Artificial intelligence (AI) is reshaping ophthalmology, moving from research laboratories into everyday clinical care [1]. With its strong reliance on imaging and pattern recognition, ophthalmology is uniquely positioned to benefit from AI innovations.
These developments range from autonomous disease screening to predictive modelling and language-based tools that improve patient communication [2]. In this article, we explore the most significant, recent advances showing how AI is transforming clinical care, expanding access to services and raising important questions about ethics, equity and implementation.
Autonomous diabetic retinopathy screening: from lab to clinic
One of the most well-established applications of AI in ophthalmology is autonomous screening for diabetic retinopathy (DR) [3]. Diabetic retinopathy remains a leading cause of preventable blindness worldwide, yet screening coverage often falls short due to workforce shortages and limited access [4]. Artificial intelligence is helping bridge this gap by automating image interpretation and triaging patients for further review.
However, despite robust evidence and regulatory approvals elsewhere, AI-driven DR screening is not currently implemented within the English Diabetic Eye Screening Programme (DESP) [5]. In the UK, AI tools have largely been evaluated as triage or decision-support systems, but only in retrospective studies rather than real-world screening settings [5].
In 2018, the IDx-DR system became the first autonomous AI diagnostic tool in any medical field to receive FDA approval [6]. Unlike earlier software designed to assist clinicians, IDx-DR makes an independent decision by analysing retinal photographs to determine whether a patient has more-than-mild DR, resulting in onward referral to human clinicians. In its pivotal trial, IDx-DR achieved 87% sensitivity and 90% specificity, comparable to performance by retina specialists [6].
Other systems such as EyeArt have shown similar accuracy and are already used in primary care and pharmacy settings to identify patients needing referral without ophthalmologist oversight [7]. Real-world programmes in India and China have successfully integrated AI into community screening, maintaining sensitivities and specificities above 85–90% while improving access in rural regions [7].
Recent innovations emphasise portability and point-of-care deployment. The AEYE-DS device combines a handheld, non-mydriatic fundus camera with built-in AI software that produces instant screening results in under one minute [8]. This enables DR screening to occur in community clinics or during diabetes reviews, reducing the need for staff with specialist training.
Challenges remain around reimbursement, integration into established care pathways, and medico-legal responsibility. Nonetheless, autonomous screening represents one of the first validated examples of AI delivering safe, effective and scalable clinical benefit in ophthalmology.
From deep learning to foundation models
Most current AI systems in ophthalmology use deep learning, where networks are trained with labelled images to perform single tasks such as identifying diabetic retinopathy or glaucoma. While powerful, these models are limited by the need for large, labelled datasets and often perform less well on images from new devices or populations.
A newer approach, known as foundation models, seeks to overcome these limitations. These large models are pretrained on millions of mostly unlabelled images, learning the general structure and patterns of the retina before being fine-tuned for specific clinical tasks. The most prominent example is RETFound, developed in 2023 by Moorfields Eye Hospital and University College London [9]. RETFound was trained on approximately 1.6 million retinal images using self-supervised learning, a method in which the model predicts missing parts of images to learn from unlabelled data. Once pretrained, it was fine-tuned for a range of applications including DR and glaucoma detection, as well as predicting systemic disease risk from retinal images. Studies have shown that RETFound can outperform conventional supervised models while requiring fewer labelled images [9].
“Autonomous screening represents one of the first validated examples of AI delivering safe, effective, and scalable clinical benefit in ophthalmology”
In simple terms, RETFound acts as a foundation model for ophthalmic AI, analogous to how ChatGPT provides a general-purpose base model for language applications. It can be adapted to a wide range of ophthalmic and systemic diseases, offering improved flexibility and efficiency compared with task-specific models. Other emerging ophthalmic foundation models have been developed as well, for example, multimodal visual-language models such as EyeCLIP, which learns shared representations across multiple imaging modalities, and approaches that enhance diagnostic models for conditions like myopic maculopathy using RETFound-based architectures [10]. However, challenges remain, including data bias, substantial computational requirements, and the need for robust validation across diverse populations. Despite these limitations, foundation models represent a significant step toward AI systems capable of analysing multiple diseases and imaging modalities within a single, unified framework.
Large language models: new voices in patient communication
While image-based AI has led most ophthalmic advances, language models are quietly transforming how clinicians communicate with patients.
Large language models (LLMs) such as ChatGPT can understand and generate human text. In ophthalmology, researchers have begun adapting them to simplify complex information and enhance patient engagement. A 2024 study showed that ChatGPT could rewrite peer-reviewed ophthalmology articles into plain-language summaries at about a seventh-grade reading level without significant loss of accuracy [11]. Another group developed EyeGPT, a chatbot capable of answering common eye health questions in accessible language [12].
In clinical practice, LLMs may soon generate patient information leaflets, summarise consultation notes or draft referral letters. Although early adoption occurred primarily in the US, these systems are now proliferating rapidly across the UK and other healthcare systems [13]. Several NHS trusts are actively piloting or scaling AI scribe technologies to address documentation burden, improve clinic efficiency and mitigate clinician burnout [14]. In the absence of autonomous AI screening within UK ophthalmology services, this represents the most tangible and impactful use of AI currently embedded in everyday practice [5].
However, language models must remain tools under clinician supervision. They can produce inaccurate or fabricated statements, known as hallucinations, which makes oversight essential [13]. While they currently cannot interpret images, research into multimodal AI aims to integrate visual and text understanding in future systems. With careful use, these tools can improve efficiency and enhance communication, particularly for patients with varying literacy levels.
Looking ahead: making AI equitable, explainable and useful
As AI moves from research into routine clinical practice, its success will depend not only on technical accuracy but also on fairness, transparency and integration. Models trained on limited or unrepresentative datasets may underperform in populations underrepresented in the training cohort, risking inequities in care. Indeed, bias in medical imaging AI has been well documented, and strategies to mitigate it are an active area of research [15].
Transparency is equally important. Clinicians must be able to understand why an AI system reaches a particular decision and how confident it is in that output. Methods such as attention maps and saliency visualisations may enhance interpretability, helping to build trust and support informed clinical use [15].
Successful implementation also relies on practical integration into everyday workflows. AI tools that complicate processes or disrupt patient flow are unlikely to achieve sustained use. Systems must integrate efficiently with electronic health records, demonstrate cost-effectiveness, and operate within clear regulatory frameworks to ensure responsible and sustainable adoption.
Conclusion
Artificial intelligence is no longer an abstract concept in ophthalmology. From autonomous DR screening to LLMs that improve communication, AI is already transforming how eyecare is delivered.
If used responsibly, AI may help clinicians detect disease earlier, personalise treatment, and even gain insights into systemic health from retinal images. The key to success will be balance: embracing innovation while maintaining human oversight, ethical standards and patient trust.
TAKE HOME MESSAGES
-
Artificial intelligence is moving from research into everyday ophthalmic care.
-
Autonomous diabetic retinopathy screening systems like IDx-DR and EyeArt have proven clinical accuracy.
-
Foundation models such as RETFound enable generalisable disease detection.
-
Large language models can enhance patient understanding and streamline documentation.
-
Successful artificial intelligence implementation depends on fairness, transparency and integration into workflows.
References
1. Topol EJ. High-performance medicine: the convergence of human and artificial intelligence. Nat Med 2019;25(1):44–56.
2. Ting DSW, Pasquale LR, Peng L, et al. Artificial intelligence and deep learning in ophthalmology. Br J Ophthalmol 2019;103(2):167–175.
3. 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.
4. Yau JWY, Rogers SL, Kawasaki R, et al. Global prevalence and major risk factors of diabetic retinopathy. Diabetes Care 2012;35(3):556–64.
5. Wahlich C, Chandrasekaran L, Chaudhry UAR, et al. Patient and practitioner perceptions around use of artificial intelligence within the English NHS diabetic eye screening programme. Diabetes Res Clin Pract 2025;219:111964.
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.
7. https://www.eyenuk.com/en/articles/
diabetic-retinopathy/uk-nsc-report
8. Iapoce C. FDA authorizes autonomous AI for portable diabetic retinopathy screening. HCPLive [Internet] 2024:
https://www.hcplive.com/view/fda-authorizes
-autonomous-ai-portable-diabetic-retinopathy-screening
9. Zhou Y, Chia MA, Wagner SK, et al. UK Biobank Eye & Vision Consortium. A foundation model for generalizable disease detection from retinal images. Nature 2023;622(7981):156–63.
10. Shi D, Zhang W, Yang J, et al. A multimodal visual language foundation model for computational ophthalmology. NPJ Digit Med 2025;8(1):381.
11. Riazi Esfahani P, Ward J, Yong A, et al. Evaluating ChatGPT’s accuracy and readability in responding to common ophthalmology questions. Cureus 2025;17(7):e87920.
12. Seyyed Kalantari L, Zhang H, McDermott MBA, et al. Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under served patient populations. Nat Med 2021;27(12):2176–82.
13. Haniff Q, Meng Z, Pongkemmanun T, et al. Use of artificial intelligence to transcribe and summarise general practice consultations. J Med Artif Intell 2025;8:9753.
14. https://www.gosh.nhs.uk/news/
researchgosh-led-trial-of-ai-scribe-technology
-shows-transformative-benefits-for-patients
-and-clinicians-across-london
15. Seyyed-Kalantari L, Zhang H, McDermott M, et al. Underdiagnosis bias of artificial intelligence algorithms applied to medical imaging. Nat Med 2021;27(12):2176–82.
[All links last accessed February 2026]
Declaration of competing interests: None declared.
