Topics covered by speakers in the Saturday afternoon Euretina session on artificial intelligence (AI) covered model development, applications, and cautions about the need for developers and users to behave safely and responsibly.
Paul Nderitu, MD, PhD (England) provided an update on Global RETFound, a large international consortium study that is aiming to create the world’s first global medical foundation model using a novel synthetic image privacy-preserving data sharing strategy. It is also seeking to improve the representativeness of AI datasets used for developing ophthalmic AI foundation models and increase both data scale and diversity. Dr. Nderitu’s takeaway message was that synthetic data/sharing is driving a new approach to scaling and diversifying data for AI foundation model development. He expressed the hope that Global RETFound and its novel approaches will become a blueprint for other medical specialities working to build foundation models on a global scale and that its inclusivity will help to improve AI health equity.
Rahul Jonas, MD, PhD (Germany) spoke about clinical applications for AI in pathologic myopia. He listed seven applications – image quality assessment, classification, lesion localization and segmentation, myopic macular neovascularisation detection, longitudinal progression analysis, individual risk prediction, and axial length estimation and described various challenges for developing AI models for these tasks. As conclusions, he noted that existing classifications for pathologic myopia and myopic macular degenerations have gaps that create problems for AI training purposes; there is a need for improving axial length estimation from fundus imaging to improve model precision; and quality assessment for retinal imaging should always reflect real world clinical data.
Aude Couturier, MD, PhD (France) shared findings from EviRed, which is a multicentre study conducted in France with the aim of developing an AI algorithm integrating multimodal imaging and clinical data to predict the risk of progression to proliferative diabetic retinopathy or diabetic macular oedema in eyes with diabetic retinopathy. She noted that a tool providing more precise risk assessment could enable personalised treatment and follow-up decisions. After discussing the results, Dr. Couturier provided a list of work that remains to be done, including algorithm validation on an external dataset.
Aliénor Vienne-Jumeau, MD, PhD (France) spoke about where AI can enter the adaptive optics (AO) workflow. She explained that AO enables high-resolution in vivo retinal imaging by correcting ocular aberrations, but the increasing speed and reproducibility of image acquisition have shifted the principal bottleneck toward robust, quantitative, and scalable image analysis. Her examination of the integration of AI into AO workflows focused on automated segmentation and quantification of retinal vascular structures, reconstruction of capillary networks from AO-OCT/AO-OCTA, photoreceptor detection and spatial analysis, RPE cell characterisation, quality control, and longitudinal image registration. Dr Vienne-Jumeau’s main take-home message was that AI has the potential to transform AO from a high-resolution imaging modality into a reproducible quantitative platform for cellular-scale phenotyping and longitudinal assessment of retinal microstructure.
In an update on generative AI for retina, Daniel Ting, MD, PhD (Singapore) noted the speed at which AI models are being developed and issued words of caution about the need for users and developers to proceed safely and responsibly. He pointed out risks of AI and potential benefits, including increasing workflow efficiency and the possibility that applying AI could shorten the life cycle for new drug development. Dr. Ting concluded his talk by presenting the S.A.F.E.R Strategy – Safety, Accuracy and accountability, Fairness and ethics, Ease of use, Return on investment – as a framework for safe and responsible AI.
The session concluded with a talk by Lorenzo Ferro Desideri, MD (Switzerland) who provided some practical ideas about integrating AI predictive models into the clinical workflow of retina practice. Dr. Desideri emphasised that predictive models matter only when they reach the patient, and he noted that reliable standardised ground truth is the foundation of any clinically useful model. Addressing agentic AI, Dr. Desideri said that clinicians can increasingly build and update models themselves, always with the expert in the loop.
