Google’s AI algorithm for helping to screen for breast cancer will now be part of commercial mammograms.
On November 28, the company announced it licensed its AI technology to iCAD, a medical technology company that provides breast cancer detection services to health care facilities around the world.
While iCAD already includes AI-based strategies in its cancer screening services, it will now also incorporate Google’s algorithm, which Google has been testing with researchers at Northwestern University. “It’s an inflection point for us,” says Greg Corrado, co-founder of the Google Brain team and principal scientist on Google’s AI health care team. “We’re moving from academic research to being able to deploy our algorithm in the real world.”
In an earlier study published in 2020 in Nature, Google’s algorithm for mammograms performed better than radiologists in logging fewer false positives and false negatives in reading the images. The study involved mammograms from more than 91,000 women in the U.S. and the U.K. In the U.S., where most women ages 50 to 74 are recommended to be screened every two years, Google’s system lowered the false positive rate by 6%, and in the U.K., where women ages 50 to 70 are advised to get screened every three years, by 1.2%. The machine learning algorithm also decreased false positives by 9% in the U.S. and nearly 3% in the U.K.
That benefit will now be available commercially for the first time to the 7,500 mammography sites globally, including university health systems, that use iCAD’s services. While Corrado declined to detail how Google’s algorithm differs from those being tested by other researchers and companies in the field, he said the system incorporates data from a wide range of images, even beyond those of breast tissue, to refine the machine learning process. iCAD and Google will continue to develop and refine the technology as part of the partnership agreement.
The algorithm is not designed to replace radiologists, at least not in the near term. But in Europe, says Stacey Stevens, president and CEO of iCAD, it could help to relieve the burden on radiologists, since many nations (including the U.K.) require two readings of a mammography image. iCAD is working with health regulators to earn the proper authorization so that the company’s AI-based interpretation could eventually be one of them, she says. In the U.S., Stevens expects the first product including Google’s algorithm to be rolled out in early 2024.
Stevens also anticipates that the AI-based system will bring mammography to more people around the world, particularly in lower-resource areas that could not support the infrastructure required of hosting hardware related to mammography image storage. With Google’s cloud-based storage capabilities, she says, “we have the ability to expand to new geographies and new regions of the world and to scale our tools across a greater number of patients in areas of the world constrained by infrastructure challenges.”
As with any machine learning system, the more data from mammograms that are fed into the algorithm, the better it gets at detecting the smallest differences that distinguish normal tissue from potentially cancerous tissue. Women receiving mammograms using the AI-based system will have their information fed back into the algorithm, minus any identifying data. At the moment, most people getting mammograms likely aren’t aware that an AI-system might be in the background complementing the radiologist, since for now, no regulatory agencies have signed off on an entirely AI-based interpretation of mammograms. But as more AI algorithms like Google’s enter the market, that may change, and radiologists may end up discussing with patients how their images are interpreted.
Ultimately, such machine-based readings could begin to pull out patterns that human eyes can’t see. Stevens says iCAD’s current AI-based algorithm already detects the presence of minute calcifications in the breast tissue that scientists are beginning to link to a heightened risk of heart disease. If that association is confirmed, mammograms could also become a tool for assessing women’s heart disease risk.
For now, adding an AI perspective to mammograms could begin to improve how women’s risk of breast cancer is determined. AI systems can better distinguish, for example, differences that are unique to specific racial and ethnic groups; in the U.S., African-American women are at higher risk of developing more aggressive types of breast cancer and are more likely to die of the disease than other women, so training an AI system to track down the first signs of these cancers could lead to better outcomes. “We are finding that there are many cases of women with what seems like a normal mammogram, but there are things in those images that can’t be seen with the human eye,” says Stevens. If those differences can be picked up by an AI algorithm, then those women could be sent for additional screening to figure out whether they are at higher risk of developing cancer. That could set them on a path to receiving treatment sooner, which ultimately leads to a better chance of survival. That could also mean less expensive medical services, which would translate back to cost savings for the health system. “We are in the early innings of breast cancer risk assessment with AI,” says Stevens, “but we are excited about its potential.”