Neuro-ophthalmic disorders, such as optic neuritis and papilledema, present significant diagnostic challenges because they often share non-specific clinical symptoms. Traditionally, diagnosis relies heavily on expert interpretation of optic nerve head (ONH) imaging. Given that misdiagnosis can lead to irreversible vision loss or the failure to identify life-threatening underlying conditions, there is a growing interest in objective diagnostic aids.
AI, specifically deep learning, has shown potential in automating the analysis of fundus photography and optical coherence tomography (OCT). A study recently published in BMC Ophthalmology aimed to clarify the current performance, techniques and limitations of AI models specifically targeting these nuanced neuro-ophthalmic conditions.
Methodology
The researchers conducted a systematic review of studies published up to December 2025. The methodology involved:
-
Search Strategy: A comprehensive search of PubMed, Embase and the Cochrane Library.
-
Inclusion Criteria: Studies were included if they evaluated AI and deep learning models for diagnosing neuro-ophthalmic conditions using fundus photography or OCT.
-
Analysis: The review identified 32 eligible studies. The researchers categorized these studies by the specific disorder (66% focused on papilledema; 34% on optic neuritis), the imaging modality used (OCT was utilized in 75% of the studies) and the diagnostic accuracy, which was measured by area under the curve (AUC).
-
Objectives: The primary goals were to assess diagnostic accuracy, categorize the deep learning techniques employed and identify the clinical gaps that currently prevent widespread adoption.
Results and Implications for Eye Care Practitioners
The findings highlight both the accuracy and the current clinical limitations of AI in neuro-ophthalmology:
-
Papilledema Detection: AI models demonstrated exceptional performance in binary classification. For fundus photography, deep learning models achieved an AUC > 0.98. In OCT analysis, models were highly effective at distinguishing established papilledema from normal discs.
-
Optic Neuritis: Models were effective at detecting chronic retinal nerve fiber layer (RNFL) thinning (AUC > 0.95). However, a significant gap was identified in the diagnosis of acute optic neuritis, for which few models currently exist.
-
Differential Diagnosis Challenges: While AI excels at “normal vs. abnormal” screening, performance was more modest (AUC 0.85–0.92) when tasked with the more complex clinical problem of differential diagnosis—such as distinguishing mild papilledema from pseudopapilledema, or differentiating optic neuritis-related atrophy from glaucoma or non-arteritic anterior ischemic optic neuropathy (NAION).
What this means for practitioners:
For the eye care professional, AI currently serves best as a “red flag” or screening tool for detecting optic disc swelling and chronic atrophy. However, because current models struggle with the differential diagnosis of look-alike conditions, the “human-in-the-loop” remains essential.
Practitioners should view AI as a supportive tool for identifying pathology rather than a replacement for clinical reasoning in complex cases. The field is still in its early stages, and future developments will need to focus on multi-class classification to better assist in the nuanced differential workups required in clinical practice.
Author
