Using A.I. to Transform Breast Cancer Care

This post was originally published on this site

How could a researcher in computer science improve future cancer care, I wondered, when a trip to Boston afforded me the opportunity to converse with Regina Barzilay, a professor at the Massachusetts Institute of Technology and the recipient in 2017 of a prestigious MacArthur Fellowship, known as a “genius grant.” After a breast cancer diagnosis in 2014, Dr. Barzilay, who has a doctorate in computer science, began directing her work in artificial intelligence toward helping other patients.

She and her team have developed algorithms to predict whether a patient is likely to develop breast cancer in the next five years. Their model is designed to spot the tiny changes on mammograms that turn into tumors. And it detects them regardless of the patient’s race, a significant concern in light of the racial divide in breast cancer mortality.

During a conversation in an office notable for a vertical tower of books, Dr. Barzilay used an analogy to help explain her efforts to apply machine learning to the disease. “Consider the algorithms used by market researchers for Amazon,” she said. “They are based on your clicking on this or that product.”

Just as machines can be programmed to track a pattern of clicking so as to predict my eccentric tastes in merchandise, computers can be employed to map the medical history of many patients so as to predict and better treat new cancers — and maybe to prevent them.

The enthusiasm that Dr, Barzilay brings to this undertaking is fueled by her dismay at current approaches to cancer care. While being treated at Massachusetts General Hospital, she was struck by the high degree of uncertainty surrounding treatment of her disease. Why did her questions go unanswered about how other patients at the same hospital with similar tumors fared with this or that drug or with this or that surgery? Why was there so little information?

According to Dr. Barzilay, oncologists ground their treatment regimens in clinical trials; however, these trials enroll only about 3 percent of the eligible population. The experiences of as many as 97 percent of cancer patients are therefore disregarded. To Dr. Barzilay, such a “primitive” practice seems “a travesty,” especially because large volumes of information about patients accumulate in every hospital. One problem is that hospital data are written in what is called “free-text”— English rather than a structured format (like a database form) that computers can process — which limits their utilization.

To obtain targeted evidence for patients, Dr. Barzilay created a detailed database of pathology reports from three decades on more than 100,000 patients with breast cancer at Massachusetts General and developed an algorithm to parse them. New patients will be empowered by learning how tumors with particular characteristics responded to specific treatments. Machines accessing subsets of the population will also make it faster and cheaper for clinicians to identify patients with particular disease characteristics and to enroll them in clinical trials.

As always with A.I., ethical concerns arise. Just as “weaponized” surveillance cameras could be abused, cancer data could be misused by insurance companies and employers. Yet Dr. Barzilay believes the payoffs are worth the risks: “Oncologists should be reaching out to researchers in A.I.” Finding no evidence that physicians are doing so, she has embarked on a second project reaching out to them.

Like A.I. specialists at Google, who use computers to interpret CT scans of lung cancer, Dr. Barzilay teaches computers how to analyze medical images of breast cancer. She does so because even high-resolution pictures can produce indeterminate results or false positives or they can be misread by radiologists. Opening her laptop, Dr. Barzilay showed me her own mammograms from 2012, 2013 and 2014. Although her cancer was not diagnosed until 2014, there is evidence of it in the 2012 and 2013 images.

What if her cancer had been caught earlier? That question quickly converted a personal misfortune into a professional mandate.

With her collaborators — Dr. Constance Lehman, the chief of breast imaging at Mass General; Dr. Kevin S. Hughes, also at Mass General; and Adam Yala, her graduate student at M.I.T. — Dr. Barzilay has taught computers to generate detailed information from mammogram images using data from over 60,000 patients.

“Machines work more effectively than human eyes,” Dr. Barzilay explained. “They can register subtle changes in tissue — influenced by genetics, hormones, lactation, weight changes — that we cannot see.”

In one study, patients identified as high risk by their model were 3.8 times more likely than the average patient to develop breast cancer. Their model was significantly more accurate than the current clinical standard and more accurate than approaches based on breast density. It identified 31 percent of patients with future breast cancer as high risk, while the current clinical standard identified only 18 percent Dr. Barzilay and her collaborators want to usher in the day when no woman is surprised by a late-stage diagnosis and when all breast cancers are curable. They also hope to solve the problems of over- and under-testing. Instead of a one-size-fits-all practice, the frequency of screenings and biopsies could be customized with sufficient data.

At that moment, I thought of the young women I know who have inherited a BRCA mutation, and specifically of their distress over whether or when to undergo prophylactic double mastectomies. Needless to say, they dread such an operation since they have no certainty that it is truly necessary. “Exactly,” Dr. Barzilay agreed. “With a CD of their scan, we would be able to tell them their personal risk.”

While thanking Dr. Barzilay for taking the time to meet with me, I pointed to the vertical tower of books piled high on top of each other and asked, “How do you get a volume out from the middle of the stack, like that bright yellow book midway between the top and bottom?” Laughing, she stepped up to the massive pile and pulled the yellow book out … without a catastrophic collapse. There must be hidden shelves between every four or five volumes, I realized.

In a flash, I recognized what my human eyes could not see. I wished Dr. Barzilay the best of luck with research-in-progress that has already proven how successfully she has transformed the trauma of diagnosis into a quest to democratize cancer care, to extract from the amassed experiences of many information that aids a single individual: e pluribus unum.

Susan Gubar, who has been dealing with ovarian cancer since 2008, is distinguished emerita professor of English at Indiana University. Her latest book is “Late-Life Love.”