FLORetina 2025: AI in practice, discussed by Marion Munk and Quan Dong Nguyen

Artificial intelligence (AI) has moved rapidly from experimental use to real-world deployment in retinal care, now forming an integrated ecosystem across screening, diagnosis, monitoring, research, and surgery.

During the 2025 FLORetina meeting, Marion R Munk, MD, PhD, FEBO and Quan Dong Nguyen, MD, MSc, FAAO, FARVO, FASRS, founding members of the Society for Artificial Intelligence in Vision and Ophthalmology (SAIVO), presented on AI in vision and ophthalmology.

The Eye Care Network asked Prof Munk and Prof Nguyen about AI's presence in retinal care and what that looks like.

How is AI currently being applied to retinal disease diagnosis and management?

AI has already moved well beyond the proof-of-concept stage in retina. Today, deep learning systems are being used to automatically detect diabetic retinopathy (DR) and other retinal diseases from fundus photographs and OCT scans with performance approaching that of expert graders. Some of these tools, such as autonomous systems for diabetic retinopathy screening, are now FDA-cleared and CE-marked and are used in primary care settings to identify patients who need referral to an eye specialist or a retina specialist.¹

In daily clinical practice, AI is increasingly embedded in imaging platforms to segment fluid, quantify atrophy, and track biomarkers longitudinally in neovascular age-related macular degeneration (nAMD) and geographic atrophy (GA). Such application supports more personalised retreatment strategies and may optimise visit frequency and injection intervals.²

The Society for Artificial Intelligence in Vision and Ophthalmology (SAIVO) was established in May 2025.³ At our first SAIVO course at FLORetina 2025, we were showcasing how broad this spectrum has become. We discussed AI-based home monitoring for nAMD (Prof Anat Loewenstein), AI-driven OCT evaluations for macular neovascularization and atrophy (Prof Daniel Pauleikhoff), and AI applications in clinical trials for retinal diseases (Dr Daniela Ferrara), as well as AI in GA from pegcetacoplan trials (Prof Diana Do), among other topics. These examples illustrate that AI in ophthalmology is no longer a single “algorithm,” but a full ecosystem spanning screening, diagnosis, monitoring, and research.

What are the most promising machine learning models for ophthalmic imaging analysis?

The most promising approaches in ophthalmic imaging are deep learning architectures—particularly Convolutional Neural Networks (CNNs) and increasingly transformer-based and multimodal models—that can integrate different image types and even clinical metadata. Recent work has shown that multimodal models combining OCT and fundus photography can improve multilabel disease prediction and better reflect the complexity of real-world patients.⁴

We also see a growing interest in foundation and self-supervised models that can be pre-trained on very large, weakly labelled datasets and then fine-tuned for specific tasks. Techniques such as fundus-enhanced disease-aware distillation are being explored to transfer information between modalities and improve OCT-based disease classification without requiring perfectly paired data, which is more realistic in clinical workflows.⁵

Our SAIVO Symposium reflects this evolution: from LucIA, a RAG-based virtual ophthalmology assistant (Prof Eric Souied), to AI that assists with multimodal image acquisition in the CORD-IV study (Dr Maria Cristina Savastano) and AI-based variant interpretation in genetic retinal disease (Dr Jia-Horung Hung). Together these talks highlight how models are moving from single-task classifiers to intelligent systems that support image acquisition, interpretation, and decision-making.

How do you address bias and variability in AI training datasets for eye care?

Bias and variability are central concerns in ophthalmic AI because our algorithms are only as fair and robust as the data we train them on. We know that retinal disease prevalence, phenotype, and imaging characteristics differ across populations and devices. If our training dataset is narrow, performance can drop significantly when the model is deployed elsewhere. Recent work comparing performance across centres and ethnicities has illustrated how limited generalisability can be if diversity is not built in from the start.

To address this, we emphasize several principles in SAIVO:

• Diverse, multicentre datasets that include different ethnicities, age groups, disease stages, and imaging platforms.
• Transparent reporting of dataset composition and external validation on truly independent cohorts, not just random splits of the same data.
• Continuous post-deployment monitoring to detect drift when disease patterns, cameras, or care pathways change.

In our Symposium at FLORETINA 2025, we also highlight how population differences can impact classifiers—for instance, in retinopathy of prematurity (ROP) screening, where Prof Darius Moshfeghi discussed how population characteristics influence algorithm performance—and how biomarker definitions (e.g., photoreceptor integrity in AMD, as presented by Prof Sobha Sivaprasad) must be standardised to avoid embedding subjective biases into ground-truth labels.

What regulatory challenges exist for clinical deployment of AI tools in ophthalmology?

From a regulatory perspective, ophthalmic AI sits at the intersection of medical device regulation, data protection, and emerging AI-specific frameworks. In the United States, AI-based ophthalmic devices are regulated under the FDA’s existing device pathways, but the absence of a dedicated AI act means sponsors must navigate evolving expectations around algorithm change, real-time learning, and cybersecurity.⁶

In Europe and the UK, AI tools must comply with medical device regulations and data protection law and will increasingly fall under horizontal AI regulations that emphasise transparency, risk management, and human oversight. Professional bodies such as the Royal College of Ophthalmologists have called for iterative adoption with robust clinical governance, audit, and inclusive patient engagement to ensure safe and equitable implementation.⁷

A particular challenge for ophthalmology is how to handle “adaptive” algorithms that update over time as more images are acquired. Regulators and professional societies are still defining frameworks for lifecycle management, performance monitoring, and responsibility when an AI system’s recommendation conflicts with clinical judgment. SAIVO aims to be a neutral platform where clinicians, industry, and regulators can discuss practical standards and post-market surveillance strategies specifically tailored to eye care.

How do you envision AI integrating with teleophthalmology for global screening programs?

AI and teleophthalmology are naturally synergistic. In low-resource or remote settings, teleophthalmology programs already rely on technicians acquiring fundus images that are graded centrally. Embedding AI into this pathway allows for near real-time triage, identifying patients with sight-threatening diabetic retinopathy, AMD, or glaucoma, and directing them to appropriate care levels. Autonomous AI systems for DR screening have demonstrated that it is feasible to perform high-quality screening outside traditional eye clinics, including in primary care and pharmacy settings.⁸

For global screening programs, we see AI acting as a force multiplier rather than a replacement for ophthalmologists. Algorithms can pre-screen large volumes of images, flag high-risk cases, and standardize grading, while human experts focus on complex cases, treatment, and patient communication. At our SAIVO Symposium, Dr Yasir Sepah presented on AI-powered patient communication and triage, illustrating how conversational AI and automated triage can be coupled with image-based algorithms in teleophthalmology workflows. Coupled with home-monitoring solutions for nAMD (Prof Loewenstein), such finding points towards a continuous, distributed model of retinal care.

What role does explainability play in clinician adoption of AI-driven decision support?

Explainability is crucial for clinician trust and for integrating AI into evidence-based ophthalmic practice. Many ophthalmologists are understandably wary of “black box” systems that provide a probability score without showing why a decision was made. Concerns around accountability, bias, and medico-legal responsibility are magnified when we cannot understand or contest an algorithm’s output.⁹

Methods such as saliency maps, heatmaps on OCT or fundus images, and feature-level explanations can help, but explainability is not only technical. It also includes clear performance metrics, transparent validation on local populations, and understandable user interfaces that show levels of confidence and suggested next steps. Recent surveys of ophthalmologists indicate that trust, perceived utility, and education about AI strongly influence willingness to adopt these tools.¹⁰

Within SAIVO, we view explainability as part of a broader culture of augmented intelligence: AI tools should help clinicians see patterns they might otherwise miss, while leaving final responsibility and judgment with the human expert. Our symposium’s focus on imaging biomarkers, GA trials, and genetic retinal disease is deliberately paired with discussions about how to present AI outputs in a way that fits naturally into clinical reasoning rather than replacing it.

What future applications of AI could transform surgical planning or outcomes?

In vitreoretinal and cataract surgery, AI is poised to influence the full perioperative pathway. Preoperatively, predictive models can use multimodal imaging and clinical data to estimate surgical difficulty, likely intraoperative complications, and postoperative visual outcomes. This can inform patient counselling, case selection, and choice of technique or instrumentation.¹¹

Intraoperatively, we are beginning to see integration of AI with augmented reality and robotic platforms to assist with tasks such as identifying tissue planes, highlighting critical structures, or stabilising instruments. Research groups are exploring AI-assisted vitreoretinal surgery that could enhance precision and safety in complex detachment or macular surgery.¹²

Postoperatively, AI-driven analysis of imaging and visual function could enable earlier detection of adverse events and more tailored rehabilitation. For SAIVO, an exciting frontier is how surgical video and imaging archives can be used to train models that provide real-time feedback to surgeons, similar to flight simulators in aviation. Our course lays the foundation by focusing on high-quality imaging and analytics—key ingredients for these future surgical applications.

What is the broad mission of SAIVO? What steps—policy proposals, educational opportunities, etc.—will members take to achieve that mission?

The Society for Artificial Intelligence in Vision and Ophthalmology (SAIVO) was founded with a simple but ambitious mission: to ensure that AI in eye care is scientifically robust, clinically meaningful, and ethically and equitably implemented for patients worldwide.We would like for SAIVO to be as inclusive as possible of experts and colleagues who have significant interests and/or conduct studies and research in ophthalmic AI.

To achieve this, our activities fall into several pillars:

Education and capacity-building

• Started with the inaugural SAIVO meeting held 17 October 2025, in Orlando, Florida, USA, courses like our inaugural SAIVO Symposium at FLORetina—“AI in Ophthalmology: Breaking New Ground with Imaging and Analytics”—bring together retina specialists, uveitis experts, cornea surgeons, geneticists, data scientists, and industry partners to discuss concrete, real-world applications of AI, from clinical trials to home monitoring and telehealth. Other SAIVO symposia are being planned at upcoming congresses such as the 2025 Asian Pacific Vitreoretinal Society (APVRS) in Manila, 2026 Asian Pacific Academy of Ophthalmology in Hong Kong, and 2026 EURETINA in Vienna.
• We plan longitudinal educational offerings: online modules, hands-on workshops, and curated reading lists to help clinicians, residents, and researchers understand both the opportunities and limitations of AI.
• The 2026 SAIVO Annual Meeting will be held 1 May 2026, in Denver, Colorado, USA. Abstract submissions will open very shortly.

Standards, best practices, and policy input

• SAIVO aims to contribute to consensus statements on topics such as dataset quality and diversity, validation and generalisability, bias mitigation, and the role of explainable AI in clinical decision making, in alignment with broader professional guidance in ophthalmology.
• We seek dialogue with regulators, patient organizations, and industry to help shape practical frameworks for lifecycle management, post-market surveillance, and responsible integration of adaptive algorithms.

Collaboration and research

• By connecting clinicians who understand disease, engineers who build algorithms, and trialists who generate evidence, SAIVO wants to catalyse multicentre, multiethnic collaborations in areas such as GA, nAMD, ROP, uveitis, and inherited retinal diseases—many of which are represented in our FLORETINA 2025 program.
• We also encourage open science and data-sharing models that respect patient privacy but allow rigorous external validation and benchmarking of AI tools.

Ultimately, SAIVO’s vision is that AI becomes an integral, trusted component of eye care—one that extends, rather than replaces, the expertise and judgment of ophthalmologists, including retina specialists. Our first Symposium at FLORetina 2025 is a milestone in this journey: it sets the tone for a society that is practical, multidisciplinary, and firmly focused on patient benefits.

Marion R Munk, MD, PhD, FEBO and Quan Dong Nguyen, MD, MSc, FAAO, FARVO, FASRS

Professor Munk and Professor Nguyen are founding members of SAIVO.

References

1. IDx-DR - Healthvisors. Healthvisors. Published June 3, 2020. Accessed December 9, 2025. https://www.healthvisors.com/en/idx-dr/?utm

2. Lim JI, Rachitskaya AV, Hallak JA, Gholami S, Alam MN. Artificial intelligence for Retinal Diseases. Asia-Pacific Journal of Ophthalmology. 2024;13(4):100096-100096. doi:https://doi.org/10.1016/j.apjo.2024.100096

3. AI. SAIVO -- Society for AI in Vision and Ophthalmology. SAIVO -- Society for AI in Vision and Ophthalmology. Published 2019. Accessed December 9, 2025. https://www.saivo.org/

4. Zedadra A, Mahmoud Yassine Salah-Salah, Ouarda Zedadra, Guerrieri A. Multi-Modal AI for Multi-Label Retinal Disease Prediction Using OCT and Fundus Images: A Hybrid Approach. Sensors. 2025;25(14):4492-4492. doi:https://doi.org/10.3390/s25144492

5. Multimodal Deep Learning for Improved Disease Diagnosis in Ophthalmology. Github.io. Published 2023. Accessed December 9, 2025. https://multi-eye.github.io/

6. AI-Powered Ophthalmic Devices. AI-Powered Ophthalmic Devices. CRSToday. Published 2025. Accessed December 9, 2025. https://crstoday.com/articles/feb-2025/ai-powered-ophthalmic-devices

7. Artificial Intelligence in Ophthalmology Position Statement. https://www.rcophth.ac.uk/wp-content/uploads/2024/05/240521-Position-statement-artificial-intelligence.pdf

8. Cholakova M. AI in Ophthalmology, Existing Tools and What’s Coming. Ophthalmology24. Published May 8, 2025. https://www.ophthalmology24.com/ai-in-ophthalmology

9. Jin K, Yuan L, Wu H, Grzybowski A, Ye J. Exploring large language model for next generation of artificial intelligence in ophthalmology. Frontiers in Medicine. 2023;10. doi:https://doi.org/10.3389/fmed.2023.1291404

10. Christoph Tappeiner. Artificial Intelligence in Ophthalmology: Acceptance, Clinical Integration, and Educational Needs in Switzerland. Journal of Clinical Medicine. 2025;14(17):6307-6307. doi:https://doi.org/10.3390/jcm14176307

11. Poh SSJ, Sia JT, Yip MYT, et al. Artificial Intelligence, Digital Imaging, and Robotics Technologies for Surgical Vitreoretinal Diseases. Ophthalmology Retina. 2024;8(7):633-645. doi:https://doi.org/10.1016/j.oret.2024.01.018

12. The Future of Vitreoretinal Surgery | Duke Department Of Ophthalmology. Duke.edu. Published 2024. Accessed December 9, 2025. https://dukeeyecenter.duke.edu/news-and-events/vision-magazine/vision-2025/future-vitreoretinal-surgery