A novel technique for analyzing mammograms has demonstrated a higher accuracy in identifying individuals at elevated risk of breast cancer compared to the traditional questionnaire-based approach. This groundbreaking method, utilizing artificial intelligence, holds the potential to facilitate earlier cancer diagnoses and inform decisions regarding earlier screenings, additional imaging, or preventive medications.
A recent study conducted by Washington University School of Medicine in St. Louis introduces an advanced way to analyze mammograms, significantly enhancing the ability to forecast the risk of breast cancer over the next five years. By using up to three years of prior mammograms, this innovative approach accurately identified high-risk individuals 2.3 times more effectively than the conventional method, which relies solely on questionnaires that evaluate clinical risk factors like age, race, and family history of breast cancer.
The findings were released on December 5 in JCO Clinical Cancer Informatics.
“Our goal is to enhance early detection as it greatly improves treatment outcomes,” stated Graham A. Colditz, MD, DrPH, the study’s senior author, who serves as the associate director of the Siteman Cancer Center affiliated with Barnes-Jewish Hospital and WashU Medicine, in addition to being the Niess-Gain Professor of Surgery. “Improving risk predictions may also benefit prevention research, enabling us to identify better strategies for women at high risk to lower their five-year breast cancer risk.”
This risk assessment approach builds on previous research by Colditz and the primary author, Shu (Joy) Jiang, PhD, a statistician and associate professor of surgery in the Division of Public Health Sciences at WashU Medicine. They discovered that prior mammograms contain critical early indicators of breast cancer development, often undetectable even by highly trained professionals. This involves subtle fluctuations in breast density, which signifies the ratio of fibrous to fatty tissue in the breasts.
For this new study, the researchers developed an AI-based algorithm that can identify these subtle nuances in mammograms, aiming to pinpoint women at the highest risk of developing a new breast tumor within a specific period. Besides breast density, their Machine-Learning tool analyzes various image patterns, including texture, calcification, and any asymmetrical features in the breasts.
“Our advanced method can detect minute changes over time in repeated mammogram images that are otherwise invisible,” noted Jiang, emphasizing that these variations provide crucial insights for identifying high-risk patients.
Currently, options for risk reduction are limited, typically involving medications like tamoxifen that decrease risk but may present side effects. Women identified as high-risk generally receive more frequent screenings or have the chance to incorporate additional imaging techniques, such as MRI, to detect cancer as early as possible.
“Right now, we lack a reliable means to predict who is likely to develop breast cancer based solely on their mammogram images,” remarked Debbie L. Bennett, MD, a co-author and associate professor of radiology, who leads breast imaging at the Mallinckrodt Institute of Radiology at WashU Medicine. “This research is thrilling because it suggests we can derive valuable information from both current and past mammograms utilizing this algorithm. While predictions will never be perfect, this study implies that the new algorithm significantly outperforms existing methods.”
AI Advances Predictions for Breast Cancer Risks
The research team trained their machine-learning algorithm using mammograms from over 10,000 women who underwent breast cancer screenings at Siteman Cancer Center from 2008 to 2012. These women were monitored until 2020, during which time 478 of them were diagnosed with breast cancer.
The researchers subsequently applied their predictive method on a different group of over 18,000 women receiving mammograms at Emory University in the Atlanta area from 2013 to 2020. In this cohort, 332 women were diagnosed with breast cancer during the follow-up period that concluded in 2020.
The novel prediction model revealed that women in the high-risk category were 21 times more likely to develop breast cancer over the next five years compared to those in the low-risk group. In the high-risk category, 53 out of every 1,000 screened women were diagnosed with breast cancer in the following five years, whereas only 2.6 women per 1,000 in the low-risk group were diagnosed within the same timeframe. Under previous questionnaire-based methods, only 23 women per 1,000 were accurately classified in the high-risk category, which meant the older method overlooked 30 breast cancer cases that this new approach successfully identified.
The mammograms were conducted at both academic medical centers and community clinics, confirming the method’s reliability across diverse settings. Notably, the algorithm was developed with a strong representation of Black women, a group often underrepresented in breast cancer risk model research. The accuracy of the predictions remained consistent across racial demographics. Of those screened at Siteman, the majority were white, with 27% being Black. Of those screened at Emory, 42% were Black.
The research team continues to evaluate the algorithm across women of varied racial and ethnic backgrounds, including Asian, Southeast Asian, and Native American populations, to ensure equal accuracy for all groups.
The researchers are collaborating with WashU’s Office of Technology Management to secure patents and licenses for this innovative method with the intention of making it widely accessible wherever screening mammograms are performed. Colditz and Jiang are also exploring the establishment of a startup company centered around this technology.
Jiang S, Bennett DL, Rosner BA, Tamimi RM, Colditz GA. Development and validation of a dynamic 5-year breast cancer risk model using repeated mammograms. JCO Clinical Cancer Informatics. Dec. 5, 2024.
This research was supported by Washington University School of Medicine in St. Louis.
Jiang and Colditz currently hold pending patents related to this work on predicting disease risk through radiomic imaging.
