Google’s artificial intelligence technology, DeepMind, beat the world champion at the ancient Chinese game “Go” in 2016. It was a major AI victory, arriving nearly a decade earlier than most experts had predicted. Now, the same technology has a new goal: improving reading breast-cancer screenings, which could directly affect millions across the globe.
Nearly 1.7 million cases of breast cancer were diagnosed in 2012 worldwide (though global figures can be fuzzy since countries have different philosophies about when and how often to screen) and it is the fifth most common cause of death from cancer in women. Mammograms, a type of X-ray screening, have been widely promoted as a tool to catch breast cancer early on.
But according to Breast Cancer Action, a US-based activist organization, mammograms are far too inaccurate. About 20% of mammograms result in a false-negative, meaning the doctor or technician doesn’t find cancer that was, in fact, there. False-positive mammograms, when the doctor or technician reads the screening as having found cancer, but is proven wrong upon further testing, are also a problem. In the US, more than 50% of those women who get an annual mammogram for 10 years in a row will have a false-positive result during those 10 years. While the dangers of false negatives are obvious, false positives can also be perilous, leading to unnecessary tests (including biopsies) and even unneeded chemotherapy in some cases. All of that can create stress for patients and “add pressure and costs to health services around the world,” says Dominic King, the clinical lead at DeepMind’s health division
In November 2017, the Cancer Research UK Imperial Centre announced that a consortium of top AI researchers and health care professionals, led by King and DeepMind Health, would be put to the task of using Google’s AI tech to improve the reading and assessment of mammograms. The partnership aims to address some of the current issues and learn whether its machine-learning algorithms can be applied “to analyze [mammograms], to spot signs of cancerous tissue, and alert radiologists more accurately than current techniques allow,” according to a press release.
DeepMind develops machine-learning algorithms that take a data set and learn from it, eventually developing the ability to make predictions. When AlphaGo was trained to master Go, for instance, it studied millions of games that had been played by humans in the past in order to make predictions about possible moves in future games. It then learned through trial and error, seeing which predictive moves were successful and which weren’t. DeepMind doesn’t just play games; it has been applied to solve a range of problems, from reducing the power consumed by Google’s energy centers to some of the biggest public health challenges of the 21st century
A number of tech companies are now using AI to try to improve the way breast cancer screenings are read. The French startup Therapixel, part of Nvidia’s Inception accelerator program, says its tech can cut down false-positive rates by 5% compared with current state-of-the-art diagnostic tools. An AI program developed at the Houston Methodist Research Institute in Texas recently proved 99% accurate in identifying signs of breast cancer risk by reading mammograms—and could read those images 30 times faster than humans. And in 2017, an annual nine-month-long competition—the Digital Mammography Challenge—attracted over 1,200 participants, illustrating how many AI companies are interested in tackling the challenge of improving breast cancer screening.
Health-tech companies are applying similar AI-driven diagnostic approaches to other cancers as well: When the Chinese AI company BioMind analyzed brain cancer images in June 2018, it was correct 87% of the time, compared with medical professionals, who were correct 66% of the time—and cut the time of diagnosing in half.
While most of these technologies are still in the lab, some AI-based cancer tools are already being used to identify risk factors for breast cancer in clinical settings. For example, deep-learning algorithms currently in use in commercial software are now accurately able to look at mammograms and correctly identify increased breast density, a strong risk factor for breast cancer, according to a 2018 study involving over 6,300 women in the US.
DeepMind Health is currently training its AI on data from Optimam, a database of over 80,000 digital images collected via the UK’s National Breast Screening System—all stripped of any personal identifiers—funded by the charity organization Cancer Research UK. And while other technology companies are focusing on the issue of mammograms, DeepMind’s machine learning tech is arguably one of the most advanced to be working on this challenge, and, by virtue of being owned by Google, reportedly the “most valuable company in the world,” whatever software it produces will likely have a wide reach.
If the AI is successful, it could save significant time for doctors reading screenings, potentially reducing hundreds of hours of work to mere minutes. With more than 30 million mammograms performed each year in the US alone, this could have massive effects.
It should be noted that 3D mammography, which creates a 3D image of the breast using multiple X-rays, was approved in the US in 2011, and is still being explored as a potential improvement on standard mammography. So other other improved imaging technologies. But it’s not clear how much more effective, if at all, these methods are. While its current project focuses on mammography, a DeepMind Health representative says the AI its working on will have applications across all breast imaging methods. It’s also unclear if using AI to scan mammograms will keep costs down, or if it is best used in conjunction with improved imaging techniques.
DeepMind Health’s current research isn’t focused on two areas that experts say are critical in breast cancer treatment. One is identifying which cancers actually need treatment (though a DeepMind representative says the current research could be useful in this application down the road). The other is that current mammograms cannot catch interval cancers—the fast-developing and often-deadly cancers that emerge between screenings. However, DeepMind says “the algorithm is being tested in a way that also considers interval cancers as important,” so again, it could be useful for this application in the future.
Additionally, DeepMind Health’s mammogram project is based on the “early-detection” model, which assumes catching anything that appears to be a tumor early on is always in the best interest of patients. But this has, and continues to be, the subject of endless debate.
Author Peggy Orenstein, who has written extensively on some of the problems with the current mammography industry, is skeptical of the early-detection approach. “If the issue is that [the technology] will find tumors earlier, you have to ask: ‘What impact will it have on survival? What impact will it have on overdiagnosis?’” She points out that assuming women whose cancer is detected “early” automatically live longer has proved “patently untrue.” That claim is backed up by the research of H. Gilbert Welch, who, in a 2012 study examining three decades of mammogram screenings found “a larger relative reduction in mortality among women who were not exposed to screening mammography than among those who were exposed.” The study concluded that any trends in decreasing mortality rates in breast-cancer patients is mostly the result of improved treatment, and not additional screening.
Karuna Jaggar, executive director at Breast Cancer Action says the evidence shows that the philosophy around early detection using annual mammograms is deeply flawed, and not achieving its desired goals.
This doesn’t mean there’s no benefit from improving mammograms. Jaggar points out that in addition to screening asymptomatic women for early signs of disease, mammograms are also used more extensively for what professionals call “surveillance”: to watch women who have already had a breast cancer diagnosis or are at heightened risk for breast cancer, such as those who have a known BRCA genetic mutation. (Current protocol for those who are high-risk is to alternate mammograms and MRI every six months.) Mammograms are also essential in diagnosing suspicious symptoms that women and their health care providers have found.
But Jaggar says that decades of research have shown breast screening to be “deeply flawed,” often leading to overdiagnosis, overtreatment, and failure to prevent death—and that “technology alone cannot overcome these fundamental issues.”
In 2016, a panel of medical experts in the US revised the recommendations for mammograms, citing that the screenings are safe for women 50 years and older, and should be done every other year after that. But many doctors and insurance providers still urge women to begin screenings at age 40, and to have them annually.
DeepMind is not certain when the results of the collaboration will be available, says King, but whenever that point comes, it will become more clear what the DeepMind AI is learning from the de-identified mammograms. If the result of the current research is to encourage more mammograms, especially when they are shown not to benefit younger women, it is worth assessing who among those affected—doctor, patient, hospital, and tech company—stands to benefit most from the new technology.
Correction, Sept. 10, 2018: This article has been updated to clarify that while DeepMind Health’s current research is not focused on identifying which cancers actually need treatment or to identify interval cancers, the current work could be useful when tackling these applications in the future. In addition, a previous article incorrectly said the results of the project will be available in November 2018. In fact, DeepMind is not sure when they will be available. This article has also been corrected to clarify Karuna Jaggar’s views on mammograms.
This story is one in a series of articles on the impact of artificial intelligence on health care and medicine. Click here to sign up to get alerted when new stories are published.