Researchers have developed a new, interpretable model for predicting five-year breast cancer risk based on mammograms, according to a recent study published in the journal Radiology.
One in eight women, approximately 13% of the female population in the U.S., develops invasive breast cancer during their lifetime, and one in 39 women (3%) dies from the disease, according to the American Cancer Society. Regular mammogram screenings can significantly reduce the risk of death from breast cancer. However, accurately predicting which women will develop breast cancer based solely on screening remains a challenge.
Mirai, a state-of-the-art deep learning algorithm, has shown promise as a tool for breast cancer prediction. However, due to the limited understanding of its decision-making process, there is a risk that radiologists might over-rely on it, potentially leading to misdiagnoses.
“Mirai is a black box—a very large and complex neural network, similar in construction to ChatGPT—and no one knows exactly how it makes its decisions,” said Jon Donnelly, BS, lead author of the study and a PhD candidate in the Department of Computer Science at Duke University in Durham, NC. “We developed an interpretable artificial intelligence method that allows us to predict breast cancer from mammograms 1 to 5 years in advance. AsymMirai is much simpler and easier to understand than Mirai.”
For this study, Donnelly and his colleagues compared their newly developed deep learning model, called AsymMirai, with Mirai’s predictions of breast cancer risk over 1 to 5 years. AsymMirai is built on the “front end” of Mirai’s deep learning framework, but the rest of the complex method has been replaced with an interpretable module: local bilateral dissimilarity.
“Previously, differences between left and right breast tissue were used only to assist in detecting cancer, not in predicting it,” Donnelly said. “We found that Mirai relies on comparisons between the left and right sides, which allowed us to design a much simpler network that also makes side-by-side comparisons.”
FOLLOW US ON FACEBOOK:
Findings May Influence Mammogram Frequency
For the study, researchers compared 210,067 mammograms from 81,824 patients using the Emory Breast Imaging Dataset (EMBED) between January 2013 and December 2020, testing both Mirai and AsymMirai models. They discovered that their simplified deep learning model performs almost as well as Mirai in predicting breast cancer risk over 1 to 5 years.
The results also highlighted the clinical significance of breast asymmetry, emphasizing the potential of bilateral dissimilarity as a future marker for breast cancer risk assessment.
Because AsymMirai’s reasoning is easy to interpret, it could be a valuable addition to human radiologists in diagnosing breast cancer and predicting risk, Donnelly noted.
“With surprising accuracy, we can predict whether a woman will develop cancer in the next 1 to 5 years based solely on localized differences between her left and right breast tissue,” he said. “This could have public health implications, as it might influence how often women undergo mammograms in the future.”
READ MORE:
SERBIAN GENERAL HONORED IN CANADA: A Mountain Named After Our Hero of World War I! (PHOTO)
Source: RSNA, Photo: Pixabay