Background: In Australia and elsewhere, there is a growing interest in delivering more personalised, risk-based breast cancer screening protocols. This requires reliable, feasible and accurate estimates of risk. The US National Cancer Institute Breast Cancer Risk Assessment Tool (BCRAT) and the AutoDensity fully automated mammographic density measurement tool have each been shown to stratify women into groups according to their risk of breast cancer; the AutoDensity tool also provides information on the likely sensitivity and specificity of mammographic screening tests. The Australian 'lifepool' cohort of over 53,000 women recruited predominantly through BreastScreen Australia screening program offers an opportunity to validate these tools and examine how they can be combined to estimate various risks. Aim: To validate BCRAT and AutoDensity on a large Australian population, and examine how the tools can be combined to provide information on breast cancer risk and the accuracy of the screening test. Methods: We use lifepool cohort questionnaire data and linked screening records and mammograms, cancer registrations and death records to describe the association between BCRAT and AutoDensity scores assessed at the time of screening and future breast cancer diagnosis. We use hazards models to account for censoring and describe outcomes according to mode of detection (screen-detected, interval cancers or other). Our primary analysis is restricted to women in the historical screening target age range of 50-69 with no prevalent breast cancer diagnosis on entry to the lifepool cohort. Results: The primary analysis included approximately 40,000 women with a median follow-up period of 4.5 years (1.1-6.5 years). The BCRAT tool generated a median 5-year breast cancer risk score of 1.5% (range 0.6%-22.0%). Compared with women in the lowest quintile of this score, women in the highest quintile had a 2.3-fold risk (95% CI 1.7-3.0, P < 0.001) of incident invasive breast cancer. For the approximately 35,000 women with digital screening mammograms on enrolment, women in the highest quintile of AutoDensity values had a 1.5-fold risk (95% CI 1.1-2.0 P = 0.011) of incident invasive breast cancer and a 2.6-fold risk (95% CI 1.1-6.2, P = 0.034) of an interval cancer compared with women in the lowest quintile. With BRCAT and AutoDensity measurements weakly correlated (r2= 0.003, P = 0.05), we demonstrate various approaches to combining this information to stratify women according to breast cancer risk and risk of an interval cancer. Conclusion: The US National Cancer Institute Breast Cancer Risk Assessment Tool and the AutoDensity mammographic density tool can be used to stratify breast cancer screening participants into risk groups according to their future breast cancer risk and the risk of an interval cancer. This is likely to be of interest to screening program managers and policy-makers, and women considering screening participation.
Using a convolutional neural network to predict readers' estimates of mammographic density for breast cancer risk assessment
