MGH, MIT dive deeper to better determine breast cancer risk through deep learning

Portfolio

MGH, MIT dive deeper to better determine breast cancer risk through deep learning

September 2, 2022

By David Godkin

Researchers at Massachusetts General Hospital (MGH) and the Massachusetts Institute of Technology (MIT) have developed a deep learning computer model that expands on the ability to identify discrete data contained in breast imaging to better predict a woman’s chances of developing
breast cancer. Derived from more than 245,000 mammograms from over 80,000 patients over seven years, MGH’s AI-powered algorithm provides a fuller understanding of breast density, tissue patterns and other factors within the complex weave of breast imaging.

“It’s been challenging to help women who are so worried they’re going to develop breast cancer. But knowledge is power,” Connie Lehman, MGH’s senior physician scientist and professor of radiology at Harvard Medical School told BioWorld. “With this research we’re entering a new era where we can much more rapidly evaluate methods for reducing and maybe even removing a woman’s risk of breast cancer.”

Going deeper

Back in the ‘50s, radiologists and breast cancer specialists had an ‘aha’ moment: X-rays of the breast, they discovered, could detect cancers before they were felt with the fingers. Suddenly, doctors had found a way to help women survive breast cancer. “But not only survival,” said Lehman,“but better options for treating breast cancer.” “Disfiguration and the incredible morbidity associated with treating advanced breast cancer in the ‘50s and ‘60s really started to recede in the ‘70s and ‘80s.” Successful embryonic X-ray technology drove medtechs to design other diagnostic technologies and improve X-ray’s capacity to diagnose breast cancer tissue for removal and to preserve the breast. Scroll forward three decades and signs of calcification and millimeter-size masses in breast tissue have become part of the foundation of early breast cancer diagnosis. Ditto digital processing of mammograms images to penetrate very dense breast tissue and more recently 3D tomosynthesis. MGH believes it has taken another major step forward by “leveraging the power of deep learning to extract and process information the human eye and brain cannot detect,” said Lehman. The object is to look beyond existing cancers to cancers that might emerge in a much higher risk patient over five years. “Read the mammogram and see if anything suspicious should be worked up now, but we can also start separating out those women at high risk for future cancer from those women at very low future risk.” Traditional on-line risk models enable a woman to enter her age, family history, biopsy record etc., to obtain a risk score relative to those generated from other women’s medical profiles. This is used to guide the way forward to enhanced screening and risk reduction strategies, said Lehman. When Lehman used MGH’s database to examine the performance of traditional risk models, she learned something disturbing “I was really disheartened to see the racial disparities.” She and her colleague breast radiologist Leslie Lamb found that white women were two and a half to three times more likely to be diagnosed for increased risk of breast cancer than Hispanic, African American and Asian-American women.

Racial equity, better science

Historically it has been assumed African American women are more likely to die of breast cancer than Caucasian women because they have less access to or seek less frequently mammogram screening. Then investigators discovered African American women actually are more likely to have a different type of breast cancer and are typically diagnosed at a younger age. “What we found was the risk models that are guiding women at higher risk to better services effectively exclude women of color because the models themselves are biased,” Lehman noted. Citing a recent New England Journal of Medicine article “Hidden in Plain Sight” (https://www.nejm.org/doi/full/10.1056/NEJMms2004740) Lehman said structural racism is endemic to diagnostic risk assessment models. “My `aha’ moment was when I saw it in the very patient at Mass General that I had the responsibility for taking care of,” said Lehman. The other was when she and her MIT colleagues found a deep learning tool “that has no racial bias that performed equally well whether the woman was African American, Asian, Hispanic or Caucasian.” “Our idea was for a deep learning model trained to say `This is an image from a woman who developed breast cancer, this is an image from a woman who didn’t develop breast cancer.’”
Studies have shown that genetic predisposition, higher levels of estrogen, different tissue patterns and past biopsies are all manifest in breast cancer images – some in faint signals “that say pay attention,” Lehman said. What the deep learning tool she and her group have developed “amplifies those signals.” Amplified signals from tissue, higher BMI and environmental exposure emanate from large amounts of data from more than 245,000 mammograms amassed by MGH “and known, labelled outcomes of Yes/No breast cancer,” said Lehman.

Getting it out of the lab

The opportunity to test this came with the reopening of screening facilities after the precipitous drop in screening during the pandemic and a request from Massachusetts’ governor that patients at increased medical risk be screened first. Lehman’s concern: use of traditional biased risk models would place white patients at the front of that line. “I wanted to use our AI models to bring some balance and some equity to those we invited back.” In a study (https://pubmed.ncbi.nlm.nih.gov/35876790/) published July 25, 2022 by the Journal of the National Cancer Institute, Lehman and her co-authors compared the accuracy of a deep learn- MGH, MIT dive deeper to better determine breast cancer risk through deep learning risk score derived from the patient’s prior mammogram to traditional risk scores to prospectively identify patients with cancer in a cohort due for screening. The report found that the deep learning model detected more cancers per thousand patients at increased risk compared to the traditional Tyrer-Cuzick and NCI’s BCRAT risk models. The study’s conclusion: risk-based screening is feasible and effective when supported by a deep learning model for assessing breast cancer risk. Added the report, “deep learning is superior to traditional models to identify patients destined to develop cancer in large screening cohorts.” This shows it is possible in real time to get the science right, said Lehman, and achieve greater equity when assessing who is at higher risk of acquiring or re-acquiring breast cancer. “At the same time to achieve global impact, we want to have a commercial product,” Lehman stressed. “We can’t just keep this within the research domain.” The person designated to help make that happen is Joe Cunningham, a partner at Santé Ventures Capital Group that has raised $9 million in series A funding to support Clarity Inc., co-founded with Lehman and her group to carry its breast cancer screening approach into the commercial realm. “Joe said I really want to support this work going forward,” said Lehman. “Our goal is to bring access to this technology with a really high level of quality and do it globally.”