In a dataset enriched for African American women, BRCA mutation carriers and those with benign breast disease, a mammography-based deep learning model demonstrated a five-year AUC of 63 percent for predicting breast cancer in comparison to 54 percent for BI-RADS assessment.
Emerging study findings suggest a mammography-based deep learning (DL) model may be beneficial in detecting pre-cancerous changes in women at high-risk for breast cancer.
For the retrospective, case-control study, recently published in Radiology: Artificial Intelligence, researchers assessed the use of Mirai, a DL model trained on approximately 211,000 screening mammograms, for 6,266 mammography exams from 2,043 women (median age of 56.4 years), including 910 African American women and 853 White women. Out of 342 women who had genetic testing information, 62 had BRCA mutations, according to the study.
The researchers found that the DL model demonstrated a one-year area under the receiver operating characteristic curve (AUC) of 71 percent and a five-year AUC of 65 percent for predicting breast cancer. While the Breast Imaging Reporting and Data System (BI-RADS) system had a higher one-year AUC than the DL model (73 percent versus 68 percent), the DL model outperformed the BI-RADS system for longer-term breast cancer prediction at five years (63 percent AUC versus 54 percent AUC).
The DL model had higher five-year AUCs for low and intermediate breast cancers (64 percent and 66 percent respectively) in comparison to high-grade breast cancer (60 percent AUC), according to the study. The researchers also noted that imaging of the breast with future cancer plays a key role with the AI model, citing mirroring experiments that showed a 62 percent AUC for positive mirroring and a 51 percent AUC for negative mirroring.
“Results from selective mirroring, along with better performance of the DL model for near-versus longer-term prediction, suggests that DL may detect premalignant or early malignant changes before they become apparent,” wrote study co-author Dezheng Huo, Ph.D., a professor in the Departments of Public Health Sciences and Medicine at the University of Chicago Medical Center, and colleagues.
The researchers posited that adjunctive use of the DL model may facilitate improved accuracy in assessing patients at higher risk of developing breast cancer.
“In our study, the combined DL and BI-RADS scores showed better short-term discrimination than either model evaluated alone, and while this difference was not statistically significant, it suggests that mammography DL tools could supplement the assessment of screening mammograms for more accurate near-term risk stratification,” noted Huo and colleagues.
In regard to study limitations, the researchers acknowledged that beyond Black and White women who comprised the majority of the cohort, there was limited sample size for other self-reported racial groups. Limitations with sample size and granular detail on the BI-RADS assessment categories 4a, 4b and 4c hampered comparisons of the AI model and BI-RADS for short- and long-term predictions, according to the study authors. Noting the assessment of high-risk women with an AI model trained on a dataset with a likely lower risk profile, the researchers cautioned that their findings may not be readily extrapolated to women at average risk for breast cancer.