Spearheaded by researchers from the National University of Singapore, the Chinese University of Hong Kong, Moorfields Eye Hospital, and University College London, more than 100 study groups across 65 countries have joined forces to launch the Global RETFound initiative. The collaboration aims to create the world’s first globally representative artificial intelligence (AI) foundation model in medicine, trained on over 100 million eye images.
To learn more about the initiative and its potential to transform medical AI, The Ophthalmologist spoke with Paul Nderitu, Medical Retina Fellow at Moorfields Eye Hospital NHS Foundation Trust, Senior Research Fellow at UCL Institute of Ophthalmology, and co-author of the Nature Medicine paper describing the project.
What initially motivated the creation of the Global RETFound initiative?
In September 2023, our research group, led by professor Pearse Keane at Moorfields Eye Hospital and UCL, developed RETFound, the first foundation model in ophthalmology, trained on anonymised imaging data from Moorfields. The foundation model was made freely available for non-commercial research, which allowed vision researchers worldwide to utilize it for developing AI tools suited to their requirements. At the ARVO 2024 conference in the US, a discussion with colleagues from the National University of Singapore and the Chinese University of Hong Kong led to the idea of expanding RETFound’s training data to achieve global representation. This discussion initiated the Global RETFound project. Global RETFound involved reaching out internationally through research networks, ultimately recruiting over 100 study groups to participate. Concurrently, tools for finetuning generative AI models and a two-tier data sharing protocol were developed, considering that some research groups may have limited computational and technical resources. The process from the initial concept to the official launch of the initiative at ARVO 2025 took approximately twelve months.
Why is building a globally representative dataset so critical for medical AI, particularly in ophthalmology?
There is considerable under-representation of certain populations – particularly those in Southeast Asia, Central Asia, Latin America, Africa, and the Middle East – in current AI research and ophthalmology, relative to the prevalence of eye disease within these regions. Variations in population demographics, disease prevalence, and imaging technology may limit the effectiveness of AI models, especially in groups that are presently underrepresented. This lack of diverse representation could perpetuate existing health inequities. Consequently, there is an urgent need for large-scale collaboration and innovative data-sharing practices to establish globally representative datasets for AI development. The Global RETFound initiative has the potential to serve as an exemplar for establishing international AI collaborations for other medical specialties.
What safeguards are in place to protect patient privacy when data is shared across continents?
Two robust yet adaptable data-sharing frameworks are employed by the Global RETFound consortium. The first data sharing strategy utilises synthetic data generation through local finetuning of a generative AI model, with only the model weights being securely shared, allowing for the use of synthetic data. No real data is exchanged, and all the source data remains within originating institutions, ensuring maximal patient anonymity. The second data sharing strategy consists of the transfer of anonymised real data within a protected infrastructure. Transfers are conducted via secure cloud-based platforms, governed by stringent protocols. Both data-sharing sharing frameworks are supported by comprehensive governance procedures and legally binding agreements.
Beyond ophthalmology, what systemic diseases might Global RETFound also help to detect?
The initial ophthalmic foundation model, RETFound, was assessed for its effectiveness in predicting ischemic stroke, myocardial infarction, heart failure, and Parkinson’s disease, as well as ocular conditions. We also aim to assess the performance of the Global RETFound foundation model in detecting or predicting cardiac and neurological conditions. Additionally, we will assess for the detection of kidney disease, and the foundation models’ ability to extract clinically relevant biomarkers from retinal images, including HbA1C for diabetes mellitus and blood pressure for hypertension.
Beyond our planned internal evaluations of Global RETFound, we aim to release the foundation model as an open-access non-commercial research tool, and encourage other groups to assess its utility in detecting or predicting a wide range of systemic and ocular diseases. Given the large and diverse dataset that will be used to develop Global RETFound, we believe the foundation model will demonstrate better performance on systemic disease detection and prediction tasks, in turn advancing the emergent field of Oculomics.
How will clinicians and researchers around the world – especially those in lower-resource settings – be able to access and use the model once it is released under a Creative Commons licence?
Our objective is to make the foundation model highly accessible. As with RETFound, researchers will be able to obtain the model through open access platforms, such as GitHub and HuggingFace, for non-commercial purposes. Comprehensive instructions and user guides will be available on the release portal to facilitate straightforward downloading and implementation.
An important benefit of the foundation model, particularly in resource-limited environments, is its strong 'out of the box' performance. Users may utilize the foundation model without additional finetuning of the whole model by leveraging features extracted by the model for downstream tasks, such as disease detection, using computationally efficient techniques like 'linear probing.' This approach substantially reduces computational requirements, making it well-suited for settings with limited resources. Furthermore, we anticipate that this will enable broader adoption of on-device and edge applications for mobile devices. We also hypothesize that Global RETfound will demonstrate improved 'out of the box' performance and generalizability owing to its larger and more representative pretraining dataset.
Do you think this type of initiative could become a blueprint for global foundation models in other specialties?
Yes, that is one of our primary objectives in developing this initiative. We are currently preparing a manuscript of our overarching study protocol, which will provide comprehensive details about our consortium framework, resources used to develop our secure data sharing strategies including synthetic AI models, and the foundation model development and validation methodology. Our intention is for other medical specialties to use these resources to create similar initiatives within their domains. Additionally, we anticipate that as consortium ophthalmologists become familiar with the data sharing techniques, they will help to disseminate these techniques to their colleagues across various medical specialties worldwide.
Is there anything else you would like to add?
We invite researchers interested in participating in the Global RETFound collaboration to contact us directly at paulnderitu@nhs.net (Dr Paul Nederitu) or alex.black3@nhs.net (Alex Black, Research Communications Lead at Moorfields Eye Hospital). We view this as a multi-year and continually expanding initiative that will advance global AI health research.
