BY BÁRBARA PINHO | 18 OCTOBER 2023
An AI tool could diagnose a range of health conditions from eye scans.
In a study recently published in Nature, an international team of researchers introduced RETFound, an AI tool to diagnose multiple health conditions from eye scans. The technology outperforms existing AI disease-detection models.
‘People, both in the UK and globally, are going blind because of delays in being seen and treated,’ says Pearse Keane, an ophthalmologist at UCL, UK, and one of the study’s authors. ‘And so, as a doctor, I was trying to think of new ways to address these issues.’
Keane and his team built the model by training it with thousands of eye scans from the Moorfields diabetic image dataset, as well as other public datasets. Moorfields is a specialist NHS eye hospital in London. Because labelling each scan by hand would be too laborious, they used a training technique called self-supervised learning to train the model with the raw data. The creators of ChatGPT use a similar approach to train the large language model.
‘We’ve trained it [RETFound] on just under 2m retinal images, and then we fine-tuned it. So, it can be used in different downstream clinical tasks like diagnosing diabetic retinopathy or predicting the progression of age-related macular degeneration,’ he adds.
Besides detecting ocular diseases, RETFound predicted other health conditions such as Parkinson’s disease, heart attacks or strokes, as eye scans can provide information on other systems of the human body.
Keane and his colleagues are looking forward to seeing RETFound in action but are careful to manage the expectations and excitement associated with AI. ‘We’re trying to balance it with caution, and scepticism and saying, “Look, we can’t deploy these systems until they’re actually safe and have been well validated in the real world”,’ he adds.
To Veronika Cheplygina, an associate professor at ITU Copenhagen, Denmark, more testing would have made the study more complete. ‘The idea of semi-supervised learning has been around for a while; I think this is a nice study due to the variety of data/tasks involved, but wouldn’t say it is a paradigm shift,’ she says.
She also adds subgroup analysis – seeing how the tool performs in male vs female scans, for example – would be useful. ‘This has been shown to be a problem in other machine-learning based methods.’
Visit Blog
