro‐ophthalmologists in classifying optic disc appearance.MethodsThe deep learning system was previously trained and validated on 14,341 ocular fundus photographs from 19 international centers. The performance of the system was evaluated on 800 new fundus photographs (400 normal optic discs, 201 papilledema [disc edema from elevated intracranial pressure], 199 other optic disc abnormalities) and compared with that of 2 expert neuro‐ophthalmologists who independently reviewed the same randomly presented images without clinical information. Area under the receiver operating characteristic curve, accuracy, sensitivity, and specificity were calculated.ResultsThe system correctly classified 678 of 800 (84.7%) photographs, compared with 675 of 800 (84.4%) for Expert 1 and 641 of 800 (80.1%) for Expert 2. The system yielded areas under the receiver operating characteristic curve of 0.97 (95% confidence interval [CI] = 0.96–0.98), 0.96 (95% CI = 0.94–0.97), and 0.89 (95% CI = 0.87–0.92) for the detection of normal discs, papilledema, and other disc abnormalities, respectively. The accuracy, sensitivity, and specificity of the system's classification of optic discs were similar to or better than the 2 experts. Intergrader agreement at the eye level was 0.71 (95% CI = 0.67–0.76) between Expert 1 and Expert 2, 0.72 (95% CI = 0.68–0.76) between the system and Expert 1, and 0.65 (95% CI = 0.61–0.70) between the system and Expert 2.InterpretationThe performance of this deep learning system at classifying optic disc abnormalities was at least as good as 2 expert neuro‐ophthalmologists. Future prospective studies are needed to validate this system as a diagnostic aid in relevant clinical settings. ANN NEUROL 2020;88:785–795