30 November 2021
Vachon CM, Scott CG, Winham S, Norman A, Hruska CB, Brandt KR, Kerlikowske K
Commercially available AI system for breast cancer detection shows promise for risk prediction, including among women with dense breasts
AIM AND OBJECTIVE
AI algorithms have been developed to improve detection of breast cancer at the time of mammography. Whether these algorithms can assist in risk prediction has not been well studied.
MATERIALS AND METHOD
We identified 3,387 women with invasive breast cancer (n=462 interval, 2309 screendetected) and 7,140 controls, matched on age and date of mammogram, from two US mammography cohorts between 2007 and 2017. We obtained 2D full-field digital mammograms an average 3.6 years (6 months to 10 years) prior to cancer diagnosis and assessed a malignancy risk score using a commercially available AI system (Transpara, ScreenPoint Medical) and volumetric density using Volpara. We used conditional logistic regression (odds ratios (OR) and 95% confidence (CI) intervals), adjusted for age and BMI, and C-statistics (AUC), to describe the association of the continuous Transpara score (per 1 unit increase) with invasive breast cancer. We performed analyses stratified by dense breasts (BI-RADS density a-b vs. c-d), time to cancer (≤3.6 years vs. >3.6 years), and interval vs. screen-detected breast cancer.
Adjusted for age and BMI, a one unit increase in the Transpara risk score was associated with 23% greater odds of invasive breast cancer [OR=1.23 (1.21-1.25); AUC=0.65 (0.64-0.67)]. Associations were similar among women with dense [OR=1.24 (1.20-1.27)] and non-dense breasts [OR=1.20 (1.18-1.23); (p=0.30); AUCs= 0.65-0.66] and for interval and screen detected cancers [ORs both 1.24, (p=0.98) and AUCs both 0.66]. There were stronger associations for the Transpara risk score assessed on mammograms within 3.6 years prior to breast cancer diagnosis [OR=1.30 (1.26-1.33); AUC=0.68 (0.67-0.70)] vs. greater than 3.6 years [OR=1.18 (1.15-1.20); (p<0.001); AUC=0.63 (0.61-0.64)]. Inclusion of volumetric density measures in models with the Transpara risk score improved discrimination, but the increase was only statistically significant for prediction of interval cancers; the AUC increased from 0.66 to 0.72 [change in AUC=0.06 (0.03-0.09)] with inclusion of percent volumetric density.
Transpara’s imaging-based measures combined with volumetric density improved discrimination of invasive breast cancer. Including these measures in risk models could better inform tailored screening and supplemental imaging strategies. Commercially available AI system for breast cancer detection shows promise for risk prediction, including among women with dense breasts