AI-assisted diagnosis of neuro-ophthalmic disorders: a systematic review from optic neuritis to papilledema.

backgroundNeuro-ophthalmic disorders like optic neuritis and papilledema are diagnostically challenging, with non-specific symptoms and reliance on expert interpretation of optic nerve head imaging (Liu TYA et al. J Neuroophthalmol. 41(3):368–74, 2021). Misdiagnosis can lead to severe consequences, including irreversible vision loss or failure to detect life-threatening conditions. Artificial intelligence (AI), particularly deep learning (DL), shows promise but its application in this nuanced field is not well-defined.

objectiveThis systematic review evaluates the current state of AI in diagnosing key neuro-ophthalmic disorders, focusing on optic neuritis and papilledema. We aim to: (1) assess the diagnostic accuracy of AI models; (2) categorize the imaging modalities and DL techniques used; and (3) identify critical challenges and future opportunities for AI in clinical neuro-ophthalmology.

methodsA systematic search of PubMed, Embase, and Cochrane Library for studies published up to December 2025 on AI models for diagnosing neuro-ophthalmic conditions using fundus photography or OCT was conducted. We extracted data on study design, AI model architecture, and diagnostic accuracy, performing a qualitative synthesis of the findings.

resultsWe identified 32 eligible studies (Christopher M et al. Ophthalmology 127(3):346–56, 2020, Thompson AC et al. Front Ophthalmol. 2:937205, 2022, Katuru A et al. Ophthalmol Sci. 8(4):389–98, 2025, Yousefi S et al. Ophthalmol Sci. 32(10):e1–10,2023, Medeiros FA et al. Ophthalmol Glaucoma 6(4):432–8, 2023, Razaghi G et al. Sci Rep. 12(1):17109, 2022). Most (66%) focused on papilledema, while 34% addressed optic neuritis. OCT was the most common imaging modality (75%). For papilledema detection from fundus photos, DL models achieved high accuracy (AUC > 0.98). For OCT-based diagnosis, AI models distinguished established papilledema from normal discs well, but performance was modest in differentiating mild papilledema from pseudopapilledema. In optic neuritis, AI models detected chronic RNFL thinning effectively (AUC > 0.95) but struggled to differentiate the pattern of atrophy from other conditions like NAION or glaucoma (AUCs 0.85–0.92). A major gap was the lack of models for diagnosing acute optic neuritis.

conclusionAI shows significant potential as a diagnostic aid in neuro-ophthalmology, particularly in detecting optic disc swelling and chronic optic atrophy. However, the field is in its early stages. Current models excel at binary classification but struggle with the core clinical problem of differential diagnosis. Future progress depends on curating large, multi-center datasets, developing sophisticated multimodal models, and focusing on solving the core clinical problem of differential diagnosis.