AI in Ophthalmology
What is Artificial Intelligence?
The term artificial intelligence (AI) has been around since the 1950s, popularized by a professor at Dartmouth College, John McCarthy1. As defined by Britannica, AI is "the ability of a digital computer or computer-controlled robot to perform tasks commonly associated with intelligent beings."2 In today's ever-changing world, AI has become a major buzzword in research and consumer technology in recent years. Large language models (LLM) such as ChatGPT and image generation models such as Stable Diffusion XL have forever changed how we interact with and create data.
Why the sudden popularity in AI?
The term "AI" seems to be everywhere lately, from new research papers to smartphone apps and consumer devices. A key driver behind this explosion is the accessibility of deep learning (DL), a powerful subset of AI that only became widely usable in the last decade. IBM defines DL as "a subset of machine learning [which is a subset of AI] that uses multilayered neural networks, called deep neural networks, to simulate the complex decision-making power of the human brain."3
The breakthrough came with the advancement of graphical processing units (GPUs), primarily driven by the gaming industry. Unlike CPUs, which are optimized for general-purpose tasks with a small number of complex cores, GPUs contain thousands of smaller, simpler cores designed specifically for handling large-scale linear algebra operations, such as matrix multiplication, with high efficiency.4 These operations, which are used to render 3D graphics in video games, are fundamentally the same mathematical operations at the heart of deep learning, where every layer of a neural network transforms input data into higher-level representations through a series of mathematical operations on large matrices. As a result, GPUs can train neural networks much faster and more efficiently than CPUs.
Companies like NVIDIA recognized this potential early and began developing software frameworks (such as CUDA) that allow developers to harness GPU power for deep learning tasks. Combined with the availability of large datasets, open-source libraries like TensorFlow and PyTorch, and more affordable hardware, this convergence made it possible for researchers, startups, and even the general public to train powerful AI models without needing a supercomputer. In short, GPU development was a catalyst in bringing AI into the mainstream.
How does deep learning work?
Deep learning is a specialized subset of machine learning that attempts to replicate the way the human brain processes information by using artificial neural networks. While the mathematical foundations of deep learning have existed for years, its practical implementation was historically limited by computational constraints, particularly during the training phase of models. Training a deep neural network involves massive numbers of matrix multiplications and tensor operations to adjust millions (or even billions) of parameters based on input data, a process that was not reasonably feasible on early computers. The DL networks consist of multiple layers of interconnected nodes, each capable of transforming input data through a series of weighted connections.5 As data moves through the layers, the model extracts increasingly abstract and complex features. For instance, in the context of ophthalmology, a DL system analyzing a fundus photograph may initially recognize basic visual elements like edges or color gradients in the early layers. As the data progresses deeper into the network, the system begins to identify more sophisticated structures such as blood vessels, the optic disc, or the macula. The model learns to recognize patterns and make predictions by being trained on large datasets with labeled examples, adjusting its internal weights through a process known as backpropagation.6 This iterative learning mechanism allows the network to improve over time by minimizing error. Recent advances in graphics processing units (GPUs) and the availability of extensive annotated datasets have dramatically improved the efficiency and accuracy of DL training, making sophisticated AI applications in medicine increasingly viable.
AI in Ophthalmology
Since as early as the 1970s, AI has appeared in ophthalmology research to enhance medical diagnostic decision-making.7 This field has emerged as one of the most promising domains for AI integration, largely due to the data-rich nature of the specialty. Imaging technologies like fundus photography, OCT, and fluorescein angiography provide highly structured visual data that is well-suited for machine learning.
Diabetic retinopathy (DR) is a prominent cause of preventable blindness globally. Regular screening of diabetic individuals for DR is critical for early detection and intervention. AI-powered systems have achieved substantial success in automating DR screening and grading from fundus images.8 These systems can detect specific lesions characteristic of DR, such as microaneurysms, hemorrhages, exudates, and neovascularization. Companies such as IDx Technologies and AEYE Health Systems were among the first to receive FDA clearance for AI models used in diabetic retinopathy screening.9 These systems can analyze retinal images without requiring input from an ophthalmologist, enabling early detection of disease in primary care settings where specialist access may be limited.
Beyond DR, AI models exhibit considerable promise for the detection of other prevalent ocular diseases. For age-related macular degeneration (AMD), AI algorithms can analyze fundus images and OCT scans to identify key features like drusen, geographic atrophy, and neovascularization.10 One study reported an area under the curve (AUC) of 0.936 for AMD detection, with an accuracy of 86.3%.¹⁰ Similarly, for glaucoma, AI systems evaluate optic disc morphology, neuroretinal rim loss, and retinal nerve fiber layer (RNFL) thickness from fundus photographs and OCT scans. AI algorithms have achieved an AUC of 0.863 for referable glaucomatous optic neuropathy, with 80.2% accuracy.10 Automated segmentation of retinal vessels from digital fundus images also demonstrates high sensitivity (0.81) and specificity (0.97) for disease diagnosis and screening. These advancements indicate AI's potential to enhance screening programs for AMD and glaucoma, supporting early identification and timely referral to specialists, which is crucial for disease management and vision preservation.
AI's utility extends to quantitative analysis, providing objective measurements that support diagnosis and monitor disease progression in both anterior and posterior segment disorders. In the anterior segment, AI can analyze images from slit-lamp examinations, OCT, and corneal topography to quantify features such as corneal curvature, anterior chamber depth, and cataract grading.11 This enables precise assessment of conditions like keratoconus and glaucoma. Automated quantitative evaluation can standardize measurements, reducing variability inherent in manual assessments. For the posterior segment, AI excels at segmenting and quantifying specific retinal structures and lesions from various imaging modalities. This includes detailed measurement of retinal layer thicknesses from OCT scans, which is valuable for monitoring progression in conditions like macular edema or geographic atrophy.12
AI has demonstrated significant promise in analyzing retinal fundus photographs not only for the detection of ocular diseases but also for the prediction of various systemic parameters. Beyond identifying retinal pathology, AI algorithms can potentially extract information indicative of a patient's demographic characteristics such as age and gender, and even detect signs associated with systemic diseases, including cardiovascular, hematological, neurological, and metabolic disorders. This capability stems from the unique anatomical and physiological properties of the eye, particularly the retina, which shares microvascular similarities with other vital organs like the brain, kidneys, and heart. Consequently, the retina serves as a valuable, non-invasive window into systemic health.
Despite this potential, the use of fundus imaging for systemic disease prediction remains relatively underexplored, largely due to the complexity of model development and the novelty of machine learning in medical applications. However, recent literature reveals growing interest and expanding research in this field. For example, multiple studies have attempted to predict demographic variables such as age and gender from retinal images.13 These studies showed variable accuracy and generalizability, with predictive success often dependent on the inclusion of key retinal structures such as the optic disc and macula. Notably, one study found that gender could only be reliably predicted when both the fovea and optic disc were visible.14
In terms of cardiovascular assessment, AI models have shown encouraging results in estimating parameters such as blood pressure, retinal vessel calibers, and even coronary artery calcium (CAC) scores. These models offer a potential non-invasive alternative to traditional techniques like CT scans, with some showing comparable predictive power for cardiovascular risk stratification. For instance, the RetiCAC model derived from fundus photographs was proposed as a viable substitute for cardiac CT in low-resource settings due to its strong performance in predicting major adverse cardiovascular events.15
Hematological conditions such as anemia have also been targeted using fundus images. Models have achieved moderate success in predicting hemoglobin levels and hematocrit, with the optic nerve head often cited as a critical region due to its color and pallor. Interestingly, one study demonstrated that even when high-resolution detail was removed or images were blurred, the model retained predictive accuracy, suggesting that overall retinal coloration contributes significantly to the detection of anemia.16
Neurological diseases represent another emerging application. A particularly notable study used retinal fundus images to predict Alzheimer's disease, achieving an accuracy of 82%.17 This underscores the potential of retinal imaging to capture early neurovascular changes associated with neurodegeneration. A saliency map analysis further revealed that small vessel morphology, rather than large vessels, played a pivotal role in the model's decision-making process, aligning with current understanding of microvascular dysfunction in Alzheimer's pathology.
Future directions
These findings collectively illustrate a paradigm shift in how ophthalmic imaging, traditionally focused on ocular health, is increasingly being recognized as a rich source of systemic health data. The integration of AI holds promise for developing non-invasive, cost-effective, and accessible screening tools across a diverse range of medical specialties.
The future of AI in ophthalmology lies not in replacing clinicians, but in augmenting and streamlining diagnostic workflows. By acting as a preliminary screening tool, AI can triage patients more effectively, reduce diagnostic delays, and free up specialist time for more complex cases. Moreover, AI can help democratize eye care, particularly in low-resource settings where access to trained ophthalmologists is limited. Cloud-based diagnostic tools can be deployed via portable fundus cameras or smartphone-based platforms, bringing quality screening to underserved populations. As models become more interpretable and trustworthy, integration with EHRs and real-time decision support systems will become more common. The incorporation of multiple data sources - combining clinical notes, lab values, and imaging - will allow AI to offer more holistic insights into ocular and systemic health.
However, widespread adoption will require continued regulatory oversight, validation across diverse populations, and strong collaboration between computer engineers, clinicians, and patients to ensure transparency, equity, and ethical deployment. Challenges remain, particularly regarding data standardization, external validation of models, and the translation of these innovations into routine clinical practice. Addressing these challenges will be essential for realizing AI's full potential in transforming ophthalmic care.
References
1. The birth of Artificial Intelligence (AI) research | Science and Technology. Accessed July 16, 2025. https://st.llnl.gov/news/look-back/birth-artificial-intelligence-ai-research
2. Artificial intelligence (AI) | Definition, Examples, Types, Applications, Companies, & Facts | Britannica. July 15, 2025. Accessed July 16, 2025. https://www.britannica.com/technology/artificial-intelligence
3. What Is Deep Learning? | IBM. June 17, 2024. Accessed July 16, 2025. https://www.ibm.com/think/topics/deep-learning
4. Sevilla J, Heim L, Ho A, Besiroglu T, Hobbhahn M, Villalobos P. Compute Trends Across Three Eras of Machine Learning. In: 2022 International Joint Conference on Neural Networks (IJCNN). IEEE; 2022:1-8. doi:10.1109/ijcnn55064.2022.9891914
5. Kriegeskorte N, Golan T. Neural network models and deep learning. Curr Biol. 2019;29(7):R231-R236. doi:10.1016/j.cub.2019.02.034
6. Sarker IH. Deep Learning: A Comprehensive Overview on Techniques, Taxonomy, Applications and Research Directions. SN Comput Sci. 2021;2(6). doi:10.1007/s42979-021-00815-1
7. Weiss SM, Kulikowski CA, Safir A. A model-based consultation system for the long-term management of glaucoma. In: Proceedings of the 5th International Joint Conference on Artificial Intelligence - Volume 2. IJCAI’77. Morgan Kaufmann Publishers Inc.; 1977:826-832. doi:10.5555/1622943.1623010
8. Ting DSW, Pasquale LR, Peng L, et al. Artificial intelligence and deep learning in ophthalmology. Br J Ophthalmol. 2019;103(2):167-175. doi:10.1136/bjophthalmol-2018-313173
9. Jeon S, Liu Y, Li JPO, Webster D, Peng L, Ting D. AI papers in ophthalmology made simple. Eye. 2020;34(11):1947-1949. doi:10.1038/s41433-020-0929-6
10. Zapata MA, Royo-Fibla D, Font O, et al. <p>Artificial Intelligence to Identify Retinal Fundus Images, Quality Validation, Laterality Evaluation, Macular Degeneration, and Suspected Glaucoma</p>. Clin Ophthalmol. 2020;Volume 14:419-429. doi:10.2147/opth.s235751
11. Ting DSJ, Foo VH, Yang LWY, et al. Artificial intelligence for anterior segment diseases: Emerging applications in ophthalmology. Br J Ophthalmol. 2021;105(2):158-168. doi:10.1136/bjophthalmol-2019-315651
12. Lee CS, Tyring AJ, Deruyter NP, Wu Y, Rokem A, Lee AY. Deep-learning based, automated segmentation of macular edema in optical coherence tomography. Biomed Opt Express. 2017;8(7):3440. doi:10.1364/boe.8.003440
13. Betzler BK, Rim TH, Sabanayagam C, Cheng CY. Artificial Intelligence in Predicting Systemic Parameters and Diseases From Ophthalmic Imaging. Front Digit Health. 2022;4. doi:10.3389/fdgth.2022.889445
14. Munk MR, Kurmann T, Márquez-Neila P, Zinkernagel MS, Wolf S, Sznitman R. Assessment of patient specific information in the wild on fundus photography and optical coherence tomography. Sci Rep. 2021;11(1):8621. doi:10.1038/s41598-021-86577-5
15. Rim TH, Lee CJ, Tham YC, et al. Deep-learning-based cardiovascular risk stratification using coronary artery calcium scores predicted from retinal photographs. Lancet Digit Health. 2021;3(5):e306-e316. doi:10.1016/S2589-7500(21)00043-1
16. Mitani A, Huang A, Venugopalan S, et al. Detection of anaemia from retinal fundus images via deeplearning. Nat Biomed Eng. 2019;4(1):18-27. doi:10.1038/s41551-019-0487-z
17. Tian J, Smith G, Guo H, et al. Modular machine learning for Alzheimer’s disease classification from retinal vasculature. Sci Rep. 2021;11(1):238. doi:10.1038/s41598-020-80312-2