AI Model Predicts 5-Year Breast Cancer Risk from Mammograms

Approximately 13% of U.S. women, or one in every eight, are predicted to develop invasive breast cancer over their lifetime, with 1 in 39 women (3%) succumbing to the illness, according to the American Cancer Society. Mammography screening remains a vital tool for early breast cancer detection, offering the most effective treatment window. Regular mammogram appointments can significantly reduce breast cancer mortality risks. Nonetheless, the challenge remains in accurately predicting which individuals will contract breast cancer solely through screening methods. Mirai, an advanced deep learning algorithm, has been recognized for its ability to predict breast cancer risk, although its decision-making process remains largely unexplained, creating risks of overreliance and misdiagnoses by radiologists. Now, researchers have developed an innovative, interpretable artificial intelligence (AI) model capable of predicting the five-year risk of breast cancer based on the analysis of mammograms.

In the study, researchers at Duke University (Durham, NC, USA) conducted a comparative study utilizing their newly devised deep learning model, dubbed AsymMirai, against Mirai's one to five-year breast cancer risk assessments. AsymMirai inherits its deep learning "front end" from Mirai but incorporates an interpretable module called local bilateral dissimilarity, focusing on the tissue contrast between the left and right breasts. This study analyzed 210,067 mammograms from 81,824 patients from the EMory BrEast imaging Dataset (EMBED) spanning from January 2013 to December 2020, employing both the Mirai and AsymMirai algorithms.

The findings revealed that the simplified deep learning model, AsymMirai, nearly matched the performance of the state-of-the-art Mirai algorithm in predicting breast cancer risk from one to five years. Moreover, this study highlighted the significance of bilateral asymmetry as a vital clinical indicator, suggesting its potential as a novel imaging marker for assessing breast cancer risk. The transparency behind AsymMirai's decision-making process makes it an invaluable tool for radiologists, enhancing the accuracy of breast cancer diagnosis and risk prediction.

"We can, with surprisingly high accuracy, 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," said the study's lead author, Jon Donnelly, B.S., a Ph.D. student in the Department of Computer Science at Duke University. "This could have public impact because it could, in the not-too-distant future, affect how often women receive mammograms."

Related Links:
Duke University