An eye for AI – state of the art digital technology to improve healthcare
‘You can lose your eyesight in just a few days,’ says Professor Sir Peng Khaw, consultant ophthalmic surgeon at Moorfields Eye Hospital NHS Foundation Trust and the UCL Institute of Ophthalmology. ‘So what we are developing now and in the future is, ultimately, a way to prevent unnecessary blindness.’ Sir Peng is describing Moorfields work with DeepMind – creating an artificial intelligence (AI) tool designed to provide very rapid, highly expert assessment that can move people into the right treatment as quickly and efficiently as possible, and minimise their chance of going blind.
The work is taking place against a backdrop of what Sir Peng describes as ‘an explosion in the number of people with eye diseases’, including macular degeneration, and diabetic retinopathy. ‘I recently chaired an Eye Health consortium for Lord Darzi and the WISH foundation, which highlighted that the increase is caused by our rapidly ageing population, compounded by lifestyle factors,’ he says. ‘We’ve now got a huge number of people who need to be screened, diagnosed and decisions made about whether they need treatment, often very rapidly. Worldwide, the demand is beginning to overwhelm services, and it’s only going to get worse.’
The Moorfields team, led by Dr Pearse Keane, consultant ophthalmologist at Moorfields Eye Hospital NHS Foundation Trust and NIHR Clinician Scientist at the UCL Institute of Ophthalmology, wanted to develop a way to read the complex retinal scans automatically and then make treatment referral decisions on the spot.
This ‘on the spot’ feature is a vital part of the problem, as running multiple appointments not only involves treatment delays and possible loss of vision, with associated anxiety and disruption to patients’ lives – but it also increases clinic load, and hence cost. ‘There are huge advantages in having diagnosis on the spot,’ says Sir Peng. ‘But to do that, you need expertise. And we don’t have enough expert consultant ophthalmologists even in the UK, let alone some of the largest countries in the world, like China and India.’
Birth of the project
As they explored the problem, the Moorfields team began to realise that AI could provide an answer. So Pearse Keane approached the UK company DeepMind – often considered the world’s leading AI research company – in 2016 to discuss the possibility of working with them. And so the project was born.
‘Our role is to provide the clinical expertise and patients, while theirs is the AI and also considerable expertise in clinical interface technology’ explains Sir Peng. ‘It’s that combination which is amazing: we couldn’t do what they do, and they couldn’t do what we do.’
For DeepMind, the project offers a lot of scope. As the largest eye hospital in the developed world, Moorfields can provide rich data, with more than 700,000 patient visits in the past year alone, and is also the most productive eye research site in the world, together with UCL. The quality of this dataset makes it easier to apply AI technology to read them.
How it works
The process works by gathering a set of anonymised scans, each containing millions of pieces of information Then the algorithm ‘segments’ the different types of macula retinal pathology in the scan. A much larger clinical data set, with confirmed diagnosis and referral decision available, was used to train the algorithm. This trained algorithm was then compared against eight clinical experts and was able to match the two best-performing experts.
‘Imagine coming to see a doctor who’s an expert in something because they’re very experienced and then multiplying that diagnostic experience by many more lifetimes – well, that’s what this tool can do,’ explains Sir Peng. ‘The advantage of the algorithm is that it doesn’t forget.’ And because it can continue to learn, it is agile, so it does not simply apply a formula. In that way, is it not dissimilar to how real human clinicians learn.
Importantly, the tool also shows its workings, so that healthcare professionals can understand and check how it reached its assessment. It produces a multicolour picture showing all the different pathologies on the scan and how it came to its conclusions. Like a doctor, it gives a differential diagnosis and probability, and can then triage cases into urgent, semi-urgent routine and observation only, and calculates the chance of each of these suggestions being correct. This transparency reflects DeepMind’s interest in addressing the ‘black box problem’ in medical technology – a commitment to show its workings. This is really important, says Sir Peng, ‘because the worst thing about AI could be that you don’t know why the machine is telling you a diagnosis and you triage or treat without understanding why.’
In August 2018, the first results were published in Nature Medicine, showing that the system could match the highest performing expert clinicians, diagnosing more than 50 eye conditions and referring with 94 per cent accuracy.
The team is now planning pathway studies to make sure the technology will genuinely make a positive impact – for example, making sure people are referred as rapidly and accurately as possible, to enable more people to be seen with an even higher quality of care. ‘It’s one thing having the world’s most intelligent system, but the technology is useless unless you refine the patient pathway too,’ says Sir Peng. ‘That’s where UCLPartners really helps innovators – supporting research in to how health services and systems can run even better.’
The innovation will have obvious benefits locally, but Sir Peng’s vision is for algorithms like this to become part of the standard eye toolkit. to ensure rapid, accurate treatment worldwide: ‘Our mission for the eye division is “Changing Lives in London, Britain and the World”,’ he says. ‘Some people think we are in ivory towers, but you need high-tech expertise to develop this sort of innovation which can then be deployed on a laptop on a more achievable scale to a much wider population throughout Britain and around the world. The only way we’re going to deal with the huge and increasing worldwide disparity between demand and resource is through innovation.’