more_vert
An artificial intelligence algorithm used to detect breast cancer in screening scans can predict women’s cancer risk over the next four years, Australian research has found.
The research published in the Lancet Digital Health journal found that the AI-based risk score, called BRAIx, identified women at high risk of breast cancer more accurately than traditional factors doctors rely on, such as breast density and family history.
Nearly one in ten of those high-risk women who scored in the top 2% by the tool were diagnosed with breast cancer, despite being given the all-clear by other methodologies.
The AI tool was developed using mammograms from nearly 400 000 women, then tested on data from 96,000 women in Australia, and further confirmed in an independent Swedish population of more than 4500 women.
Lead researcher University of Melbourne Associate Professor Helen Frazer, who is also a clinician at St Vincent’s Hospital and breast radiologist, has worked in the population screening program for more than 20 years across three Australian states.
“We’re lucky in Australia. We’ve got a national programme. It’s been going for over 30 years. So we’ve got these fantastic data sets that are unique globally,” Associate Professor Frazer says.
“In the screening programme, we’re able to work with AI as a partner to come in and work alongside, but still have every single clinical decision made by a medical specialist. And I think that’s a really important thing for the readers is that there’s no autonomous component. A human is in charge of every single decision.”
Associate Professor Frazer started working with AI in image classification for detecting cancer on mammograms about ten years ago, but says this development is incredibly significant.
Although more studies are needed before it is considered for use in routine care, she says this AI tool could improve early cancer detection, and reduce false alarms, all without increasing costs.
“We do find that unlike the human who is confused or confounded by breasts that have high mammographic density, the algorithms work effectively across all densities,” she says.
“And it’s an area that we want to study more closely, but that’s very exciting because the algorithms manage to learn density, but at a pixel level, still be able to identify well beyond the scope of human visual acuity patterns, patterns that might signal future risk, increased risk as well as very early cancer.”
“So that’s the really exciting area that we would like to push into prospective studies of the risk tool to see if we can personalise the program and basically significantly improve cancer outcomes.”
She says the BRAIx risk score has the potential to reduce deaths from breast cancer by detecting cancers that could otherwise be missed, particularly “interval cancers” that present between screenings, which often have the worst outlook.
“We really want to make sure we can make a dent in that,” she says. “We can pick them up earlier so that they don’t present late and we can hopefully have much broader treatment options for those women and also save their lives through earlier detection.”
“What we discovered during the studies is that it’s also able to pick up or to detect signals for risk of breast cancer and even very early breast cancer that might not be actionable to the radiologist, might not be visible yet in the population datasets.
“That’s a really, really important finding because the population program is one size fits all, a mammogram every two years, it’s not personalised to women’s individual risk of developing breast cancer.”
While population breast cancer screening has been very successful, cutting breast cancer deaths by around 40-50 per cent in women aged 50-74, it still largely takes a one-size fits all approach without taking into account personal risk.
Associate Professor Frazer says the AI-based risk scores have the potential to make breast screening more personalised, helping to identify women at high risk of developing breast cancer while also identifying those at very low risk who may require less frequent screening.
“That will enable us to be able to take those very first steps to personalise the population screening programme, with improved detection of cancers earlier within those same funding envelopes that we have,” she says.
“So that’s the game changer that can hopefully give us that quantum step to improve, to be able to detect every cancer early and save every life.”
Associate Professor Frazer says the significance of the findings are that we can now see a future use of AI to personalise screening, as they embark on two years of randomised controlled trials.
“It is the first randomised controlled trial for AI, working with clinicians, in a high consequence healthcare decision setting at scale within Australia, we believe,” she says.
“And it’s one of the first few globally. So it is a really exciting moment.”
“I think the era of double reading of a mammogram by radiologists and then an arbitration read if they’re discordant, I think that era is coming to a close.”
“I think the algorithms, the technology is definitely at that level of performance where it can be safely implemented.”
“We do know through screening that if cancers are detected early before they spread outside the breast or to lymph nodes, that every life can be saved.”
“And I think that’s a big motivator for all of us to improve outcomes.”
Nance Haxton was a journalist at the ABC for nearly 20 years. She’s also worked as an Advocate at the Disability Royal Commission helping people with disabilities tell their stories and as a senior reporter for the National Indigenous Radio Service.
In that time she’s won a range of Australian and international honours, including two Walkley Awards, and three New York Festivals Radio Awards trophies.
Now freelancing as The Wandering Journo, Nance is independently producing podcasts including her personal audio slice of Australia “Streets of Your Town”.
Subscribe to the free InSight+ weekly newsletter here. It is available to all readers, not just registered medical practitioners.