A physician looks at a mammogram on a computer screen.

Using Neural Networks to Improve Breast Cancer Prediction

December 10, 2021

Researchers from the Herbert Irving Comprehensive Cancer Center presented new research on breast cancer risk prediction at the San Antonio Breast Cancer Symposium, which took place December 7–10. The poster, titled “Improving breast cancer risk prediction using a convolutional neural network-based mammographic evaluation in combination with clinical risk factors,” describes a hybrid model of determining breast cancer risk using both clinical factors from the Breast Cancer Surveillance Consortium (BCSC) model and a novel, fully automated convolutional neural network (CNN)-based breast cancer risk model. The method, which is used with digital mammograms, has an overall accuracy of 72 percent in predicting subsequent invasive breast cancer.

Presenting the poster were Alissa Michel, MD; Vicky Ro, MD; Julia E. McGuinness, MD; Simukayi Mutasa, MD (University of California, Irvine); Richard Ha, MD, MS; and Katherine D. Crew, MD, MS. Drs. Crew and Ha are 2018 Velocity Fellows; they received grants from Velocity, Columbia’s Ride to End Cancer, to fund a clinical study to evaluate the use of CNN to assess response to hormonal therapy among women with early-stage breast cancer.

The researchers conducted a retrospective cohort study of 23,552 women, ages 35–74, who underwent screening mammography from 2014 to 2018 at Columbia University Irving Medical Center. Twenty-seven percent of the women were non-Hispanic white, 9 percent non-Hispanic Black, 36 percent Hispanic, 5 percent Asian, and 23 percent other/unknown race/ethnicity. From this cohort, they identified 206 women who subsequently developed invasive breast cancer by linkage to the tumor registry.

The clinical risk factors the researchers used to calculate the women’s five-year risk of invasive breast cancer were age, race/ethnicity, first-degree family history of breast cancer, prior benign breast biopsy, and breast density (from radiology reports). The researchers calculated two predictions: one based solely on the clinical factors, the other using the hybrid model of the clinical factors and the women’s CNN score.

The hybrid model more accurately predicted breast cancer risk than the BCSC model alone, particularly among non-Hispanic Blacks. Thus, the CNN model could be used to efficiently predict breast cancer risk and inform risk-stratified breast cancer screening and prevention strategies.

“Harnessing the power of AI to delve more deeply into the mammographic data has been pivotal to progress in breast cancer research, particularly in the area of risk prediction,” says Dr. Ha.

“In the future,” adds Dr. Crew, “we plan to apply our AI algorithm to large epidemiologic cohort studies, particularly for non-Hispanic Black women. We also plan to expand to multiple NewYork-Presbyterian Hospital sites.”