Where will AI take us? Photo / Getty Creative
“Hungary has become a major testing ground for AI software to spot cancer, as doctors debate whether the technology will replace them in medical jobs,” the New York Times reported earlier this week.
Trials have found that AI (artificial intelligence) software “is showing an impressive ability to spot cancer at least as well as human radiologists”. It can even spot signs of potential cancer that radiologists miss when manually analysing X-rays or scans.
And we’re just at the beginning. The sudden popularity of ChatGPT from OpenAI - co-founded by Elon Musk and recently fuelled by a US$10 billion investment from Microsoft - has started an AI gold rush in Silicon Valley targetting healthcare, and many other fields.
But med-tech experts here say AI has the potential to help medical professionals analyse scans faster - helping save them time in an age when most clinicians are overworked, and most hospitals are short of doctors.
Another factor is that while some AIs are clever - their smarts, for now, are in a very narrow field.
“We’re just now at the point of operationalising narrow AI,” says Will Hewitt, founder of HeartLab - an Auckland-based startup that’s developing AI to automate analysis of echocardiograms (heart scans), backed by millions in venture capital funding from the likes of Peter Thiel’s Founders Fund and the Peter Beck-backed Outset Ventures.
“AI models have been designed to detect one single specific pathology,” Hewitt says.
“Human doctors have the far harder problem of finding all sorts of different pathology - then integrating that with the non-imaging information they have on the patient - to ultimately decide what is best for the patient.”
There’s also what might be called an emotional element. Just as a Tesla crashing in self-driving mode causes more outrage than a human driver totalling a car (however much Elon Musk touts statistics that the former is far more rare, per million miles driven), people could have a much lower tolerance for a medical blunder caused by an AI.
“There’s a risk equation that we just haven’t decided on yet - humans make mistakes, and maybe in some narrow use-cases an AI model performs just as well as a human, but we’re ultimately more comfortable with human error than AI error,” Hewitt says.
He adds, “We’re still very much at the stage of AI-assisted doctors, I don’t see doctors wholesale losing their jobs to an AI model anytime soon. The job of a doctor goes well beyond looking at an image and pointing out some disease.”
New Zealand already uses AI tools in daily clinical practice, Hewitt says.
“Look at Canterbury DHB’s and their use of Blackford Analysis, New Zealand companies like Formus Labs, Volpara Health and HeartLab have also been driving the adoption of this technology.”
Blackford - owned by the multinational Bayer - makes an AI platform for analysing CT and MRI scans.
Wellington-based, ASX-listed Volpara makes software for analysing breast density - a key indicator of cancer risk - and recently entered an AI partnership with Microsoft.
Auckland-based Formus Labs raised $7m to fund its development of AI software to help and assist orthopaedic surgeons.
Hewitt says all are complimentary to carbon-based health professionals.
“The right answer is combining the two; use the broad experience of a human doctor, and the high precision of an AI model.”
Human clinicians’ job insecurity won’t be helped by the Times article, which quotes Dr László Tabára, a leading mammography educator in Europe saying, “I’m dreaming about the day when women are going to a breast cancer centre and they are asking, ‘Do you have AI or not?’”
Tabára can stop dreaming.
“This is happening now,” Volpara CEO Teri Thomas says.
“We have a US customer - Radnet - direct marketing to women an AI analysis of their images alongside the radiologist analysis. It’s called EBCD, which stands for Enhanced Breast Cancer Detection. We are proud to be part of this and are supporting other customer efforts to provide all of the tools at our disposal to the best cancer detection and even prevention (via routing appropriate women for genetic testing.”
Thomas says there has been a fear that AI will replace radiologists for breast screening since the advent of the first CAD (computer-aided design” application in mammography in 2003.
“This 20-year fear has ebbed and flowed, and likely has had an impact on radiology as a choice for new medical school residents in the US, as it is an area experiencing a current staffing shortage,” Thomas says.
But she says the reality is that healthcare providers are struggling to catch up with huge backlogs
“Instead of radiologists sitting around with time on their hands because of the work AI takes off of their shoulders, we find that they are behind, or struggling to keep up with their workloads.” She sees AI software moving from “nice-to-have” to “must-have” as they scramble to catch up.
Here, Volpara has something of an ace up its sleeve via its research partnership with Microsoft, inked in June last year; relatively recently, just before the world went into a frenzy over the Microsoft-backed ChatGPT.
“I just had a great call with the Microsoft team and that relationship is healthy and growing,” Thomas told the Herald earlier this week.
“We’re exploring opportunities with them related to their big investment in AI and actually doing a joint presentation on AI later this week in Australia.”
Volpara’s data science head is applying Microsoft tools to the Kiwi firm’s library of 75 million images, the better to train new algorithms.
Thomas added, “My belief is that radiologists will be supported, and their skills supplemented by AI.”
There are two types of AI you will see in our space - diagnostic and operational, she said.
“The operational type of AI is interesting right now. The kind of AI you see with ChatGPT is more operational. It will speed up doctors in a way that goes beyond highlighting suspicious areas - the realm of traditional CAD - to enhancing and speeding up their workflows. Think about things like what time of day you’re most ‘on it’ and how you focus your time. Do you want to read the most challenging images during your post-lunch slump? Radiologists miss cancer sometimes. That’s just a fact. Can AI help reduce how often that happens? Perhaps.”
Thomas sees AI replacing humans for simple “reads” - those where there is a clear and simple image, which she says will free up radiologists to take the time needed for a thorough read of the cases that require the most skill.
“But the ‘art’ of combining image analysis with the assessment of the person in the room, their comorbidities and family history will still live with a medical doctor for a long long time,” she says.
"I was really wrapped up with what I was doing was who I was."