Artificial intelligence in ophthalmology: current applications, challenges, and future directions—a narrative review

Introduction

Artificial intelligence (AI) has emerged as a transformative force in healthcare, with particular relevance to ophthalmology—a specialty distinguished by its reliance on imaging and quantitative assessments for diagnosis, monitoring, and treatment planning. The increasing availability of digital retinal imaging, optical coherence tomography (OCT), and visual field testing has resulted in exponential growth in ophthalmic data, creating an ideal environment for the application of machine learning (ML) and deep learning (DL) techniques that excel in image recognition and pattern detection (1).

Background

Ophthalmology has been at the forefront of clinical AI adoption due to the structured and image-centric nature of its diagnostic workflows. In recent years, AI applications in healthcare have gained global attention for their potential to improve diagnostic accuracy, reduce clinician workload, and address gaps in healthcare access (2). Within ophthalmology, this has translated into substantial research and commercial development focused on automating the detection of leading causes of vision loss, including diabetic retinopathy (DR), age-related macular degeneration (AMD), and glaucoma (3-5).

A landmark milestone was achieved in 2018 with the U.S. Food and Drug Administration’s (FDA’s) approval of the first autonomous AI system for DR detection, underscoring the clinical readiness of AI-based diagnostic tools and catalysing broader interest in their deployment across eye care settings (6).

Rationale and knowledge gap

Despite rapid technological advances and promising diagnostic performance reported in controlled studies, the translation of AI systems from research environments into routine ophthalmic practice has been inconsistent (7). While many algorithms demonstrate high accuracy under ideal conditions, their real-world performance often varies due to differences in patient demographics, imaging devices, disease prevalence, and clinical workflows (7).

Moreover, much of the existing literature focuses either on technical algorithm development or provides high-level overviews of AI applications, with limited integration of implementation challenges, ethical considerations, health system impacts, and equity implications. There remains a clear gap for narrative reviews that critically examine not only where AI performs well, but also why adoption remains uneven and how emerging solutions can address persistent structural, regulatory, and clinical barriers.

Objectives

The objective of this narrative review is to provide a clinically relevant synthesis of AI in ophthalmology. Specifically, the review aims to: (I) summarize current AI applications across major ophthalmic disease areas with attention to diagnostic performance and clinical readiness; (II) examine the real-world impact of AI on eye care delivery and health systems; (III) identify key challenges related to implementation, ethics, regulation, and equity; and (IV) outline future directions to support responsible, effective, and equitable integration of AI into ophthalmic practice. We present this article in accordance with the Narrative Review reporting checklist (available at https://aes.amegroups.com/article/view/10.21037/aes-25-61/rc).

Methods

This narrative review synthesizes key developments in AI in ophthalmology, with a focus on clinical applications, performance evidence, implementation challenges, and future directions. A structured literature search was conducted using PubMed, Scopus, and Web of Science to identify relevant peer-reviewed articles published in English between January 2013 and March 2025. Search terms included combinations of “artificial intelligence”, “machine learning”, “deep learning”, “ophthalmology”, “diabetic retinopathy”, “glaucoma”, “age-related macular degeneration”, “optical coherence tomography”, “screening”, and “tele-ophthalmology” (Table 1).

In addition to database searches, reference lists of key review articles and landmark studies were manually screened to identify additional relevant publications. Priority was given to high-quality studies, including large validation studies, randomized or pragmatic trials, systematic reviews and meta-analyses, and regulatory evaluations relevant to real-world implementation. Given the narrative nature of this review, no formal risk-of-bias assessment or quantitative synthesis was performed. Instead, the literature was critically appraised to highlight major trends, strengths, limitations, and gaps in the current evidence base.

This approach was chosen to provide a clinically relevant overview of AI in ophthalmology, while allowing integration of technical, clinical, ethical, and health system perspectives that are often examined separately in more narrowly focused reviews.

Discussion

A brief history of AI in ophthalmology

The application of AI in ophthalmology has evolved in parallel with advances in digital imaging and computational capabilities. In the 1990s and early 2000s, traditional ML algorithms were applied classify fundus images, particularly for DR (8). However, these models were often limited by manual feature extraction and small, homogeneous datasets, which restricted their generalizability and accuracy (3).

A major turning point came with the advent of DL, particularly convolutional neural networks (CNNs), which demonstrated remarkable performance in image classification tasks without the need for hand-crafted features. In 2016, Gulshan et al. published a landmark study in JAMA demonstrating that a DL algorithm could detect referable DR in retinal fundus photographs with performance on par with ophthalmologists, using a dataset of nearly 130,000 images (9). This breakthrough was enabled by greater computational power, large annotated datasets, and CNNs. This study catalysed widespread interest and investment in AI applications within ophthalmology.

Subsequent years saw rapid advancements, including the development of autonomous AI systems capable of clinical decision-making without specialist oversight. The most prominent example is IDx-DR (now marketed as LumineticsCore™), which became the first FDA-approved autonomous AI system in 2018 for detecting more than mild DR in primary care settings (6). This approval was a significant milestone, signalling regulatory endorsement of AI’s potential to operate independently in specific clinical scenarios.

Parallel developments occurred in other subspecialties, including the use of AI to interpret OCT scans for AMD and macular oedema. DeepMind’s collaboration with Moorfields Eye Hospital led to the development of an AI model that could not only match expert performance in triaging retinal OCT scans but also provide explainable diagnostic outputs, thus addressing the ‘black box’ concern of DL systems (10).

More recently, AI applications have extended into glaucoma detection, retinopathy of prematurity (ROP) assessment, cataract grading, and even myopia prediction (11-13). These advances reflect a transition from disease-specific diagnostic tools to multifunctional platforms capable of supporting multiple ophthalmic conditions. The integration of AI into teleophthalmology, especially in low-resource settings, highlights its evolving role from a research tool to a practical solution for workforce shortages and geographical disparities (14). These developments also foreshadow the broader challenge of ensuring equitable access to validated AI systems globally.

Current developments and disease-specific applications

AI in ophthalmology has progressed from narrow, single-disease algorithms to sophisticated, multimodal systems capable of assisting in diagnosis, triage, and clinical decision-making across various eye conditions. Advances in DL, particularly CNNs, have been central to these developments, offering state-of-the-art performance in image-based diagnostics. The most active areas include DR, AMD, glaucoma, ROP, cataract, and refractive error. This section summarizes key developments across these disease areas with focus on AI performance and clinical readiness.

DR

DR remains the most extensively studied condition in ophthalmic AI. Numerous algorithms, trained on large fundus image datasets, have achieved high diagnostic accuracy. The seminal work by Gulshan et al. demonstrated a DL system with an area under the receiver operating characteristic curve (AUC) of 0.99 for detecting referable DR (9). Following this, multiple studies validated DL models across diverse populations and imaging devices, confirming their robustness (15-19).

The U.S. FDA’s approval of IDx-DR in 2018 marked the first instance of an autonomous AI system authorised for use without physician oversight, setting a precedent for regulatory acceptance (6). IDx-DR demonstrated sensitivity of 87.2% and specificity of 90.7% for detecting more-than-mild DR, in a prospective multicentre study involving primary care settings. Since then, several other AI systems such as EyeArt (Eyenuk Inc., Lose Angeles, CA, USA), AEYE-DS (AEYE Health, Inc., New York, NY, USA), and RetCAD (RetinAI, Boston, MA, USA) have also achieved European Conformity (CE) marking or FDA clearance, contributing to broader uptake in screening programs.

A recent systematic review and meta-analysis of 34 real-world studies reported a pooled sensitivity of 88% and specificity of 90% for AI-based DR screening. Performance was consistent across low-, middle-, and high-income countries (16). Besides this, a recent study has reported the development and validation of a DL system to predict time to DR progression within 5 years from fundus images (20). Another study published in 2024 has demonstrated that a system integrating image-based DL and language models can enhance primary diabetes care and DR screening (21). These findings support the readiness of AI systems for population-level screening, particularly in regions with limited ophthalmology workforce.

AMD

AI applications for AMD, particularly those using OCT, have advanced rapidly. DeepMind’s AI system, developed in collaboration with Moorfields Eye Hospital, demonstrated high accuracy in triaging retinal OCT scans and identifying urgent AMD cases (10). The model also provided interpretable visual saliency maps and diagnostic labels, addressing the growing need for explainability in clinical AI. Other studies have shown promising results in classifying drusen, pigment epithelial detachment, and geographic atrophy using AI systems, with AUCs consistently above 0.95 (22). These capabilities are particularly useful in longitudinal monitoring and stratification of patients for treatment with anti-vascular endothelial growth factor (VEGF) therapies.

Glaucoma

AI-based glaucoma detection is inherently more complex due to the need to integrate structural (optic nerve head) and functional (visual field) data. Nonetheless, several models using fundus photographs have achieved promising results. For example, Li et al. developed a DL system with sensitivity and specificity above 90% in detecting glaucomatous optic neuropathy (12). Recent approaches incorporate multimodal data, such as OCT retinal nerve fibre layer (RNFL) thickness, intraocular pressure, and patient demographics, to improve diagnostic precision, progression forecasting and surgical outcome prediction (23,24). However, the absence of a universal reference standard continues to limit clinical adoption compared with DR, and large-scale validation in routine practice is still needed.

ROP

ROP screening is labour-intensive and requires expert interpretation. AI-based tools such as the i-ROP DL model have shown potential in detecting plus disease, a critical marker for treatment-requiring ROP (25). A study by Brown et al. demonstrated that DL could classify plus diseases with performance comparable to expert graders with an AUC of 0.98 (11). Such tools can support non-specialist graders in neonatal units and are particularly relevant for scaling ROP screening in low-resource settings, where the availability of trained ophthalmologists is limited.

Cataract

AI has been explored in anterior segment imaging for cataract detection, surgical planning, and refractive error prediction (26). DL models can grade cataract severity from slit-lamp or fundus images and suggest surgical referral thresholds (27). Furthermore, AI-based biometry has shown promise in improving intraocular lens (IOL) power prediction, with some studies supporting lower mean absolute error compared to traditional formulas (28).

Refractive conditions

In refractive error assessment, ML models utilizing ocular biometric data, such as axial length and corneal curvature, have been developed to estimate cycloplegic refractive errors with high accuracy and can facilitate myopia screening in school-based programs (29,30). Additionally, ML applied to wavefront aberrometer data, combined with demographic information, has proven effective in predicting subjective refraction, further enhancing the precision and scalability of non-cycloplegic refraction screening (13,31). Together, these advancements suggest that AI-driven, non-invasive screening tools can significantly improve early identification of refractive errors in paediatric and adult populations.

Collectively, AI applications across DR, AMD, glaucoma, ROP, cataract, and refractive error demonstrate the breadth of potential for automated diagnostic support in ophthalmology. While DR remains the most mature area with several regulatory-approved systems in active use, emerging evidence for conditions such as AMD and glaucoma suggests that AI could soon augment, or in some cases partially automate, the diagnostic process in other subspecialties. Across conditions, AI performs best when embedded in structured workflows, supported by high-quality imaging, and validated in diverse populations—principles that underpin emerging multi-disease platforms for universal eye health screening and management.

Impact of AI in ophthalmology

The integration of AI into ophthalmic practices has the potential to transform patient care, service delivery, and health-system efficiency. While much of the early discourse has focused on algorithmic performance, the true measure of impact lies in real-world implementation, where AI influences access, diagnostic timelines, clinical decision-making, and patient outcomes.

Enhancing access to eye care

One of AI’s most immediate and measurable impacts is its ability to extend specialist-level screening to underserved or remote populations. Autonomous AI systems for DR, such as IDx-DR, EyeArt, and Eyetelligence Assure+, have enabled screening in primary care endocrinology clinics and community health settings without onsite ophthalmologists (6,18,32). This model reduces the need for patients to travel to tertiary centres, an important consideration in rural and low-resource settings where specialist access is limited (14). Evidence from deployment studies in the United States, India, and Australia shows that integrating AI-based screening into non-specialist clinics significantly increases screening uptake rates (16,18).

Reducing diagnostic delays and workload

AI systems can process and interpret images within minutes, compared to the days or weeks often needed for human grading in large-scale programs (9,33). This accelerated turnaround can expedite referrals for sight-threatening conditions, potentially preventing disease progression. By automating triage of normal or low-risk cases, AI reduces workload on ophthalmologists, allowing them to focus on complex or high-risk patients (34). In large screening programs, this workload redistribution translates into operational cost savings and improved service throughput.

Supporting clinical decision-making

Beyond screening, AI is increasingly being integrated into diagnostic and treatment pathways. For example, AI systems for OCT interpretation can assist in determining the urgency of referral for AMD or macular oedema (10). Similarly, AI-based glaucoma detection tools can flag suspicious optic disc changes for closer evaluation, and AI-driven surgical planning tools are emerging for cataract and refractive surgery (35-38). In these contexts, AI functions as a decision-support partner rather than replacing clinical judgement.

Economic and system-level benefits

From a health economics perspective, AI-enabled screening can be cost-effective when deployed at scale, especially in high-prevalence populations or where human grading costs are high (37-39). Studies modelling DR screening programs have shown that AI-based strategies can be more cost-efficient than traditional human grader models (38,40). Furthermore, integrating AI into teleophthalmology workflows enables centralised review and quality assurance, reducing variability in diagnoses and improving standardisation across large networks (14,41). However, realizing these benefits in routine care depends on overcoming barriers in interoperability, clinician trust and sustainable financing models—challenges discussed in the following section.

Equity and public health implications

At a population level, AI can narrow disparities by decentralizing screening and diagnosis, bridging the urban-rural divide and enabling earlier detection in disadvantaged communities (3,18). However, this impact is contingent on equitable access to the technology itself, including affordable devices, robust internet connectivity, and trained personnel to oversee the workflow. Future efforts should prioritize affordable devices, offline-capable platforms, and workforce training to ensure that AI reduces rather than widens disparities in eye care access.

Challenges in implementing AI in ophthalmology

Despite the rapid progress and encouraging results from clinical trials, the real-world implementation of AI in ophthalmology faces persistent challenges that are not merely technical but structural in nature. These include ethical and accountability concerns, data bias, limited explainability, and, critically, barriers related to health system organization, reimbursement models, and entrenched digital infrastructures. Rather than isolated obstacles, these challenges reflect deeper misalignments between AI innovation and existing clinical, regulatory, and economic ecosystems. Table 2 summarizes these challenges, their underlying causal mechanisms, and solution pathways proposed in recent literature.

Ethical, privacy, and accountability considerations

The deployment of AI in ophthalmology raises important ethical, privacy, and accountability challenges that extend beyond algorithmic performance. AI systems rely on large volumes of images and clinical data often collected retrospectively, creating concerns around data governance, informed consent, secondary data use, and patient privacy (7,14). These issues are particularly salient in low- and middle-income countries (LMICs) where regulatory oversight and data protection frameworks may be less mature (42). In addition, algorithms trained on non-representative datasets risk embedding bias, leading to reduced performance in populations with different disease prevalence, fundus pigmentation, or image quality, and potentially exacerbating existing health inequities (3,43).

Accountability remains an unresolved concern, especially in autonomous or semi-autonomous screening settings. While systems such as IDx-DR and EyeArt have obtained regulatory approval, most AI tools in ophthalmology remain investigational and lack formal clearance across jurisdictions (6). Furthermore, continuous learning algorithms challenge traditional regulatory paradigms, which are designed for static medical devices (42). As AI systems evolve post-deployment, questions arise regarding responsibility for errors, including the respective roles of developers, clinicians, healthcare institutions, and regulators (7,39). These challenges underscore the need for adaptive regulatory pathways incorporating periodic re-validation, post-market surveillance, and clear governance structures to ensure patient safety while enabling innovation.

These challenges are not solely ethical in nature but reflect broader regulatory and governance misalignment, where innovation in AI has outpaced the evolution of accountability frameworks within health systems (42). Building trust in ophthalmic AI will require transparent development, robust validation across diverse populations, and explicit accountability frameworks. Ethical deployment must therefore be accompanied by clear policies on data stewardship, human oversight, and clinical responsibility, ensuring that AI functions as a decision-support tool aligned with professional standards rather than replacement for clinical judgement (3,7). Addressing these ethical and regulatory considerations in parallel with technical development is essential to achieving safe, equitable, and sustainable integration of AI into routine eye care.

Data quality, bias, and generalizability

AI systems are highly dependent on the quality and representativeness of training data. Many models have been developed using datasets from high-income countries, often with limited representation of ethnic minorities, rural populations, or patients imaged on low-cost devices. This can lead to biased outputs and reduced accuracy in diverse populations (3). Studies have shown that AI systems trained on homogeneous datasets may underperform in populations with different disease prevalence, fundus pigmentation, or image quality (43). Ensuring equitable AI requires access to large, diverse, and well-annotated datasets—a challenge in many LMICs. Approaches such as federated learning and global data-sharing initiatives may help address these challenges.

Explainability and trust

DL models are often criticized as ‘black boxes’, with limited interpretability for clinicians and patients. While saliency maps and heatmaps offer some insights, they do not always provide clinically intuitive explanations for AI decisions (10). This lack of transparency hinders clinician trust, raises medico-legal concerns, and complicates patient consent. Developing explainable AI (XAI) frameworks that can communicate decision rationale remains an active research area (44), and will be central to future adoption.

Integration into clinical workflows

Despite strong diagnostic performance, integrating AI into routine ophthalmic workflows remains persistently difficult due to structural and organizational barriers rather than purely technical limitations. A major constraint is the closed and proprietary nature of many electronic medical record (EMR) ecosystems, where high integration costs, restrictive vendor policies, and limited interoperability significantly impede seamless deployment of AI tools. Even when interoperability standards such as Fast Healthcare Interoperability Resources (FHIR) exist, their implementation is inconsistent and often insufficient for real-time clinical integration (45,46).

Economic and incentive misalignment further compounds these barriers. In many health systems, reimbursement models do not adequately recognize autonomous or semi-autonomous AI-based diagnostics, reducing institutional motivation to invest in integration and long-term maintenance. As a result, AI systems are frequently deployed as parallel or add-on tools rather than embedded components of clinical workflows.

Organizational and human-factors inertia also play a critical role. Successful adoption requires workflow redesign, clinician trust, and changes in role delineation, all of which demand time, training, and leadership commitment. In resource-constrained settings, these challenges are amplified by limited digital infrastructure, internet bandwidth and technical support capacity. Together, these structural, economic and human factors explain why workflow integration remains a dominant bottleneck despite rapid algorithmic advances (45,46).

Economic and sustainability challenges

Although AI can be cost-effective at scale (38), initial investments in hardware, software licensing, and workforce training may be prohibitive for smaller clinics and health systems (37). Furthermore, reimbursement models for AI-based screening remain inconsistent, with limited clarity on how payers should compensate autonomous systems compared to human graders (19). Without sustainable financing mechanisms, scaling AI in routine care may remain difficult.

Equity and global health implications

While AI has potential to bridge gaps in eye care access, it may also widen disparities if only high-resource settings can afford cutting-edge tools. Importantly, inequity in ophthalmic AI is not merely a technical issue of reduced algorithmic accuracy but represents a risk of active, iatrogenic harm, whereby biased systems may systematically fail to detect disease in specific populations, such as individuals with different fundus pigmentation or those imaged using lower-cost devices (19). Deployment in LMICs faces challenges related to affordability, infrastructure, and local validation (14). Ensuring equitable access to AI will require deliberate efforts in technology transfer, capacity building, and open-access datasets that include diverse populations. This creates a paradox in which regions with the greatest burden of avoidable vision loss and the highest benefit from AI are often the least equipped with the digital infrastructure, regulatory oversight and financial resources required for safe and effective deployment (19).

Global collaborations, capacity-building programs, and open-access datasets will be critical to ensuring that AI reduces, rather than exacerbates, inequities in eye care. Without rigorous local validation, ethical governance and contextual adaptation, AI deployment risks reinforcing a form of ‘equity stratification’, where proprietary and expensive AI tools preferentially benefit well-resourced private or urban centres while under-resourced public and rural systems are left behind (19,47). Addressing these risks requires treating equity as a core design and policy principle rather than an afterthought, embedding safeguards for local validation, affordability and accountability alongside technical innovation.

Future directions

The trajectory of AI in ophthalmology is moving from single-disease, image-based algorithms to multifunctional, multimodal platforms that integrate seamlessly into clinical practice. Several emerging trends are likely to shape the next decade of development and adoption. Figure 1 provides a conceptual framework summarizing this trajectory, illustrating the pathway from diverse data inputs through AI algorithms to clinical integration and health-systems outcomes. This schematic highlights how AI functions as a bridge between diagnostic technologies and broader goals of improving access, efficiency, and equity in eye care.

Figure 1 Conceptual framework illustrating the role of AI in ophthalmology. The framework illustrates the pathway from diverse data inputs (including fundus photographs, OCT scans, visual fields, demographics, and systemic health data) through AI algorithms for image analysis, multimodal integration, and risk prediction. Outputs are then integrated into clinical workflows such as decision support, autonomous screening, teleophthalmology, and EMR embedding. These processes ultimately influence health-system outcomes, including improved access, faster diagnosis, workforce optimization, cost efficiency, and equity in care delivery. AI, artificial intelligence; EMR, electronic medical record; OCT, optical coherence tomography.

One promising direction is the application of federated learning. Limited access to diverse, annotated datasets has long been a barrier to generalizability, but federated approaches allow models to be trained collaboratively across multiple institutions without sharing raw patient data (43). This preserves data privacy while expanding diversity, helping to overcome geographic and demographic biases that limit current systems (48). Importantly, federated learning also addresses structural data silos and governance constraints that have historically prevented cross-institutional data sharing, which are key root causes of limited generalizability in ophthalmic AI (7).

Future systems are also expected to evolve beyond fundus photographs and OCT by incorporating multimodal data streams, including visual fields, biometric measurements, EMRs, and even genomic and systemic health information (13,29,36,46). Such holistic integration could enable more accurate disease detection and personalized risk prediction. For example, combining retinal imaging with systemic parameters such as blood glucose and blood pressure may substantially improve prognostic models for DR progression (49). By integrating fragmented clinical data streams that currently reside across disconnected systems, multimodal AI directly responds to workflow inefficiencies and diagnostic silos that constrain real-world clinical adoption (46).

Equally important will be advances in explainability. To foster clinician trust and adoption, future platforms will likely embed XAI methods such as saliency maps, counterfactual explanations, and concept activation models (50). These approaches will help clinicians validate AI outputs and use them as ‘augmented intelligence’ tools that flag high-risk cases, reduce workload, and support evidence-based decision-making.

Integration into clinical workflows will also remain a decisive factor for success. Persistent integration challenges are driven not only by technical limitations but also by proprietary EMR ecosystems, high implementation costs, lack of reimbursement incentives and organizational resistance to workflow changes, which future AI platforms must explicitly address (2,46,47). Future AI platforms are expected to leverage interoperability standards like FHIR to connect seamlessly with EMRs and imaging archives. Cloud-based and offline-compatible solutions will provide flexibility across high- and low-resource settings. AI embedded directly into fundus cameras and slit lamps is already being piloted, foreshadowing a new generation of ‘smart devices’ capable of real-time decision support at the point of care (14,18).

At the same time, regulatory frameworks will need to accommodate the unique characteristics of AI, particularly continuous learning systems that adapt post-development. Dynamic pathways involving periodic re-validation and post-market surveillance are already being considered, balancing patient safety with the need to encourage innovation (7). Aligning regulatory pathways with real-world deployment cycles is essential to reducing uncertainty that currently deters institutional investment and clinical uptake of AI systems (42).

Future Research and implementation efforts should also prioritise robust ethical and privacy-preserving governance frameworks. This includes the development of transparent patient authorisation mechanisms for secondary data use, secure data storage architectures and privacy-preserving approaches such as federated learning and on-device inference. Embedding ethical oversight, clear accountability structures and human-in-the-loop safeguards into AI deployment pathways will be essential to maintain public trust and ensure that innovation proceeds in alignment with professional and societal expectations.

Finally, ensuring equity will be central to the future of ophthalmic AI. The greatest promise lies in addressing inequities in access to care, which requires prioritizing affordable and adaptable tools validated in LMICs. Partnerships between academia, industry, and public health organizations will be essential to ensure that AI benefits are distributed globally rather than concentrated in high-resource settings (51).

Strengths and limitations of this review

This narrative review has several strengths. First, it provides a clinically oriented synthesis of AI applications across major ophthalmic disease areas, integrating evidence from real-world validation studies, regulatory experiences, and health systems perspectives. Unlike purely technical reviews, this manuscript explicitly examines implementation challenges, ethical considerations and equity implications, offering a system-level perspective relevant to clinicians, researchers and policymakers. The inclusion of a conceptual framework and an analytical summary of root causes and solution pathways further strengthens its translational relevance.

However, this review also has limitations inherent to its narrative design. It does not follow a systematic search or formal quality appraisal process and therefore may not capture all published studies or provide quantitative comparisons of algorithmic performance. Selection of examples and themes was guided by clinical relevance and policy significance rather than exhaustive coverage. In addition, as the field of ophthalmic AI is rapidly evolving, regulatory approvals, performance evidence, and implementation models may change over time. These limitations underscore the need for complementary systematic reviews and prospective implementation studies to further inform evidence-based adoption of AI in eye care.

Conclusions

AI is rapidly transforming ophthalmology, with multiple systems now matching or, in some contexts, surpassing expert performance in detecting conditions such as DR, AMD, glaucoma and ROP. Real-world deployments, including U.S. FDA-approved autonomous systems and emerging multi-disease platforms, demonstrate that AI has moved beyond proof-of-concept to become a tangible adjunct to clinical practice.

However, widespread adoption remains constrained by persistent challenges, including dataset bias, limited explainability, workflow integration barriers, misaligned reimbursement models, and evolving regulatory frameworks. Without careful oversight, these limitations risk reinforcing existing inequities in access to eye care rather than alleviating them.

Looking ahead, the greatest opportunities lie in developing AI systems that are equitable, transparent, and scalable. Advances in federated learning, multimodal integration, and XAI, supported by adaptive regulatory approaches and global collaboration, will be essential to building trustworthy and sustainable solutions. By synthesizing evidence across disease applications, implementation challenges, and ethical considerations, this review provides a conceptual and practical framework to guide future research, policy development, and responsible clinical integration of AI into ophthalmology.

Ultimately, while AI will not replace ophthalmologists, it is poised to become an indispensable partner in delivering timely, accurate, and equitable eye care, provided its deployment is guided by robust evidence, ethical governance, and a commitment to reducing, rather than deepening, health disparities.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the Narrative Review reporting checklist. Available at https://aes.amegroups.com/article/view/10.21037/aes-25-61/rc

Peer Review File: Available at https://aes.amegroups.com/article/view/10.21037/aes-25-61/prf

Funding: None.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://aes.amegroups.com/article/view/10.21037/aes-25-61/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Ting DSW, Liu Y, Burlina P, et al. AI for medical imaging goes deep. Nat Med 2018;24:539-40. [Crossref] [PubMed]
2. Topol EJ. High-performance medicine: the convergence of human and artificial intelligence. Nat Med 2019;25:44-56. [Crossref] [PubMed]
3. Ting DSW, Pasquale LR, Peng L, et al. Artificial intelligence and deep learning in ophthalmology. Br J Ophthalmol 2019;103:167-75. [Crossref] [PubMed]
4. Wu Y, Qian B, Li T, et al. An eyecare foundation model for clinical assistance: a randomized controlled trial. Nat Med 2025;31:3404-13. [Crossref] [PubMed]
5. Zhou Y, Chia MA, Wagner SK, et al. A foundation model for generalizable disease detection from retinal images. Nature 2023;622:156-63. [Crossref] [PubMed]
6. 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. [Crossref] [PubMed]
7. He J, Baxter SL, Xu J, et al. The practical implementation of artificial intelligence technologies in medicine. Nat Med 2019;25:30-6. [Crossref] [PubMed]
8. Mookiah MR, Acharya UR, Chua CK, et al. Computer-aided diagnosis of diabetic retinopathy: a review. Comput Biol Med 2013;43:2136-55. [Crossref] [PubMed]
9. 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;316:2402-10. [Crossref] [PubMed]
10. De Fauw J, Ledsam JR, Romera-Paredes B, et al. Clinically applicable deep learning for diagnosis and referral in retinal disease. Nat Med 2018;24:1342-50. [Crossref] [PubMed]
11. Brown JM, Campbell JP, Beers A, et al. Automated Diagnosis of Plus Disease in Retinopathy of Prematurity Using Deep Convolutional Neural Networks. JAMA Ophthalmol 2018;136:803-10. [Crossref] [PubMed]
12. Li Z, He Y, Keel S, et al. Efficacy of a Deep Learning System for Detecting Glaucomatous Optic Neuropathy Based on Color Fundus Photographs. Ophthalmology 2018;125:1199-206. [Crossref] [PubMed]
13. Zuo H, Huang B, He J, et al. Machine Learning Approaches in High Myopia: Systematic Review and Meta-Analysis. J Med Internet Res 2025;27:e57644. [Crossref] [PubMed]
14. Li JO, Liu H, Ting DSJ, et al. Digital technology, tele-medicine and artificial intelligence in ophthalmology: A global perspective. Prog Retin Eye Res 2021;82:100900. [Crossref] [PubMed]
15. 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;318:2211-23. [Crossref] [PubMed]
16. Joseph S, Selvaraj J, Mani I, et al. Diagnostic Accuracy of Artificial Intelligence-Based Automated Diabetic Retinopathy Screening in Real-World Settings: A Systematic Review and Meta-Analysis. Am J Ophthalmol 2024;263:214-30. [Crossref] [PubMed]
17. Keel S, Li Z, Scheetz J, et al. Development and validation of a deep-learning algorithm for the detection of neovascular age-related macular degeneration from colour fundus photographs. Clin Exp Ophthalmol 2019;47:1009-18. [Crossref] [PubMed]
18. Joseph S, Wang Y, Drinkwater JJ, et al. Effectiveness of artificial intelligence-based diabetic retinopathy screening in primary care and endocrinology settings in Australia: a pragmatic trial. Br J Ophthalmol 2025;110:76-82. [Crossref] [PubMed]
19. Rajesh AE, Davidson OQ, Lee CS, et al. Artificial Intelligence and Diabetic Retinopathy: AI Framework, Prospective Studies, Head-to-head Validation, and Cost-effectiveness. Diabetes Care 2023;46:1728-39. [Crossref] [PubMed]
20. Dai L, Sheng B, Chen T, et al. A deep learning system for predicting time to progression of diabetic retinopathy. Nat Med 2024;30:584-94. [Crossref] [PubMed]
21. Li J, Guan Z, Wang J, et al. Integrated image-based deep learning and language models for primary diabetes care. Nat Med 2024;30:2886-96. [Crossref] [PubMed]
22. Treder M, Lauermann JL, Eter N. Automated detection of exudative age-related macular degeneration in spectral domain optical coherence tomography using deep learning. Graefes Arch Clin Exp Ophthalmol 2018;256:259-65. [Crossref] [PubMed]
23. Nishida T, Moghimi S, Yamane MLM, et al. Vision-Related Quality of Life Among Healthy, Preperimetric Glaucoma, and Perimetric Glaucoma Patients. Am J Ophthalmol 2023;248:127-36. [Crossref] [PubMed]
24. Lan CH, Chiu TH, Yen WT, et al. Artificial Intelligence in Glaucoma: Advances in Diagnosis, Progression Forecasting, and Surgical Outcome Prediction. Int J Mol Sci 2025;26:4473. [Crossref] [PubMed]
25. Coyner AS, Murickan T, Oh MA, et al. Multinational External Validation of Autonomous Retinopathy of Prematurity Screening. JAMA Ophthalmol 2024;142:327-35. [Crossref] [PubMed]
26. Ahuja AS, Paredes Iii AA, Eisel MLS, et al. Applications of Artificial Intelligence in Cataract Surgery: A Review. Clin Ophthalmol 2024;18:2969-75. [Crossref] [PubMed]
27. Lu S, Ba L, Wang J, et al. Deep learning-driven approach for cataract management: towards precise identification and predictive analytics. Front Cell Dev Biol 2025;13:1611216. [Crossref] [PubMed]
28. Liu Z, Zou Y, Yu Y, et al. Accuracy of Intraocular Lens Power Calculation in Pediatric Secondary Implantation: In-the-Bag Versus Sulcus Placement. Am J Ophthalmol 2023;249:137-43. [Crossref] [PubMed]
29. Ying B, Chandra RS, Wang J, et al. Machine Learning Models for Predicting Cycloplegic Refractive Error and Myopia Status Based on Non-Cycloplegic Data in Chinese Students. Transl Vis Sci Technol 2024;13:16. [Crossref] [PubMed]
30. Qi Z, Li T, Chen J, et al. A deep learning system for myopia onset prediction and intervention effectiveness evaluation in children. NPJ Digit Med 2024;7:206. [Crossref] [PubMed]
31. Hernández CS, Gil A, Casares I, et al. Prediction of manifest refraction using machine learning ensemble models on wavefront aberrometry data. J Optom 2022;15:S22-31. [Crossref] [PubMed]
32. Fleming AD, Goatman KA, Philip S, et al. Automated grading for diabetic retinopathy: a large-scale audit using arbitration by clinical experts. Br J Ophthalmol 2010;94:1606-10. [Crossref] [PubMed]
33. Joseph S, Rajan RP, Sundar B, et al. Validation of diagnostic accuracy of retinal image grading by trained non-ophthalmologist grader for detecting diabetic retinopathy and diabetic macular edema. Eye (Lond) 2023;37:1577-82. [Crossref] [PubMed]
34. Sim DA, Keane PA, Tufail A, et al. Automated retinal image analysis for diabetic retinopathy in telemedicine. Curr Diab Rep 2015;15:14. [Crossref] [PubMed]
35. 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;126:513-21. [Crossref] [PubMed]
36. Tognetto D, Giglio R, Vinciguerra AL, et al. Artificial intelligence applications and cataract management: A systematic review. Surv Ophthalmol 2022;67:817-29. [Crossref] [PubMed]
37. Bellemo V, Lim ZW, Lim G, et al. Artificial intelligence using deep learning to screen for referable and vision-threatening diabetic retinopathy in Africa: a clinical validation study. Lancet Digit Health 2019;1:e35-44. [Crossref] [PubMed]
38. Hu W, Joseph S, Li R, et al. Population impact and cost-effectiveness of artificial intelligence-based diabetic retinopathy screening in people living with diabetes in Australia: a cost effectiveness analysis. EClinicalMedicine 2024;67:102387. [Crossref] [PubMed]
39. Wang Y, Hu W, Zhou K, et al. What is the role of human decisions in a world of artificial intelligence: an economic evaluation of human-ai collaboration in diabetic retinopathy screening. arXiv:2503.20160 [Preprint]. 2025. Available online: https://arxiv.org/abs/2503.20160
40. Wang Y, Liu C, Hu W, et al. Economic evaluation for medical artificial intelligence: accuracy vs. cost-effectiveness in a diabetic retinopathy screening case. NPJ Digit Med 2024;7:43.
41. Li Z, Keel S, Liu C, et al. Can Artificial Intelligence Make Screening Faster, More Accurate, and More Accessible? Asia Pac J Ophthalmol (Phila) 2018;7:436-41. [Crossref] [PubMed]
42. Wang Y, Liu C, Zhou K, et al. Towards regulatory generative AI in ophthalmology healthcare: a security and privacy perspective. Br J Ophthalmol 2024;108:1349-53. [Crossref] [PubMed]
43. Burlina P, Joshi N, Paul W, et al. Addressing Artificial Intelligence Bias in Retinal Diagnostics. Transl Vis Sci Technol 2021;10:13. [Crossref] [PubMed]
44. Tan TF, Dai P, Zhang X, et al. Explainable artificial intelligence in ophthalmology. Curr Opin Ophthalmol 2023;34:422-30. [Crossref] [PubMed]
45. Aravazhi PS, Gunasekaran P, Benjamin NZY, et al. The integration of artificial intelligence into clinical medicine: Trends, challenges, and future directions. Dis Mon 2025;71:101882. [Crossref] [PubMed]
46. Ye J, Woods D, Jordan N, et al. The role of artificial intelligence for the application of integrating electronic health records and patient-generated data in clinical decision support. AMIA Jt Summits Transl Sci Proc 2024;2024:459-67.
47. Tan TF, Thirunavukarasu AJ, Jin L, et al. Artificial intelligence and digital health in global eye health: opportunities and challenges. Lancet Glob Health 2023;11:e1432-43. [Crossref] [PubMed]
48. Kaissis GA, Makowski MR, Rückert D, et al. Secure, privacy-preserving and federated machine learning in medical imaging. Nature Machine Intelligence 2020;2:305-11.
49. Li Z, Wang L, Wu X, et al. Artificial intelligence in ophthalmology: The path to the real-world clinic. Cell Rep Med 2023;4:101095. [Crossref] [PubMed]
50. 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:e1312. [Crossref] [PubMed]
51. Burton MJ, Ramke J, Marques AP, et al. The Lancet Global Health Commission on Global Eye Health: vision beyond 2020. Lancet Glob Health 2021;9:e489-551. [Crossref] [PubMed]