Breast cancer is the most frequently diagnosed cancer among women in 140 of 184 countries worldwide, according to World Cancer Research Fund International. In the United States, it’s the second most common among women after skin cancer. Despite great advances in medical diagnostic technologies, some patients receive a diagnosis too late.
A team from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) and Massachusetts General Hospital (MGH) is using artificial intelligence to create a deep learning model to predict breast cancer. Using a mammogram, the project’s technology can predict if a patient has a risk of developing breast cancer up to five years before receiving a diagnosis.
According to Constance Lehman, professor of radiology at Harvard Medical School and division chief of breast imaging at MGH, this project is the first to design “accurate risk assessment tools that [work] for individual women.”
Current screening tools often use factors such as genetics, hormones, diet, pregnancy and weight gain to determine cancer risk. However, most of these factors are only mildly correlated with an actual diagnosis. The MIT/MGH model uses deep learning to go directly to the source and look for patterns in the data. The team used 90,000 mammograms from over 60,000 MGH patients, resulting in a model that can detect subtle changes in breast tissue that can lead to malignant tumours.
The model has been shown to predict breast cancer risk at a significantly higher rate than current methods. It accurately placed 31 percent of cancer diagnoses in its highest-risk category, compared to only 18 percent for traditional models.
MIT professor Regina Barzilay was prompted to conduct oncology research after receiving a breast cancer diagnosis during a routine mammogram appointment in 2014. The MacArthur-grant-winning researcher told WBUR that she was not only shocked as the first member of her family to receive a cancer diagnosis, but she was also surprised at the lack of advance technologies used in cancer treatment and diagnosis.
“What was even more unexpected was how little data, and machine learning, are used in oncology,” she recounted. “I hadn’t seen any, and I was treated at an excellent place, at MGH. And I don’t think that MGH is an exception, it’s actually true from what I understood across the country. Today we are not using at all, data to select treatment, to personalize it, or to help the patient’s reduce their uncertainty about the outcomes. I really strongly felt it has to be changed.”
Instead of taking a one-size-fits-all approach to cancer screening, Regina stated that mammograms can be used in accordance with a woman’s risk of developing cancer. “For example, a doctor might recommend that one group of women get a mammogram every other year, while another higher-risk group might get supplemental MRI screening,” she explained in a blog post on the university’s web site.
The project also aims to make accurate diagnoses more racially inclusive. Typical screening tools in the US were based off white women and are less effective for other ethnic groups. The MIT/MGH model is equally accurate for white and black women, which may lessen the gap between the cancer survival rates for both groups.
“It’s particularly striking that the model performs equally as well for white and black people, which has not been the case with prior tools,” said Allison Kurian, an associate professor at Stanford University School of Medicine. “If validated and made available for widespread use, this could really improve on our current strategies to estimate risk.”
The team believes that the model has the potential to eventually predict the likelihood of developing other health conditions. Eventually, the researchers hope to apply these technologies to diseases with less effective risk models, such as pancreatic cancer.