Every two minutes, a woman is diagnosed with breast cancer. The question is, can those at high risk be identified early enough to survive?
The current standard practice in medicine is not exactly precise. It relies on age, family history of cancer, and breast density, among other factors, to determine risk. But these factors do not always tell the whole story, leaving many women to slip through the cracks. In addition, a racial gap persists in breast cancer treatment and survival. African-American women are 42 percent more likely to die from the disease despite relatively equal rates of diagnosis.
But now those grim statistics could be changing. A team of researchers from MIT’s Computer Science and Artificial Intelligence Laboratory have developed a deep learning model that can more accurately predict a patient’s breast cancer risk compared to established clinical guidelines – and it has predicted risk equally well in both white and black women for the first time.
Study results published in Radiology described how the AI software read mammogram images from more than 60,000 patients at Massachusetts General Hospital to identify subtle differences in breast tissue that pointed to potential risk factors, even in their earliest stages. The team accessed the patients’ actual diagnoses and determined that the AI model was able to correctly place 31 percent of all cancer patients in the highest-risk category of developing breast cancer within five years of the examination, compared to just 18 percent for existing models.
“Each image has hundreds of thousands of pixels identifying something that may not necessarily be detected by the human eye,” said MIT professor Regina Barzilay, one of the study’s lead authors. “We all have limited visual capacities so it seems some machines trained on hundreds of thousands of images with a known outcome can capture correlations the human eye might not notice.”
Barzilay, a breast cancer survivor herself, had abnormal tissue patterns on mammograms in 2012 and 2013, but wasn’t diagnosed until after a 2014 image reading, illustrating the limitations of human processing alone.
Next up: The MIT team is looking at training the model to detect other cancers and health risks. Barzilay recalls how a cardiologist told her during a conference that women with heart diseases had a different pattern of calcification on their mammograms, demonstrating how already existing images can be used to extract other pieces of information about a person’s health status.
Integration of the AI model in standard care could help doctors better tailor screening and prevention programs based on actual instead of perceived risk. Patients who might register as higher risk by current guidelines could be identified as lower risk, helping resolve conflicting opinions about how early and how often women should receive mammograms.
Open Questions: While the results were promising, it’s unknown how well the model will work on a larger scale, as the study looked at data from just one institution and used mammograms supplied by just one hospital. Some risk factor information was also unavailable for certain patients during the study, leaving researchers unable to fully compare the AI model’s performance to that of the traditional standard.
One incentive to wider implementation and study, however, is the bonus that no new hardware is required to use the AI model. With other institutions now showing interest, this software could lead to earlier routine detection and treatment of breast cancer — resulting in more lives saved.