A.I.'s promise for women's health

Reading a mammogram is a test in discerning overlapping shadows. Since cancer shows up on mammography images as white areas — just like glandular breast tissue — a radiologist's challenge becomes teasing out what's normal from what's not.

But radiologists have a secret weapon: artificial intelligence (A.I.). For years, mammograms have incorporated a technology called computer-aided detection that can flag spots with abnormal cells the human eye sometimes misses. The newest forms of A.I. continue to refine these abilities in ways that may be benefiting countless women.

"Every once in a while, the computer will circle something we were on the fence about or didn't notice, and it will change our approach," says Dr. Pierre Sasson, chair of the Department of Radiology at Harvard-affiliated Mount Auburn Hospital. "A.I. also reinforces questionable findings we see that make us more confident when we call a patient back for additional testing. That happens every day."

But while A.I. mammography software is heralded as one of the technology's most compelling applications in women's health, it's hardly the only one. And while A.I.'s uses — and the fury surrounding them — have exploded in recent years, earlier forms have quietly been part of American health care for decades.

"We're no longer at the dawn of artificial intelligence — I'd say we're at midmorning," Dr. Sasson says. "It's already present, already being utilized, and already helping."

Finding more breast cancers

One of A.I.'s earliest uses, in the 1970s and '80s, was in computer-aided detection systems designed to pin-point problems in fetuses on ultrasound images. Today, A.I. continues to sharpen the accuracy of women's imaging tests by helping detect and track not only breast cancer, but also endometriosis, fibroids, and cervical precancers, among other conditions.

Indeed, of the more than 700 FDA-approved A.I. and machine learning–enabled devices in health care, the vast majority are used in radiology, says Dr. Bernardo Bizzo, senior director of Harvard-affiliated Mass General Brigham AI, a diversified business that bridges academic research and A.I. product development.

"Health care uses of A.I. are still fragmented, but finally it's being integrated into our everyday practice, particularly in radiology," Dr. Sasson says. "I view it as affecting almost every part of a patient's care."

A.I.'s use in mammography has led to headline-worthy results. An August 2023 study published in The Lancet Oncology indicated that A.I.-supported mammogram readings revealed 20% more cancers than a routine double reading by two different radiologists. The technology also accomplished this without increasing the rate of false positives found on mammogram images — areas that look like cancer but turn out to be normal after additional (and often anxiety-provoking) testing.

A.I. can also alert mammography technicians whether the breast images they're capturing are high-quality, Dr. Sasson says. "That's before images even get to the radiologist," he says. "A.I. prioritizes patients to help us determine if something needs attention sooner."

Noteworthy limitations

Some experts had once predicted that A.I. would replace radiologists, but that hasn't happened — and isn't likely to, Dr. Sasson says.

"There was a lot of angst in the early years that A.I. would take over specialties like radiology and pathology, but we're now welcoming our 'computer overlords' because they're helping us do better work," Dr. Sasson says. "The saying in our world is, 'A.I. with a radiologist is better than A.I. without a radiologist, or a radiologist alone.' So it's a good combination."

One of A.I.'s major limitations, however, is precision. It isn't always able to differentiate normal from abnormal spots on various imaging scans, Dr. Sasson says. For example, his team tested software that counts and measures lung nodules on CT scans and it was a "disaster," flagging electrodes on patients' chests as suspicious lesions.

A.I. can't yet work independently, notes Dr. Bizzo, who is also an assistant professor of radiology at Harvard Medical School. "We still need radiologists to supervise and review A.I. results, and the diagnostic decision is not final without the radiologist's approval," he says. "There are no autonomous uses of A.I. approved for radiology in the United States, and I don't see the FDA accepting full automation in the near term."

"It's important to highlight that ultimately physicians must review all A.I. results related to diagnosis, treatment, or prognosis, and their expertise prevails," Dr. Bizzo adds.

What's next?

In coming years, A.I. may be able to produce personalized breast cancer risk assessment scores, which offer a more detailed picture of an individual woman's risk for the disease.

Currently, women's breast cancer risks are calculated using questionnaires that ask about factors such age, ethnicity, family history of breast cancer, breast biopsy history, breast density, when they first had a menstrual period, and — for those who have children — how old they were when they first gave birth. All of these issues can influence breast cancer risk.

The calculators used now compare each person's answers with the average for others of the same age, ancestry, and background. This comparison generates her estimated risk of developing breast cancer in both the next five years and her lifetime. But A.I. could greatly streamline this process.

"This is an exciting area of growth for A.I.," Dr. Sasson says. "We want to make sure we're using the right breast screening tests on the right patients, and A.I. will help that to happen."

Image: © mphillips007/Getty Images