Harry de Quetteville
Can Artificial Intelligence solve the cancer crisis?
Artificial intelligence (AI) has long been on a reputational whiteknuckle ride. One day it is hailed as a revolutionary cure-all. The next, as such promises go unfulfilled, despairing researchers walk away. It is a “hype cycle” that has plunged the field into repeated “AI winters” ever since Alan Turing proposed his test of machine intelligence in 1950.
When it comes to treating cancer, however, researchers insist that this time, the AI hype is justified. “I’m very excited,” says Charlie Swanton, the Royal Society’s Napier Professor of Cancer Medicine. “AI is ready for prime time now. It’s already delivering.”
The NHS has just made a computer tool called DERM that diagnoses skin cancers available to doctors across the country, for example, after twin pilot schemes showed it greatly reduced unnecessary referrals to hospital. “This is happening now,” says Neil Daly, CEO of Skin Analytics, the company behind DERM. “Not in three years. Not in five years. Now.”
Perhaps the most significant evidence of this is the range of cancers that AI – computer programmes which learn from existing data to spot problems and make connections often beyond human calculation – is now being used on. In June, it was breast cancer, with government funding granted to an AI mammogram screening project that has shown to be as effective as human radiologists; in May it was lung cancer, which a study found AI was also able to predict as well as the experts; in April came the news of AI being deployed to track mesothelioma, a cancer linked to asbestos exposure that is hard to measure because it surrounds the lungs rather than growing spherically.
Last month it was announced that AI is being used to predict the risk of notoriously hard-to-spot ovarian cancer, and that it is aiding early diagnosis of oesophageal cancer. The variety of ways it can help is enormous: in the case of ovarian cancer, it is being unleashed on huge
Researchers soon will be able to identify the women most at risk of ovarian cancer
data sets of genetic information from volunteers held at the UK Biobank in Stockport, spotting links so that researchers soon will be able to identify women who are most at risk before they develop the disease. For oesophageal cancer, AI is scanning images from endoscopies for the subtle changes in throat tissue colour and pattern that can be a concern. Trained on hundreds of thousands of cases, it can flag problems that doctors might miss. Some studies suggest that up to a quarter of early cancers in the oesophagus are missed during standard examinations.
And early diagnosis is key. As oncologist Karol Sikora says: “Stage 1 cancer, 90 per cent cure rate. Stage 3 cancer, 20 per cent cure rate.” No wonder the NHS, stretched by the pandemic, has just given the green light to 38 new projects deploying AI, many in diagnosis, all part of its £140 million “AI in Health and Care Award”.
Such advances arrive as the NHS battles a post-covid cancer crisis – with recent statistics showing 40,000 fewer people started cancer treatment last year than normal, a statistic that inevitably will lead to many deaths – and a shortage of oncologists. Research by the Royal College of Radiologists shows that this year’s newly trained consultants will only fill 55 per cent of vacancies.
Helping doctors with diagnosis, or taking the rote work of mass screening off their hands altogether, is one obvious way of easing pressure and clearing backlogs. “Reading mammograms, it’s so dull,” says Sikora. “In a day you’re doing hundreds. And boredom doesn’t always lead to good results.” Indeed it is in the areas of pathology and radiology, where experts rely on huge experience and training to spot disease in tissue samples or scans, that AI can make the most obvious inroads.
“AI is particularly good at pattern recognition,” says Prof Swanton. “It can discard all the benign cases very quickly,” leaving humans to focus on those that really need attention. “There’s a national shortage of radiologists. An AI tool filtering out cases they don’t need to look at would save lives.”
Early diagnosis is also the motivation behind one of the most startling deployments of AI in cancer detection – that of a simple blood test developed by the American company Grail. Swanton, who is also an adviser at Grail, says it works by detecting fragments of genetic material from “nascent tumours”, sometimes before any other signs or symptoms of the cancer have emerged. Critically, the AI can detect not only the presence of very early cancers, but also, through changes to those cancers’ DNA, locate where in the body they are developing. It is thought to pick up more than 50 types of cancer and is being piloted in the NHS from this autumn.
Yet, predictably, it is not all good news. One is AI’S so-called “black box” issue. “While it’s very impressive, you don’t know how it’s come up with a solution,” says Swanton. “That’s a concern and will need to be resolved.” On separate AI breakthroughs into renal failure by Google’s Deepmind, he says: “It clearly works, but we don’t know how or why.”
He thinks we humans “will inevitably give in to computers in the end, but the concern I have is that clinicians and scientists like to know how things work.” And patients, too, presumably, whose trust in new AI procedures will be key to widespread adoption. Swanton’s worry with AI’S “black box” is that “if things go wrong, it might be quite hard to disentangle why”.
The other potential pitfall for AI is the data that it is trained on. “Rubbish in, rubbish out” is the usual AI mantra. If training data are garbled, results will typically be unreliable. More pernicious, if such data are the result of pre-existing bias, then bias can be baked into future AI results. This can be as simple as algorithms trained on skin cancer images from white patients struggling to detect life-threatening lesions on those with darker skin. It is a problem that everyone in the industry is aware of.
“It’s an exceptionally difficult problem to solve,” says Daly, explaining that those with darker skin suffer less from skin cancer and are a minority in Western countries “which makes it more and more difficult to get hold of the [training] data.” Skin Analytics, he
‘There’s a shortage of radiologists. An AI filtering tool for cases would save lives’
says, has been teaming up with clinics in Africa to source data “but it’s complicated”. He insists that by “getting the data into the system, the fixes will come”.
As such cases show, the ethics of AI generally is a huge emerging field. But the stakes are particularly high in cancer care – a matter of life and death. “It’s not like a chatbot at a bank. We’re talking about decisions that dramatically impact patients,” says Daly. “We’ve always really been aware of the opportunity but also the risks associated with what we do.”
No system is perfect. So how good does AI have to be to gain widespread trust? As good as existing doctors?
Better? Daly says dermatologists spot 88-90 per cent of melanomas. “We think we should set our level at 95 per cent.”
None of which should spell the end for the men and women in white coats. “Doctors won’t go out of business,” says Sikora. “AI will allow them to do more.” Swanton adds: “The fact is patients don’t want to see a computer. They want to see a human being.” But they doubtless want to see a human being who can give them answers. And for all medical progress, which in the developed world has pushed cure rates up from a third to a half in the last 40 years, cancer remains a mystery. “I’ve got patients that have survived 25 years with metastatic cancer,” says Sikora. “What’s different about them? I want the AI to tell me, why are they surviving? It can’t be the power of their mind or anything like that. It has to be something physical. We just don’t know.”
As it crunches ever bigger numbers, wades through ever more data, AI may be able to spot the otherwise imperceptible reasons why. The NHS, with its huge, homogeneous system, its 350,000 new cancer patients each year, is perfectly placed to reap the benefit. After the last year, it will have to. “History is littered with technologies of value that never got realised,” says Daly. “But this time, with where the healthcare system is at the moment, the pressure and the pandemic, AI won’t get buried. There’s just too much need.”