CHICAGO, USA—Given the chance, a substantial number of women will opt for artificial intelligence (AI) enhanced mammograms, even if they have to pay for them on their own, a study presented today at the annual meeting of the Radiological Society of North America finds. And the use of AI appears to be worth the extra cost, increasing the probability of cancer detection by 15 per cent, the researchers determined.
That’s important, because a prior team from the University of Melbourne and Melbourne-area health facilities, recently reported that even in Australia, where mammograms are reviewed by two, and sometimes three, independent radiologists, there is still a 369 per 10,000 rate of false positives and an 18.6 per 10,000 rate of false negatives. Automated AI deep-learning processes, trained on large databases of actual cases, are widely viewed as ways to decrease the risk of such errors, reducing the need for unnecessary re-tests while also saving lives by reducing the rate of false negatives.
To find out how effective this can be, the new team, led by AI researcher Bryan Haslam, of DeepHeath, in Somerville, Massachusetts, examined 747,604 women who had mammograms at 10 different medical systems.
Even though they had to pay for the additional use of AI themselves (because US insurance companies are not yet prepared to assist), nearly a quarter of these women opted in, with the percentage increasing over the course of the one-year study. The amount they were charged was not disclosed, but the Susan G. Koman Foundation, a US breast-cancer advocacy group, estimates it as $US40 to $US100.
Overall, Haslam’s team found, the cancer detection rate in those who chose to use the AI process was 43 percent higher than in those who did not choose to use it.
Partly, they say, that was due to selection bias by patients who knew from family history (or other factors) that they were at higher risk, and were therefore more willing to spend the extra money.
But even controlling for that, Haslam’s team found that supplementing the normal mammography process with an AI screen increased the cancer detection rate by 21 per cent. Helping to improve the accuracy was an additional review by a second radiologist, in the event of disagreement between the initial radiologist and the AI.
And while not all of those women actually had cancer—there were still false positives—re-tests found that the AI-assisted process nevertheless caught 15 per cent more real cancers than the normal process.
Overall, Haslam says, “The AI-driven review program…[helps] ensure women with suspicious findings get expert-level care that could help detect many more breast cancers, early.” And, Haslam says, the number of women opting in for this program—even though they still have to pay for it themselves—“is now at 36 per cent and growing, and the rate of cancer detection continues to be substantially higher for those women.”
The next step, Haslam’s team says, is to conduct randomized trials that would eliminate any chance of self-selection bias and prove even more strongly that the AI-enhanced process is worth the money.
Kelly Scott, a primary care physician in Portland, Oregon, who was not part of the study team, is impressed with the results. “Radiology is one of the fields where AI is likely to be helpful at finding small or subtle abnormalities,” she says, “especially if backed up by radiologists’ interpretations. The big question will be the detection of true cancers versus false positives. The double-blinded study will be helpful at sorting that out.”
Meanwhile, she notes that the added cost isn’t a big deal for affluent people, but, at least under the US health-care system, “[It] raises the question of worsening disparities in breast cancer detection between wealthy and poor, if insurance doesn’t pay for it.”
Read more: AI and breast cancer
