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

Introduction

The diagnosis of neuro-ophthalmic disorders represents one of the most intellectually demanding intersections of neurology, ophthalmology, and internal medicine. The optic nerve, as an extension of the central nervous system, provides a unique window into intracranial pathology, but interpreting its subtle signs requires a high degree of subspecialty expertise [ 1 – 4 ]. Conditions such as papilledema, optic neuritis (ON), and non-arteritic anterior ischemic optic neuropathy (NAION) often manifest with overlapping clinical signs—primarily a swollen or pale optic disc—and non-specific symptoms like blurred vision or visual field loss. The correct interpretation of these findings is paramount, as the underlying etiologies, management strategies, and prognoses are vastly different [ 3 , 5 – 9 ].

Papilledema, defined as optic disc swelling secondary to elevated intracranial pressure (ICP), is a true neuro-ophthalmic emergency. It can be the presenting sign of a brain tumor, cerebral venous sinus thrombosis, or idiopathic intracranial hypertension (IIH), conditions that require urgent investigation and intervention to prevent permanent, devastating vision loss and to address the life-threatening underlying cause [ 10 , 11 ]. The clinical challenge is twofold: first, to reliably detect the presence of optic disc swelling, and second, to differentiate true papilledema from a host of mimickers, collectively known as pseudopapilledema (e.g., optic disc drusen, anomalous discs), which are benign and require no treatment [ 12 , 13 ].

Similarly, the acute, painful vision loss characteristic of typical optic neuritis, often the first demyelinating event in multiple sclerosis (MS), must be distinguished from the painless, sudden vision loss of NAION, the most common acute optic neuropathy in individuals over 50 [ 3 – 5 , 14 ]. The former often warrants immunomodulatory treatment and a full neurological workup for MS, while the latter has no proven acute treatment and requires management of vascular risk factors [ 6 , 15 ]. Misdiagnosis in this context is a significant concern, with reported error rates ranging from 15% in specialized centers to as high as 59.8% in less experienced hands [ 2 ]. Such errors can lead to a cascade of negative consequences: unnecessary and costly investigations like lumbar punctures and MRIs, inappropriate and potentially harmful treatments such as high-dose steroids for a condition that will not respond, and, most critically, delayed diagnosis of the correct underlying condition, leading to adverse patient outcomes [ 5 ].

The scarcity of trained neuro-ophthalmologists worldwide exacerbates this diagnostic challenge, creating a bottleneck that leads to diagnostic delays and limits access to expert care. This is particularly acute in emergency departments, primary care settings, and in underserved regions, where the initial evaluation of patients with neuro-ophthalmic complaints often falls to clinicians with limited training in detailed ophthalmoscopy [ 16 , 17 ].

In recent years, the convergence of medical imaging and artificial intelligence (AI), particularly deep learning (DL), has begun to offer a powerful new paradigm for addressing these long-standing challenges [ 18 ]. Ophthalmology has been at the forefront of this revolution, with landmark studies demonstrating the ability of DL algorithms to detect diabetic retinopathy and glaucomatous optic neuropathy from retinal fundus photographs with an accuracy that is comparable, and in some cases superior, to that of human experts [ 19 – 29 ]. These successes have provided a robust proof-of-concept, paving the way for exploring AI’s potential in the more nuanced and complex field of neuro-ophthalmology [ 16 , 30 ]. AI offers the promise of an objective, scalable, and accessible diagnostic tool that can augment the capabilities of human clinicians, helping to screen, diagnose, and monitor neuro-ophthalmic diseases more effectively.

Early forays into applying AI to neuro-ophthalmology have shown significant potential. The BONSAI (Brain and Optic Nerve Study with Artificial Intelligence) consortium, in a seminal New England Journal of Medicine paper, demonstrated that a DL system could accurately detect papilledema from fundus photographs with high sensitivity and specificity across a large, multi-ethnic, international population [ 31 ]. Subsequent studies have validated this system in diverse and challenging clinical settings, including emergency departments using nonmydriatic cameras operated by non-specialists, and have extended its application to pediatric populations where diagnosis is even more challenging [ 32 , 33 ]. Beyond simple detection of swelling, AI models are being developed to tackle the core clinical problem of differential diagnosis: distinguishing between various causes of optic disc swelling, such as papilledema and pseudopapilledema, and differentiating between various types of optic neuropathy using the rich, quantitative data from optical coherence tomography (OCT) and fundus images [ 12 , 34 – 36 ].

Despite this rapid progress, the clinical readiness of AI in neuro-ophthalmology is still in its nascent stages. Many published studies are retrospective, utilize highly curated and often small datasets, and lack rigorous external validation. Furthermore, the inherent “black box” nature of many complex DL models remains a significant barrier to clinical trust and widespread adoption [ 2 , 37 ]. Before these tools can be responsibly integrated into patient care, a clear-eyed assessment of the current evidence is required. This systematic review aims to provide a comprehensive and critical overview of the current state of AI for the diagnosis of papilledema and optic neuritis. We will synthesize the evidence on the diagnostic accuracy of these models, describe the imaging modalities and AI techniques employed, and identify the key challenges, limitations, and future directions for the successful and equitable integration of AI into the practice of clinical neuro-ophthalmology.

Methods

Search strategy and selection criteria

This systematic review was conducted and reported in accordance with the Preferred Reporting Items for a Systematic Review and Meta-analysis of Diagnostic Test Accuracy Studies (PRISMA-DTA) guidelines. A comprehensive and systematic search of three major electronic databases—PubMed, Embase, and the Cochrane Library—was performed to identify all relevant studies published up to December 31, 2025. The search strategy was designed to be broad and inclusive, combining Medical Subject Headings (MeSH) terms and free-text keywords related to artificial intelligence, neuro-ophthalmology, and the specific disorders of interest. The core search query, adapted for each database, was: (“artificial intelligence” OR “deep learning” OR “machine learning” OR “convolutional neural network”) AND (“neuro-ophthalmology” OR “optic neuritis” OR “papilledema” OR “optic disc swelling” OR “optic neuropathy” OR “ischemic optic neuropathy”) AND (“diagnosis” OR “detection” OR “classification” OR “differentiation”) [ 38 ].

We included original research articles that developed or validated an AI model for the diagnosis, classification, or differentiation of papilledema or optic neuritis using any relevant imaging modality (e.g., color fundus photography, optical coherence tomography (OCT), OCT angiography (OCT-A), or magnetic resonance imaging (MRI)). Studies were required to report sufficient data to allow for the assessment of diagnostic accuracy, including but not limited to sensitivity, specificity, and the area under the receiver operating characteristic curve (AUC). We placed no restrictions on study design, including both retrospective and prospective studies. We excluded case reports, narrative reviews, editorials, conference abstracts without full-text publications, and studies not published in the English language. Two reviewers (S.F. and S.C.) independently screened the titles and abstracts of all retrieved articles. The full texts of potentially eligible articles were then retrieved and assessed against the inclusion criteria. Any disagreements regarding study inclusion were resolved by consensus or, if necessary, by a third senior reviewer [ 38 ].

Data extraction and quality assessment

For each included study, a standardized data extraction form was used to collect the following information: first author, year of publication, country of origin, study design (retrospective or prospective), population characteristics (sample size, demographics, clinical setting), imaging modality or modalities used, details of the AI model architecture (e.g., ResNet, Inception, U-Net, custom CNN), the reference standard used for establishing the ground-truth diagnosis (e.g., expert consensus, clinical diagnosis with follow-up, biopsy), and the key diagnostic accuracy metrics reported (AUC, sensitivity, specificity, accuracy, precision, F1-score). Data were extracted by one reviewer and independently verified for accuracy and completeness by the second reviewer.

The methodological quality of the included studies was independently assessed by two reviewers using the Quality Assessment of Diagnostic Accuracy Studies 2 (QUADAS-2) tool [ 39 , 40 ]. This widely used tool is specifically designed for systematic reviews of diagnostic accuracy studies and evaluates the risk of bias and concerns regarding applicability across four key domains: (1) patient selection, (2) index test (the AI model), (3) reference standard, and (4) flow and timing of the study. Each domain was judged as having a low, high, or unclear risk of bias and applicability. Disagreements were resolved through discussion and consensus.

Data synthesis

Given the anticipated heterogeneity in study populations, AI model architectures, imaging modalities, and statistical reporting, a formal quantitative meta-analysis was deemed inappropriate for the main analysis. Instead, a qualitative synthesis and narrative review of the findings was performed. We first categorized the studies based on the primary neuro-ophthalmic condition addressed (papilledema or optic neuritis/optic neuropathies). Within these categories, we further stratified the studies by the primary imaging modality used. The performance of different AI models was summarized in tables and compared narratively. We systematically compiled and discussed key challenges, limitations of the current evidence, and promising areas for future research as identified in the literature and our own analysis. Where multiple studies evaluated the same AI system (e.g., the BONSAI DLS) or addressed a similar clinical question, we highlighted consistent findings and any important discrepancies. This approach allowed for a comprehensive and nuanced overview of the state of the field.

Results

Study selection and characteristics

Our initial database search across PubMed, Embase, and the Cochrane Library yielded 854 articles. After the removal of 242 duplicates, 612 unique records remained for title and abstract screening. During this phase, 544 records were excluded as they were clearly irrelevant to the scope of this review (e.g., focused on non-diagnostic applications of AI, different clinical domains, or were non-original research). This left 68 full-text articles that were assessed for eligibility. Of these, 36 were excluded for reasons including: being review articles or editorials ( n = 15), lacking sufficient data to extract diagnostic accuracy metrics ( n = 12), using private datasets without clear validation ( n = 5), and focusing on pathologies outside the review’s scope ( n = 4). Ultimately, 32 studies met our stringent inclusion criteria and were included in the final qualitative synthesis. The complete study selection process is detailed in the PRISMA flow diagram (Fig. 1 ).

Of the 32 included studies, a significant majority ( n = 21, 65 [ 39 , 40 ] 0.6%) focused on the application of AI for the diagnosis and assessment of papilledema. The remaining 11 studies (34.4%) addressed optic neuritis and the differentiation of various optic neuropathies. Fundus photography was the predominant imaging modality in the papilledema studies ( n = 18, 85.7%), reflecting its widespread availability. In contrast, OCT was the more common modality in studies on optic neuritis and its sequelae ( n = 9, 81.8%), owing to its ability to quantify axonal and neuronal loss. Most studies were retrospective in design ( n = 28, 87.5%), leveraging existing image databases. The QUADAS-2 assessment highlighted a prevalent high risk of bias in the patient selection domain across many studies ( n = 22, 68.8%). This was primarily due to the use of curated, high-quality image datasets, which may not accurately represent the spectrum of image quality and patient presentation in real-world clinical practice. Concerns regarding applicability were also frequently noted, particularly for models trained and validated on data from a single center or a genetically homogeneous population, limiting their generalizability (Fig. 2 ).

AI for the diagnosis of papilledema

The application of AI to papilledema is the most mature area of research within AI-driven neuro-ophthalmology. We have summarized the key studies in Table 1 .

Fundus photography-based models: a new standard for screening

The development of DL models for papilledema detection from fundus photographs represents a major breakthrough. The foundational BONSAI study by Milea et al. utilized a massive dataset of over 14,000 images from 19 sites across 11 countries to train and validate a DL system [ 31 ]. The model demonstrated exceptional performance, achieving an AUC of 0.99 in the internal validation set. More importantly, in a rigorous external testing phase on a completely separate dataset, the system maintained high accuracy with an AUC of 0.96, a sensitivity of 96.4%, and a specificity of 84.7% for detecting papilledema [ 31 ]. This large-scale, multi-ethnic validation set a high standard for the field.

Subsequent research has focused on translating this high performance into clinically relevant scenarios. A pivotal follow-up study by Biousse et al. applied the BONSAI system to nonmydriatic fundus photographs obtained by technicians in a busy emergency department (the FOTO-ED study) [ 32 , 41 , 45 ]. This was a critical test of real-world applicability, as nonmydriatic images are often of lower quality. The study found that the AI system could reliably identify papilledema and normal optic discs, showing its potential as a robust screening tool even under challenging imaging conditions [ 32 , 41 , 45 ]. The system has also been successfully adapted for pediatric populations, where the differential diagnosis of optic disc swelling is particularly challenging. Lin et al. showed the BONSAI system could accurately identify pediatric papilledema with an AUC of 0.98 [ 33 ], while Chang et al. developed a model specifically to differentiate true papilledema from pseudopapilledema in children, achieving an AUC of 0.94 [ 42 ].

Two recent systematic reviews and meta-analyses have solidified the evidence base. The comprehensive review by Łajczak et al., encompassing 21 studies, reported a remarkable pooled sensitivity of 0.97 and specificity of 0.98, concluding that DL models often meet or exceed the performance of human experts in this specific task [ 39 – 43 , 45 ]. A complementary meta-analysis by Karkhur et al. found a pooled sensitivity of 94.6% and specificity of 90.3%, with an overall AUC of 0.94, further confirming the high diagnostic accuracy and clinical potential of these systems [ 44 ] (Fig. 3 ).

OCT-based models: adding quantitative depth

While fundus photography is excellent for detecting the qualitative sign of disc swelling, OCT provides objective, quantitative measurements of the optic nerve head and surrounding RNFL, offering potential for more nuanced assessment and monitoring [ 1 ]. AI models using OCT data have demonstrated high accuracy in distinguishing clear papilledema from normal optic discs, with Li et al. reporting an AUC of 0.96 for risk-stratification using 3D OCT volumes [ 46 ].

A key clinical challenge, however, is differentiating mild papilledema from pseudopapilledema, particularly buried optic disc drusen, which can mimic true disc swelling. Studies leveraging enhanced depth imaging (EDI-OCT) to visualize these drusen have shown that while Deep Learning Systems (DLS) can aid in this differentiation, their performance (AUCs in the range of 0.85–0.92) is more modest compared to the near-perfect detection of moderate-to-severe papilledema. For context, in one of the key studies, the human expert AUC for the same task was approximately 0.94, highlighting that this remains a challenging area for both human and machine [ 12 , 13 ]. This remains a frontier for AI development. Multimodal models that are being developed to combine the qualitative features from fundus photography with the quantitative data from OCT are an active and promising area of research that may hold the key to improving accuracy in these diagnostically ambiguous cases [ 35 , 47 ].

AI for the diagnosis of optic neuritis and differentiation of optic neuropathies

The application of AI in optic neuritis is inherently more complex. The diagnosis is often clinical, and the imaging challenge lies in differentiating it from other optic neuropathies and identifying its underlying etiology (e.g., multiple sclerosis, MOGAD, or NMOSD) (Table 2 ).

Detection of chronic optic neuritis through atrophy patterns

A history of ON typically leaves a signature of chronic RNFL and ganglion cell layer thinning [ 34 ]. AI models, particularly those using OCT data, have proven highly effective at detecting this pattern of atrophy. Studies have consistently demonstrated that DL algorithms can identify eyes with a history of ON based on RNFL thickness patterns with AUCs exceeding 0.95 [ 50 , 51 ]. For instance, Motamedi et al. developed a DL network using peripapillary OCT ring scans that could detect prior ON with an AUC of 0.96, outperforming conventional analysis of average pRNFL thickness [ 51 ]. This has significant implications for clinical trials and for diagnosing MS, where a history of subclinical ON is a key diagnostic criterion. These findings are consistent with recent studies demonstrating the utility of AI-enhanced OCT analysis in MS diagnosis [ 48 , 49 , 52 ].

Differentiating acute optic neuropathies: the core challenge

Differentiating acute ON from NAION is a critical and often difficult clinical task, as both can present with acute vision loss and a swollen optic disc. The ONION (Optic Neuritis vs. Ischemic Optic Neuropathy) study by Liu et al. represents a significant step forward in addressing this challenge [ 5 , 53 ]. They developed a DL system using only fundus photographs to distinguish between these two conditions in the acute phase. The system achieved an impressive AUC of 0.90, with performance comparable to a retinal expert and superior to general ophthalmologists, suggesting that subtle funduscopic features invisible to the human eye can be learned by the AI [ 5 , 53 ].

Other studies have utilized the volumetric data from 3D OCT scans to classify different forms of optic atrophy [ 19 – 23 ]. A study by Szanto et al. used a 3D CNN to distinguish between glaucoma, NAION, and ON, achieving an overall accuracy of 88.9% [ 35 , 36 ]. This work is important as it shows that AI can learn disease-specific signatures of optic nerve head structural change, moving beyond simple thickness measurements and toward a more sophisticated, pattern-based diagnosis. However, it must be emphasized that this classification is based on patterns of atrophy and does not represent a de novo diagnosis of NAION, which remains a clinical diagnosis of exclusion. A full workup to rule out compressive, infiltrative, or inflammatory causes is still paramount.

Gaps in the literature: the missing link for acute ON

Despite these advances, a critical gap remains in the literature: the lack of AI models for the diagnosis of acute optic neuritis, especially in cases where the optic disc appears normal (retrobulbar ON). The diagnosis of acute ON is a clinical synthesis of patient history (e.g., pain with eye movements), demographics, and examination findings, with imaging often serving to rule out other causes. To date, no robust AI models have been developed to integrate these crucial clinical features with imaging data to assist in the initial diagnosis of a patient presenting with suspected acute ON. This represents a major and necessary frontier for future research. Furthermore, an AI system focused solely on optic disc swelling would inherently miss other significant optic neuropathies that can present with a normal or atrophic disc, such as traumatic optic neuropathy, posterior ischemic optic neuropathy, and certain compressive lesions. This highlights the limitation of a single-modality, single-feature approach and underscores the need for more holistic AI systems.

Discussion

This systematic review synthesizes the burgeoning evidence on the application of AI for the diagnosis of two cardinal neuro-ophthalmic conditions: papilledema and optic neuritis. Our findings reveal a field of rapid innovation, where AI, particularly DL, has made substantial and clinically meaningful progress. However, the maturity of the technology and its readiness for clinical deployment vary significantly between different diagnostic tasks, reflecting the complexity and nuance of neuro-ophthalmic diagnosis (Table 3 ).

A central theme emerging from this review, and a point rightly emphasized by our reviewers, is the critical distinction between an AI model’s ability to detect an anatomical sign (e.g., optic disc swelling) and its inability to independently establish a clinical diagnosis (e.g., papilledema). Papilledema is, by definition, optic disc swelling secondary to elevated intracranial pressure, a diagnosis that requires clinical correlation and the exclusion of numerous mimickers. Current AI systems are powerful feature detectors, not independent diagnosticians. This review will discuss the capabilities and limitations of these models with this crucial framework in mind.

AI as a powerful tool for papilledema detection: a paradigm shift in screening

The detection of papilledema from fundus photographs is, without question, the most significant success story for AI in neuro-ophthalmology to date. The consistently high sensitivity and specificity reported in multiple large-scale, multi-ethnic studies, and confirmed in subsequent meta-analyses, suggest that these tools are not just promising but are approaching the level of a robust, validated screening device [ 31 , 43 , 44 ]. The ability of these systems to maintain high performance even with nonmydriatic, lower-quality images obtained by non-expert technicians in chaotic emergency department settings is particularly impactful [ 32 , 41 , 45 ]. This is a critical use case where papilledema is frequently missed, with potentially devastating consequences. AI-powered screening in such settings could function as an intelligent triage system, automatically flagging patients with suspicious optic discs for urgent specialist review, thereby democratizing access to neuro-ophthalmic expertise, reducing diagnostic delays, and fundamentally improving patient safety.

Despite the remarkable accuracy for binary classification (papilledema vs [ 34 ]. normal), significant challenges remain that prevent widespread adoption. The crucial clinical differentiation of mild, early papilledema from pseudopapilledema (such as buried optic disc drusen or tilted discs) remains a formidable hurdle for both clinicians and AI [ 12 ]. While OCT-based models, with their ability to visualize subsurface structures, offer more granular data, they have yet to definitively solve this problem, with AUCs for this specific task being notably lower than for clear-cut cases [ 13 ]. The path forward likely lies in the development of sophisticated multimodal AI systems. These future models will need to integrate the qualitative gestalt of fundus photos, the quantitative precision of OCT (including RNFL thickness, B-scan morphology, and en face imaging of drusen), and perhaps even clinical data points (e.g., patient age, hyperopic refractive error) to improve the accuracy of this crucial differential diagnosis. As rightly pointed out by expert reviewers, a complete clinical assessment often requires further investigations, such as orbital ultrasound (B-scan) to detect buried optic nerve head drusen, and neuroimaging like MRI to rule out compressive lesions. Future AI systems must be designed to integrate these findings to provide a truly comprehensive diagnostic suggestion, and safe, diagnostic support.

The evolving and complex role of AI in optic neuritis and optic neuropathies

In stark contrast to the relatively straightforward task of papilledema detection, the role of AI in optic neuritis is far more complex and less mature. The primary strength of current AI models lies in detecting the chronic sequelae of ON—namely, the measurable RNFL and ganglion cell layer atrophy that follows an acute inflammatory event [ 50 , 51 ]. This capability has clear utility in monitoring disease progression in conditions like multiple sclerosis and in objectively confirming a suspected history of ON, which can be a key component of MS diagnosis. However, this addresses the aftermath of the disease, not the acute diagnostic challenge.

The development of the ONION system, which can differentiate acute ON from NAION using only fundus photos, is a significant and promising step forward [ 5 , 53 ]. This study is important because it proves that DL models can learn to recognize the subtle, often overlapping, funduscopic patterns of disc swelling that distinguish different acute optic neuropathies. Similarly, the use of 3D OCT scans to classify different patterns of optic atrophy demonstrates the potential of AI to move beyond simple thickness measurements and toward a more sophisticated, etiology-based diagnosis [ 35 , 36 ].

The major and most critical gap, as identified in our review, is the absence of AI tools designed to assist in the initial diagnosis of a patient presenting with symptoms of acute ON, especially in the 70% of cases where the optic disc appears normal (retrobulbar ON). This diagnosis is a clinical synthesis, heavily reliant on patient history (e.g., orbital pain exacerbated by eye movements), demographics, and specific examination findings (e.g., a relative afferent pupillary defect). To date, no robust AI models have been developed to integrate these crucial clinical features with imaging data. This represents the next frontier for AI in neuro-ophthalmology: the creation of true clinical decision support tools that can process and weigh multiple data streams to provide a probabilistic diagnosis and guide management.

Overarching challenges and a roadmap for the future

For AI to transition from a promising research tool to an indispensable clinical partner, several fundamental challenges must be addressed.

First, the “black box” problem remains a primary barrier to clinician trust and adoption [ 2 , 37 ]. To address this, the field is moving toward more inherently interpretable or “explainable” AI (XAI) frameworks. Rather than relying on post-hoc interpretations, these methods aim to build trust by making the model’s reasoning transparent. Specific XAI techniques currently being tested in neuro-ophthalmology include:

Saliency Maps : These heatmaps visually highlight the specific pixels or regions in an image (e.g., the optic disc margin, a specific blood vessel) that the model “focused” on to make its diagnosis of papilledema. This allows a clinician to quickly verify if the model is looking at clinically relevant features [ 54 ].

Uncertainty Quantification : Instead of a simple binary output, these models provide a “confidence score” alongside their diagnosis. A low confidence score can automatically trigger a human review, creating a safer human-in-the-loop system [ 55 ].

Clinicians need to understand the ‘why’ behind an AI recommendation to feel confident in acting upon it, especially in high-stakes decisions, and these XAI techniques are a critical step in that direction.

Second, the issue of data heterogeneity, bias, and generalizability is paramount. The performance of a Deep Learning System is only as good as the data it is trained on. Our review identified several specific examples of “real-world failures” where model performance degraded significantly when applied outside of the initial training conditions:

Poor Image Quality : Models trained on high-quality, well-focused images often fail when encountering images with media opacity (e.g., cataracts), poor illumination, or off-center optic discs, as is common in real-world screening settings [ 56 ].

Confounding Pathologies : A model trained to detect papilledema may misclassify other optic disc anomalies, such as tilted discs or optic nerve head drusen, if these were not adequately represented in the training data [ 1 ].

High Myopia : The altered optic disc morphology in highly myopic eyes is a known confounder that can lead to false positives for glaucomatous damage or other optic neuropathies.

Ethnic Cohorts : Optic disc size and pigmentation can vary across different ethnic groups. A model trained primarily on one ethnicity may not generalize well to others without specific fine-tuning and validation.

The path to robust generalizability requires prospective, multi-center validation studies in real-world clinical workflows. Furthermore, we must actively audit these models for biases related to race, ethnicity, and socioeconomic factors to ensure they do not perpetuate existing health disparities.

Third, significant regulatory and implementation hurdles must be cleared. International collaboration is needed to establish clear, standardized guidelines for the validation, approval, and post-market surveillance of AI diagnostic systems [ 16 , 57 ]. Medico-legal questions of liability and responsibility must be addressed. From a practical standpoint, seamless integration into existing electronic health records (EHRs) and picture archiving and communication systems (PACS) is non-negotiable for these tools to be used efficiently in a busy clinical environment.

Future progress will be driven by the development of multimodal AI platforms that can intelligently synthesize information from multiple sources—fundus photos, OCT, OCT angiography, visual fields, MRI, and structured EHR data—to provide a holistic assessment that more closely mirrors the complex reasoning of an expert clinician [ 2 , 35 , 53 ]. Furthermore, the advent of Large Multimodal Models (LMMs) and Generative AI in 2026 opens new frontiers for synthesizing unstructured clinical notes with imaging data, potentially closing the gap in diagnosing conditions like acute retrobulbar ON where the initial imaging may be normal. Techniques like federated learning, which allows models to be trained across multiple institutions without the need to share sensitive patient data, offer a powerful solution to the challenge of building large, diverse datasets, especially for rare diseases [ 2 , 7 – 9 , 58 ].

Conclusion

Artificial intelligence has unequivocally demonstrated its potential to be a transformative force in the diagnosis of neuro-ophthalmic disorders. For the detection of papilledema from fundus photographs, AI models have achieved a level of accuracy and robustness that strongly supports their use as powerful screening and triage tools, with the potential to significantly reduce diagnostic error and improve access to care. In the more complex arena of optic neuritis and other optic neuropathies, AI is proving valuable for detecting chronic disease-related changes and is beginning to show promise in the critical task of differentiating between acute conditions. However, the field is still in its adolescence. The successful and responsible translation of these promising technologies from research papers to the patient bedside will depend on a concerted effort from clinicians, computer scientists, and regulatory bodies. We must focus on creating robust, explainable, and multimodal AI systems that are rigorously and prospectively validated in diverse, real-world settings. Only then can we ensure that these powerful tools are not only accurate but also safe, equitable, and truly beneficial for the patients we serve.

Author contributions

S.-S.F. and S.-H.C. conceived and designed the study. S.-S.F. conducted the literature search and data extraction. S.-H.C. performed the quality assessment. S.-S.F. wrote the main manuscript text and prepared all figures and tables. S.-H.C. provided critical revisions. All authors reviewed and approved the final manuscript.

Data availability

All data generated or analysed during this study are included in this published article and its supplementary information files.

Declarations

Ethics and consent to participate

Ethics and Consent to Participate declarations: not applicable. This study is a systematic review of previously published literature and did not involve human participants, human data, or human tissue.

Competing interests

The authors declare no competing interests.

Footnotes

Publisher’s note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

References

1. Liu TYA, Ting DSW, Yi PH, et al. Detection of optic disc abnormalities in color fundus photographs using deep learning. J Neuroophthalmol. 2021;41(3):368–74. 10.1097/WNO.0000000000001358.

2. Karkhur S, Beri A, Verma V, et al. Artificial Intelligence in neuro-ophthalmology: opportunities for the diagnosis of optic neuropathies and visual pathway disorders. Cureus. 2025;17(8):e90142. 10.7759/cureus.90142.

3. Toosy AT, Mason DF, Miller DH. Optic neuritis. Lancet Neurol. 2014;13(1):83–99. 10.1016/S1474-4422(13)70259-X.

4. Biousse V, Newman NJ. Ischemic optic neuropathies. N Engl J Med. 2015;372(25):2428–36. 10.1056/NEJMra1413352.

5. Liu K, Liu S, Tan X, et al. Deep learning system for distinguishing optic neuritis from non-arteritic anterior ischemic optic neuropathy at acute phase based on fundus photographs. Front Med (Lausanne). 2023;10:1188542. 10.3389/fmed.2023.1188542.

6. Berry S, Lin WV, Sadaka A, Lee AG. Nonarteritic anterior ischemic optic neuropathy: cause, effect, and management. Eye Brain. 2017;9:23–8. 10.2147/EB.S125311.

7. Li F, Wang Z, Qu G, et al. Automatic differentiation of Glaucoma visual field from non-glaucoma visual filed using deep convolutional neural network. BMC Med Imaging. 2018;18(1):35. 10.1186/s12880-018-0273-5.

8. Kucur SS, Hollo G, Sznitman R. A deep learning approach to automatic detection of early glaucoma from visual fields. PLoS ONE. 2018;13(11):e0206081. 10.1371/journal.pone.0206081.

9. Wang M, Shen LQ, Pasquale LR, et al. An artificial intelligence approach to holistic visual field interpretation. Transl Vis Sci Technol. 2020;9(2):59. 10.1167/tvst.9.2.59.

10. Hayreh SS. Pathogenesis of optic disc edema in raised intracranial pressure. Prog Retin Eye Res. 2016;50:108–44. 10.1016/j.preteyeres.2015.10.001.

11. Rigi M, Almarzouqi SJ, Morgan ML, Lee AG. Papilledema: epidemiology, etiology, and clinical management. Eye Brain. 2015;7:47–57. 10.2147/eb.s69174.

12. Lee KM, Woo SJ, Hwang JM. Differentiation of optic nerve head drusen and optic disc edema with spectral-domain optical coherence tomography. Ophthalmology. 2011;118(5):971–7. 10.1016/j.ophtha.2010.09.006.

13. Malmqvist L, Bursztyn L, Costello F, et al. The optic disc drusen studies consortium recommendations for diagnosis of optic disc drusen using optical coherence tomography. J Neuroophthalmol. 2018;38(3):299–307. 10.1097/WNO.0000000000000585.

14. Hayreh SS. Ischemic optic neuropathies - where are we now? Graefes Arch Clin Exp Ophthalmol. 2013;251(8):1873–84. 10.1007/s00417-013-2399-z.

15. Beck RW, Cleary PA, Anderson MM Jr, et al. A randomized, controlled trial of corticosteroids in the treatment of acute optic neuritis. The Optic Neuritis Study Group. N Engl J Med. 1992;326(9):581–8. 10.1056/NEJM199202273260901.

16. Kenney RC, Requarth TW, Jack AI, et al. AI in neuro-ophthalmology: current practice and future opportunities. J Neuroophthalmol. 2024;44(3):308–18. 10.1097/wno.0000000000002205.

17. Bruce BB, Thulasi P, Fraser CL, et al. Diagnostic accuracy and use of nonmydriatic ocular fundus photography by emergency physicians: phase II of the FOTO-ED study. Ann Emerg Med. 2013;62(1):28–e331. 10.1016/j.annemergmed.2013.01.010.

18. Abràmoff MD, Lavin PT, Birch M, et al. Pivotal trial of an autonomous AI-based diagnostic system for detection of diabetic retinopathy in primary care offices. NPJ Digit Med. 2018;1:39. 10.1038/s41746-018-0040-6.

19. Christopher M, Bowd C, Belghith A, et al. Deep learning approaches predict glaucomatous visual field damage from OCT optic nerve head en face images and retinal nerve fiber layer thickness maps. Ophthalmology. 2020;127(3):346–56. 10.1016/j.ophtha.2019.09.036.

20. Thompson AC, Falconi A, Sappington RM. Deep learning and optical coherence tomography in glaucoma: bridging the diagnostic gap on structural imaging. Front Ophthalmol. 2022;2:937205. 10.3389/fopht.2022.937205.

21. Katuru A, Nair AG, Srinivasan K, et al. Deep learning with disc photos or OCT scans in glaucoma detection. Ophthalmol Sci. 2025;8(4):389–98. 10.1016/j.xops.2025.100877.

22. Yousefi S, Elze T, Guo J, et al. An artificial intelligence enabled system for retinal nerve fiber layer thickness-based glaucoma severity classification. Ophthalmol Sci. 2023;32(10):e1–10. 10.1016/j.xops.2023.100389.

23. Medeiros FA. The definition of glaucomatous optic neuropathy in artificial intelligence research and clinical applications. Ophthalmol Glaucoma. 2023;6(4):432–8. 10.1016/j.ogla.2023.01.007.

24. Medeiros FA, Jammal AA, Thompson AC. From machine to machine: an OCT-trained deep learning algorithm for objective quantification of glaucomatous damage in fundus photographs. Ophthalmology. 2019;126(4):513–21. 10.1016/j.ophtha.2018.12.033.

25. Gulshan V, Peng L, Coram M, et al. Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs. JAMA. 2016;316(22):2402–10. 10.1001/jama.2016.17216.

26. Ting DSW, Cheung CY, Lim G, et al. Development and validation of a deep learning system for diabetic retinopathy and related eye diseases using retinal images from multiethnic populations with diabetes. JAMA. 2017;318(22):2211–23. 10.1001/jama.2017.18152.

27. Ting DSW, Pasquale LR, Peng L, et al. Artificial intelligence and deep learning in ophthalmology. Br J Ophthalmol. 2019;103(2):167–75. 10.1136/bjophthalmol-2018-313173.

28. Kapoor R, Walters SP, Al-Aswad LA. The current state of artificial intelligence in ophthalmology. Surv Ophthalmol. 2019;64(2):233–40. 10.1016/j.survophthal.2018.09.002.

29. Schmidt-Erfurth U, Sadeghipour A, Gerendas BS, et al. Artificial intelligence in retina. Prog Retin Eye Res. 2018;67:1–29. 10.1016/j.preteyeres.2018.07.004.

30. Leong YY, Vasseneix C, Finkelstein MT, et al. Artificial intelligence meets neuro-ophthalmology. Asia Pac J Ophthalmol (Phila). 2022;11(2):111–25. 10.1097/APO.0000000000000512.

31. Milea D, Najjar RP, Jiang Z, et al. Artificial intelligence to detect papilledema from ocular fundus photographs. N Engl J Med. 2020;382(18):1687–95. 10.1056/NEJMoa1917130.

32. Biousse V, Najjar RP, Tang Z, et al. Application of a deep learning system to detect papilledema on nonmydriatic ocular fundus photographs in an emergency department. Am J Ophthalmol. 2024;261:199–207. 10.1016/j.ajo.2023.10.025.

33. Lin MY, Chang MY, Biousse V, et al. The BONSAI (Brain and Optic Nerve Study with Artificial Intelligence) deep learning system can accurately identify pediatric papilledema on standard ocular fundus photographs. J AAPOS. 2024;28(1):103803. 10.1016/j.jaapos.2023.10.005.

34. Razaghi G, Hedayati E, Hejazi M, et al. Measurement of retinal nerve fiber layer thickness with a deep learning algorithm in ischemic optic neuropathy and optic neuritis. Sci Rep. 2022;12(1):17109. 10.1038/s41598-022-22135-x.

35. Szanto D, Wang JK, Woods B, et al. Deep learning differentiates papilledema, NAION, and healthy eyes with unsegmented 3D OCT volumes. Am J Ophthalmol. 2025;277:249–59. 10.1016/j.ajo.2025.05.036.

36. Poonja S, Milea D, Chen JJ. 3D deep learning analysis of OCT to classify optic atrophy. JAMA Ophthalmol. 2025;143(10):811–2. 10.1001/jamaophthalmol.2025.2829.

37. Rudin C. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nat Mach Intell. 2019;1(5):206–15. 10.1038/s42256-019-0048-x.

38. Petzold A, Fraser CL, Abegg M, et al. Diagnosis and classification of optic neuritis. Lancet Neurol. 2022;21(12):1120–34. 10.1016/S1474-4422(22)00200-9.

39. Anandi L, Budihardja BM, Anggraini E, et al. The use of artificial intelligence in detecting papilledema from fundus photographs. Taiwan J Ophthalmol. 2023;13(2):184–90. 10.4103/tjo.tjo-d-22-00178.

40. Salaheldin AM, Abdel Wahed M, Talaat M, Saleh N. An evaluation of AI-based methods for papilledema detection in retinal fundus images. Biomed Signal Process Control. 2024;92:106120. 10.1016/j.bspc.2024.106120.

41. Lamirel C, Bruce BB, Wright DW, et al. Quality of nonmydriatic digital fundus photography obtained by nurse practitioners in the emergency department: the FOTO-ED study. Ophthalmology. 2012;119(3):617–24. 10.1016/j.ophtha.2011.09.013.

42. Chang MY, Biousse V, Newman NJ, et al. Artificial intelligence to differentiate pediatric papilledema from pseudopapilledema. Ophthalmol Sci. 2024;4(4):100496. 10.1016/j.xops.2024.100496.

43. Łajczak PM, Sirek S, Wyględowska-Promieńska D. Unveiling AI’s role in papilledema diagnosis from fundus images: a systematic review with diagnostic test accuracy meta-analysis and comparison of human expert performance. Comput Biol Med. 2025;184:109350. 10.1016/j.compbiomed.2024.109350.

44. Karkhur S, Singh P, Verma V, et al. Diagnostic accuracy of artificial intelligence for the detection of papilledema on fundus images: a systematic review and meta-analysis. Cureus. 2025;17(12):e99135. 10.7759/cureus.99135.

45. Browning SD, Costello JM, Dunn HP, Fraser CL. The use of fundus photography in the emergency room—a narrative review. Curr Neurol Neurosci Rep. 2025;25(1):30. 10.1007/s11910-025-01417-7.

46. Li A, Tandon AK, Sun G, et al. Early detection of optic nerve changes on optical coherence tomography using deep learning for risk-stratification of papilledema and glaucoma. J Neuroophthalmol. 2024;44(1):47–52. 10.1097/WNO.0000000000001945.

47. Frisen L. Swelling of the optic nerve head: a staging scheme. J Neurol Neurosurg Psychiatry. 1982;45(1):13–8. 10.1136/jnnp.45.1.13.

48. Ortiz M, Mallen V, Boquete L, et al. Diagnosis of multiple sclerosis using optical coherence tomography supported by artificial intelligence. Mult Scler Relat Disord. 2023;74:104725. 10.1016/j.msard.2023.104725.

49. Pérez del Palomar A, Cegoñino J, Montolío A, et al. Swept source optical coherence tomography to early detect multiple sclerosis disease. The use of machine learning techniques. PLoS ONE. 2019;14(5):e0216410. 10.1371/journal.pone.0216410.

50. Kenney RC, Requarth TW, Jack AI, et al. The role of optical coherence tomography criteria and machine learning in multiple sclerosis and optic neuritis diagnosis. Neurology. 2022;99(11):e1100–12. 10.1212/WNL.0000000000200883.

51. Motamedi S, Oertel FC, Engel L, et al. Prior optic neuritis detection on peripapillary ring scans using deep learning. Ann Clin Transl Neurol. 2022;9(11):1682–91. 10.1002/acn3.51632.

52. Chiang CYN, Milea D, Girard MJA. Empowering optical coherence tomography with AI: a new era for neuro-ophthalmic and neurological disorders. OCT and Imaging in Central Nervous System Diseases. Springer; 2025. pp. 699–736. 10.1007/978-3-031-72156-4_26.

53. Bunod R, Augstburger E, Brasnu E, et al. A deep learning system using optical coherence tomography angiography to detect glaucoma and anterior ischemic optic neuropathy. J Clin Med. 2023;12(2):507. 10.3390/jcm12020507.

54. Holzinger A, Langs G, Denk H, et al. Causability and explainability of artificial intelligence in medicine. Wiley Interdiscip Rev Data Min Knowl Discov. 2019;9(4):e1312. 10.1002/widm.1312.

55. Dongil-Moreno FJ, Ortiz M, Pueyo A, et al. Diagnosis of multiple sclerosis using optical coherence tomography supported by explainable artificial intelligence. Eye (Lond). 2024;38(8):1502–8. 10.1038/s41433-024-02933-5.

56. Nam Y, Kim Y, Kim J, et al. Deep learning-based optic disc classification is affected by optic-disc tilt. Sci Rep. 2024;14(1):1234. 10.1038/s41598-023-50256-4.

57. Lee R, Ong J, Waisberg E, et al. Applications of artificial intelligence in neuro-ophthalmology: neuro-ophthalmic imaging patterns and implementation challenges. Neuro-Ophthalmology. 2025;49(4):273–84. 10.1080/01658107.2025.2459206.

58. Topol EJ. High-performance medicine: the convergence of human and artificial intelligence. Nat Med. 2019;25(1):44–56. 10.1038/s41591-018-0300-7.

Associated Data

Data Availability Statement

All data generated or analysed during this study are included in this published article and its supplementary information files.