- Researchers in the United Kingdom have devised an innovative artificial intelligence (AI) program that uses retinal images to pick up signs of eye, heart, and neurological disorders.
- RETFound, one of healthcare’s first AI foundation models and ophthalmology’s first, used millions of eye scans to help detect and treat blindness.
- In multiple tests, RETFound surpassed existing AI systems and clinical experts in completing a range of complex diagnostic functions with less labelled data.
- RETFound also accounts for diverse populations and rare diseases, which many traditional scans and current AI systems often miss.
- Furthermore, this ‘transformative technology’ dramatically reduces the workload of human experts in analyzing and labeling retinal imaging.
Experts at Moorfields Eye Hospital and University College London (UCL) Institute of Ophthalmology in England have recently developed an AI system which can detect vision disorders more accurately and efficiently than current methods.
The scientists performed a study on RETFound, their world-first foundation model, which used millions of eye scans from the UK’s National Health Service (NHS). Their open-source initiative may serve as a template for efforts to help detect and treat blindness with AI.
“This is another big step towards using AI to reinvent the eye examination for the 21st century, both in the UK and globally. We show several exemplar conditions where RETFound can be used, but it has the potential to be developed further for hundreds of other sight-threatening eye diseases that we haven’t yet explored.”
The study appears in
A report from the British Chambers of Commerce recently referred to AI foundation models as “a transformative technology” for their use of massive quantities of data.
The launch of ChatGPT in November 2022 highlighted the potential of AI models to develop adaptable language tools.
RETFound took a similar approach with retinal images, training on millions of scans. This has enabled the construction of a versatile model for virtually unlimited uses.
AI models have largely depended on human expertise and effort. Medical News Today discussed the challenge with technology developer Dr. Steve Frank, founder of Med*A-Eye Technologies. He was not involved in this research.
Dr. Frank explained to MNT: “AI is data-hungry, and teaching an AI system to perform tasks generally requires vast amounts of training data. Worse, training usually requires the data to be labeled in some way — meaning that you’re teaching the system to distinguish one thing from another based on examples that you tell it are one thing or the other. That’s traditional ‘supervised’ learning.”
Furthermore, Dr. Frank said, experts may disagree on a piece of data, requiring time-consuming expert panel reviews.
According to the UK researchers, RETFound can match the performance of other AI programs using only 10% of human labels in its dataset.
RETFound achieved this higher efficiency with its self-supervising approach of masking parts of an image and learning to predict the missing parts by itself.
“Self-supervised learning (SSL), which underlies RETFound, dispenses with labeling altogether. With enough training data, a properly structured AI model can learn enough about the training data from the data itself to make meaningful predictions […]This approach is of particular value for healthcare AI because the cost of labeling is so high — doctors are already busy saving lives, and their time is quite precious.”
– Dr. Steve Frank
A 2023 review in the Journal of Clinical Medicine refers to the retina as “a window to the body”. The study of oculomics uses deep learning to explore correlations between retinal image characteristics and diseases.
The program could also predict systemic disorders including heart failure, stroke, and Parkinson’s disease.
Moreover, this AI technology facilitates a non-invasive view of the nervous system.
MNT discussed this study with Atropos Health co-founder Dr. Brigham Hyde, who was not involved in this research. We asked him how AI and deep learning techniques can help with detecting diseases.
“First, imaging techniques aided by AI can often detect diseases a human may miss. Second, AI and deep learning techniques applied to combinations of digital, medical, and experiential data can uncover digital biomarkers for disease leading to earlier diagnosis,” he responded.
”Lastly,” he added, ”risk scoring algorithms deployed at the physician’s office can highlight and direct care teams to patients with key risk factors earlier.”
The present study employed and evaluated RETFound, a new SSL-based foundation model for retinal images. The authors described a foundation model as “trained on a vast quantity of unlabeled data.”
In this case, Prof. Keane and his collaborators trained the AI system with a dataset of 1.6 million images from Moorfields Eye Hospital.
“We adapt RETFound to a series of challenging detection and prediction tasks by fine-tuning RETFound with specific task labels, and then validate its performance,” their paper reads.
The team considered ocular diseases including diabetic retinopathy and glaucoma, and ocular disease prognosis, in a 1-year period.
Next, they studied a 3-year prediction of heart diseases such as stroke, heart failure, and myocardial infarction, and Parkinson’s disease.
Compared to models pretrained on SL-ImageNet, SSL-ImageNet, and SSL-Retinal, RETFound demonstrated “consistently superior performance and label efficiency.”
Dr. Frank remarked: “The RETFound results are especially impressive for the sheer number of tasks their system can perform. The accuracies the researchers achieve aren’t sufficient for clinical use, but the more conventional systems they test against are mostly worse.”
The UCL-Moorfields experts said that RETFound showed equal effectiveness in finding disease across diverse ethnic groups.
PhD researcher Yukun Zhou, the study’s lead author, mentioned in a press release: “By training RETFound with datasets representing the ethnical diversity of London, we have developed a valuable base for researchers worldwide to build their systems in healthcare applications such as ocular disease diagnosis and systemic disease prediction.“
Dr. Tyler Wagner, vice president of biomedical research at Anumana, not involved in the research, had this to say about the study: “While RETFound performs better than the other models compared in the manuscript during external evaluation on a set of patients with different demographics, the authors note the decrease in performance, highlighting the importance of the patient diversity in model development.”
The study authors hope that their finding will encourage further studies, writing: “Finally, we make RETFound publicly available so others can use it as the basis for their own downstream tasks, facilitating diverse ocular and oculomic research.”