We pooled data from two nested case-control studies from women ages 35 and older receiving full field digital mammography (FFDM) screening examinations at Mayo Clinic in Rochester, Minnesota (MN) (years 2008 to 2017) and the Hospital of the University of Pennsylvania in Philadelphia, Pennsylvania (PA, UPENN) (2011 to 2014) from Hologic-Selenia mammography units. Incident invasive breast cancers were identified from linkages to institutional and state cancer registries and included at least six months after screening mammography. Both screen-detected and symptomatically detected cancers were included. Three controls without a history of breast cancer were initially matched to cases based on facility, race, age, and year of screening mammogram; however, due to various reasons, some controls dropped out due to breast area exclusions and imaging artifacts, resulting in certain cases having only one matched control in the final analyses. For this analysis, only those with non-missing information on breast density measures and BMI, and who self-identified as non-Hispanic Black or non-Hispanic White, were included. The final study population included 3699 women, of whom 526 were Black and 3173 were White, with each patient contributing one screening mammogram to the analysis.
Self-reported demographic and reproductive breast cancer risk factors, such as family history of breast cancer, were available from mammography screening questionnaires administered as part of routine practice. BMI values were extracted from electronic medical records recorded on the screening date, if available, and if not, within 1 year before or after the screening mammogram. The study was HIPAA-compliant and approved by the Institutional Review Boards at both Mayo Clinic and University of Pennsylvania, and a waiver of informed consent was granted for this review of existing clinical data.
BI-RADS breast density was obtained from radiology reports, visually determined by one interpreting radiologist, as is standard practice in the U.S. The measures were derived according to the American College of Radiology BI-RADS Atlas 4th Edition [8] definitions and provide an assessment of a woman’s breast density in one of the four ordinal categories: almost entirely fatty (a), scattered fibroglandular density (b), heterogeneously dense (c), or extremely dense (d).
Quantitative measures of volumetric breast density were obtained using Volpara™, a fully automated software that computes volumetric breast density from FFDM images. Briefly, Volpara uses an area of entirely fatty tissue as a reference point to estimate the thickness of dense tissue at each pixel in the image [26]. Estimates of dense volume are obtained by summing the estimated dense tissue across all pixels in the breast image through multiplying the estimated breast area by the breast thickness. Volumetric percent density is obtained by dividing the estimated dense volume by the total breast volume. Breast density was measured on the cranio-caudal (CC) and medio-lateral oblique (MLO) views for both left and right breasts for both cases and controls. For each woman, the estimates from all 4 views were averaged to obtain the final density values. We focused on values of dense as opposed to non-dense tissues, as we are interested in translation to the clinical setting. Though Volpara software can estimate a BI-RADS-like categorical breast density measure, we did not include this in our analysis.
We used conditional logistic regression to compare matched cases and controls to assess the association between the density measures (BI-RADS density, dense volume DV, volumetric percent density VPD) and risk of invasive breast cancer. We estimated the association of density measures using odds ratios and discriminatory accuracy using the area under the receiver operator curve (AUC) and 95% confidence intervals (CI). The AUC was calculated within matched pairs. Odds ratios for volumetric percent density and dense volume are presented per 1 standard deviation in the log-transformed measure. BI-RADS density was modeled as an ordinal trend across categories. This was done given the small sample size of Black women in the extremely dense category. AUCs when BI-RADS density was modeled as a four-level categorical variable were very similar to models using ordinal trend (See Additional File 1). Analyses were adjusted for age and BMI as continuous variables, and comparisons were made between models adjusted for age alone and those adjusted for both age and BMI. We also tested models including family history as a covariate, as well as an interaction between menopausal status (age greater than or equal to 55 or age less than 55) [18] and BMI. Analyses were performed overall and stratified by race groups (Black vs. White). Differences in associations of density measures with overall invasive cancer by race groups (UPENN Black or White Mayo Clinic and UPENN) were tested with inclusion of interaction terms between race and density in the conditional logistic regression models.
For primary analyses, AUC was compared between race groups based on results from 1000 bootstrapped samples. In analyses of age and BMI subgroups, non-conditional logistic regression analysis was used, and matching factors were also included in the adjustment variables.
As a secondary analysis, we explored the association between each of the density measures and breast cancer risk stratified by age (as a proxy for menopausal status), obesity (BMI > 30 kg/m2), and time to breast cancer diagnosis. The time to breast cancer diagnosis was used to see if breast density was better at detecting breast cancers close to diagnosis compared with predicting breast cancers in the long term. Statistical analyses were carried out using SAS version 9.4. The type I error rate for CIs and statistical tests was set at 0.05, and two-sided tests were used.
Results section
The nested case-control study consisted of 1013 invasive breast cancer cases and 2686 matched controls, with 14% Black and 83% White women. White women contributed 875 cases, and 2298 controls, and Black women contributed 138 cases and 388 controls. There were differences in both BMI, family history, and breast density between Black and White women (Table 1). Among controls, Black women had a higher mean BMI than White women (30.5 ± 9.7 kg/m2 vs. 27.2 ± 8.2 kg/m2). White women who were controls had a higher percentage of first-degree relatives with a history of breast cancer (22.5%) compared with Black controls (17.5%). Using radiologist-reported clinical BI-RADS measures density, Black women were less likely to have heterogeneously dense (18.6% vs. 31.5% for) or extremely dense (0.3% vs. 4.8%) breasts than White women. Similar findings were seen for volumetric percent density, with lower mean volumetric percent density for Black (6.0% ± 4.2%) controls compared with White (8.0%±5.3) controls. However, mean dense volume was higher for Black (64.8 ± 30.5 cm [3]) controls vs. White controls (58.7 ± 29.2 cm [3]).
BI-RADS density, dense volume, and volumetric percent density were all significantly associated with breast cancer risk in both Black and White women, when adjusted for age alone, both age and BMI, and age, BMI, and menopausal status (Table 2). Adjustment for BMI in addition to age resulted in stronger associations for BI-RADS density and volumetric percent density, but not dense volume, whose association was unchanged after BMI adjustment. The OR for BI-RADS density increased from 1.44 (95% CI 1.30–1.59) to 1.58 (95% CI 1.42–1.77), and volumetric percent density increased from 1.25 (95% CI 1.15–1.36) to 1.40 (95% CI 1.28–1.54) after additionally adjusting for BMI, suggesting negative confounding of the association by BMI. Discriminatory accuracy also increased for all models after adjustment for BMI, except for dense volume. Similar findings with BMI adjustment were also observed in the race-stratified analyses (Table 2). We additionally adjusted for family history of breast cancer and included an interaction between menopausal status and BMI. Neither significantly changed our results and thus were not included in the final models.
When comparing differences in the breast density associations by race, BI-RADS density was more strongly associated with breast cancer risk for Black women (OR 2.06, 95% (1.45, 2.91) compared with White women (OR 1.55, 95% CI 1.38, 1.74) (p-interaction = 0.04, adjusted for BMI and age). For volumetric percent density and dense volume, ORs were larger for Black women than White women, but not statistically different. AUCs were generally higher for Black compared with White women, with the exception of dense volume, which had the same AUC for Black and White women. For Black women, BI-RADS density (adjusted for age and BMI) had the highest discriminatory accuracy (AUC = 0.634) of all models, and for White women, dense volume (adjusted for age and BMI had the highest AUC (AUC = 0.603).
As a secondary analysis, we explored the association between each of the density measures and breast cancer risk stratified by age (as a proxy for menopausal status), obesity (BMI > 30 kg/m2), and time to breast cancer diagnosis (less than or greater than three years, Table 3). In the age-stratified models, ORs for BI-RADS density, volumetric percent density, and dense volume were higher for Black women < 55 years, but only statistically significantly higher for volumetric percent density (p-interaction = 0.01). There were no significant differences in ORs by age for White women. In the BMI stratified models, all density measures had stronger associations with breast cancer risk for obese women, however, the difference was only statistically significant for volumetric percent density for obese Black women (BMI ≥ 30 kg/m2 OR = 2.55 vs. BMI < 30 kg/m2 OR = 1.31, p-interaction = 0.01). In the models stratified by time to breast cancer diagnosis, there were no significant differences in breast density associations for Black or White women.