The effect of volumetric breast density on the risk of screen-detected and interval breast cancers: a cohort study

The female breast consists of fibroglandular (dense) and fat (nondense) tissue. It is known that breast density, measured as either the percentage or absolute amount of fibroglandular tissue in the breast, is a strong breast cancer risk factor [1‐3]. Besides this, high breast density is also known to hinder the detection of tumors during mammographic screening [4‐6]. These findings have led to legislation in the USA that mandates disclosure of breast density information to women undergoing mammography screening. Depending on their breast density, women may opt for supplemental screening, for example by ultrasound or magnetic resonance imaging (MRI).

When breast density measures are used as a tool to stratify women for different screening strategies, it is important that these measures are highly reproducible as it is undesirable that the likelihood of a woman being classified as having dense breasts is dependent on which radiologist she goes to. Another prerequisite is that these measures can be obtained in the daily practice of high-throughput screening.

Breast density can be measured using several methods. The most widely used methods are visual estimation using Breast Imaging - Reporting and Data System (BI-RADS) density categories and the semi-automatic area-based Cumulus method [7]. The former is mostly used by radiologists in clinical and screening practice. The latter is mostly used for research purposes. Both methods have been shown to be related to breast cancer risk. However, they also have some disadvantages. BI-RADS density classification has high inter-reader variability and Cumulus is very time-consuming. Hence, neither method is optimal for stratification in breast cancer screening [8‐12].

Nowadays, digital mammography (DM) is widely used and several software packages are available to measure breast density fully automatically on digital mammograms. Most packages report volumetric breast density measures, which are user-independent and developed with the intention to give an accurate and reproducible estimation of the amount of fibroglandular tissue in the breast using physics-based modeling based on information in the header of the raw digital images [13, 14]. This is in contrast to the aforementioned methods (BI-RADS and Cumulus) that measure the projected area of dense tissue on a mammogram. Density measured with automatic volumetric methods has also been examined in relation to breast cancer risk and results were found to be comparable to results from studies where BI-RADS or Cumulus were used to quantify breast density. In general, percentage dense volume (PDV) seemed to be a stronger breast cancer risk predictor than dense volume (DV) [15‐18].Absolute dense volume represents the amount of fibroglandular tissue in the breast and is hypothesized to be positively associated with breast cancer risk as this is considered to be the actual tissue where breast cancers develop [19, 20]. Percent dense volume is the proportion of fibroglandular (dense) tissue in the breast. This has sometimes been suggested to be a stronger risk factor than absolute dense volume, which would indicate that the ratio between the two tissues is important, or that nondense volume, which is part of the denominator of percent dense volume, is a protective factor for breast cancer risk. In a large meta-analysis, Pettersson et al. previously studied the associations between both absolute and percentage dense area and breast cancer risk. Their results suggest that percentage dense area is a stronger risk indicator than absolute dense area, but some heterogeneity between study results remained [3]. By using automatically assessed volumetric breast density measures instead of projected area measures we hope to gain more insight into these ambiguities and to find the best measure to be used in breast cancer risk prediction.

In light of discussions on the need for supplemental screening after a negative mammogram, measures that indicate the risk of a tumor that is not detected at mammography screening (interval cancer) may be even more relevant. BI-RADS density and other area-based density measures have been shown to be strongly related to the risk of interval cancer [4, 21‐26]. The relationship between automatic volumetric density methods and the risk of screen-detected and interval cancers, respectively, has not yet been studied. In this study, we used a fully automatic method to obtain several volumetric density measures, and assess their relationship with breast cancer risk. We distinguished between the risk of tumors that are detected at mammography (screen-detected cancers) or in the interval between two screening rounds (interval cancers), in a large cohort with digital mammograms from the population-based Dutch breast cancer screening program.

After excluding cases of screen-detected breast cancer, diagnosed on the basis of the first available digital screening examination, 523 breast cancers, including 299 screen-detected and 224 interval breast cancers were used for analyses. At the first digital screening mammogram the median age was 56 years (IQR 51–63) in women who did not develop breast cancer (N = 52,291), 59 years (IQR 54–64) in women who were diagnosed with screen-detected breast cancer and 55 years (IQR 50–62) in women diagnosed with interval breast cancer during follow up. The median total number of screening rounds in which women participated (analog (if applicable) and digital together) was 5 (IQR 2–8), 7 (IQR 4–9) and 4 (IQR 2–7) in women who were not diagnosed with breast cancer, were diagnosed with a screen-detected breast cancer and were diagnosed with interval cancer, respectively. The median follow-up time after the first digital mammogram until breast cancer diagnosis (event), death or end of follow up, was 4.2 years (IQR 2.0–6.2), 3.8 years (IQR 2.1–4.3) and 2.3 (IQR 1.3–4.0), in women with no diagnosis of breast cancer, screen-detected breast cancer and interval breast cancer, respectively (Table 1).

PDV was strongly negatively correlated with NDV (-0.70, p < 0.01)) and positively correlated with DV (0.27, p < 0.01). DV was positively correlated with NDV (0.50, p < 0.01). DV and PDV were both negatively correlated with age (-0.16 and -0.29, respectively, p < 0.01), whereas NDV was positively correlated with age (0.14, p < 0.01). There was strong correlation between breast measures of the left and right breast and between CC and MLO views. Correlations between the MLO views of the left and right breast for DV, PDV and NDV was 0.86, 0.91 and 0.98, respectively (N = 52,410 and p < 0.01 for all three). Correlation between MLO and CC views of the right breast was 0.86, 0.90 and 0.97 DV, PDV and NDV, respectively (N = 38,997 and p < 0.01 for all three).

In this study we found that automatically assessed volumetric breast density measures are positively associated with breast cancer risk and in particular with the risk of interval cancer. The difference in effect on interval and screen-detected cancers appeared slightly more pronounced for the percentage dense volume measures than for absolute dense volume measures. The VDG categories that are comparable to the BI-RADS categories had the highest risk estimates for association with interval breast cancers.

To our knowledge the relationship between volumetric breast density and risk of screen-detected and interval cancers, respectively, has not been studied separately before, but there are several studies on the association between volumetric breast density, determined on digital mammograms, and total breast cancer risk. An overview of their design and results is given in Additional file 3: Table S1. Four of the five presented studies examined both absolute and percentage dense volume in relation to breast cancer risk. In general, the risk estimates for percentage dense volume are somewhat higher than for absolute dense volume [15‐18]. Although our risk estimates for dense volume and percentage dense volume are in line with those of the studies presented in Additional file 3: Table S1, our percentage dense volume risk estimates (VDG, PDV) seem to be slightly lower than in the other studies [15‐18]. This could be explained by the fact that we do not have information on body mass index (BMI) or other risk factors in the routine screening program and therefore could not adjust for these risk factors. In particular the adjustment for BMI is known to increase the risk estimates for percent density measures. Despite this lack of adjustment for BMI, still the percent dense volume measures in our study appear to be more strongly related to the risk of interval cancer than the absolute dense volume measures.

As mentioned, no other studies have examined the ability of automated volumetric breast density assessment methods to identify women with a high risk of interval breast cancer. However, from a breast cancer screening perspective, this is important to know, since women with a high risk of interval breast cancer might benefit most from supplemental screening. Area-based breast density measures, such as BI-RADS density and the semi-automatic Cumulus thresholding method have been studied in relation to interval cancer risk. In these studies density was found to be more strongly related to the risk of an interval cancer than to the risk of a screen-detected cancer [4, 22‐26, 29]. Boyd et al. (2014) and Krishnan et al. (2016) were the only two studies of both absolute dense area and percent dense area in relation to the mode of breast cancer detection. They too found that breast density was significantly more strongly related to the risk of interval breast cancers than of screen-detected breast cancers. This difference in effect seemed stronger for percentage dense area than for absolute dense area, although this was not formally tested [22, 29].

Likewise, in our study percentage dense volume seems to be a better marker for interval breast cancer risk than dense volume. When looking at total breast cancer risk, this difference between percentage dense volume and absolute dense volume seems less pronounced. An explanation for this finding could be that the amount of dense tissue is etiologically relevant for breast cancer risk, and that when this dense tissue is occupying the larger part of the breast volume, these women will run a higher risk that a tumor will be masked at mammography and later diagnosed as interval cancer. If on the other hand, this same amount of dense tissue is accompanied by a large amount of fat tissue, large parts of the mammogram may still be well “readable” for cancer detection, and the probability that this tumor is detected at the screening examination will become higher.

Besides masking, another possible explanation for the relatively larger number of interval breast cancers in women with dense breasts is tumor aggressiveness. It is possible that tumors in dense breasts are more aggressive (grow faster) than tumors in nondense breasts. However, although some studies report that breast density is more strongly related to estrogen receptor (ER)-negative tumors than to ER-positive tumors, this was not confirmed in other studies [30‐34].

When interpreting our results, it should be kept in mind that we excluded prevalent breast cancer cases (screen-detected cancers diagnosed as a result of the first digital screening examination). The proportion of interval cancers is therefore larger (43%) in the current study than in the Dutch biennial screening program (26%). As interval cancers are relatively more common in dense breasts, the association between breast density and total breast cancer risk is probably somewhat overestimated. However, this is not the case for analyses separated for screen-detected and interval breast cancers.

In most European countries women are screened from the age of 50 years. In the USA most women are screened from the age of 40 years. In addition, biennial screening is common in European countries, while in the USA many women are screened yearly. We therefore looked at interval cancers diagnosed within the first year after a screening mammogram. We observed that 22% of the interval cancers in the group with the lowest breast density was found in that first year, while this was 60% in the group with the highest breast density, indicating an even stronger association between breast density and risk of interval breast cancer within the first year after screening.

A limitation of our study is that we could not compare the association between the automated volumetric density measurements and breast cancer risk with those of other commonly used breast density measurements such as BI-RADS. The reason for this is that BI-RADS density readings are not routinely performed in the Dutch breast cancer screening program. Previous studies directly comparing automated density measuring methods to BI-RADS density readings and other breast density measurements in relation to breast cancer risk identified similar positive associations for all density measurements, although they may differ in their classification of which women have dense breasts and which women have not [15‐18, 35, 36]. It remains as yet unknown which method most accurately reflects true density [17]. The method that we used in our study has been validated against MRI measurements of the fibroglandular tissue volume and by its fully automatic nature this gives objective and reproducible density measurements [13]. Although BI-RADS density is known to be strongly associated with breast cancer and interval breast cancer risk, the inter-reader reproducibility using this method is moderate [8‐12]. A critical issue now in the discussion of the value of supplemental imaging for women with dense breasts is that it may depend on the radiologist who interprets the mammograms as to whether a woman is classified as having dense breasts or not [12].

Another limitation is that at the time the mammographic examinations were performed, the CC view was not a routine view in our screening program. The CC view was only taken when women had their first screening examination, or on indication, e.g. because of high breast density or suspect lesions. Although for robustness of volumetric breast density estimates it may have been preferable to calculate density on MLO and CC views, in our study we did this on the MLO view only, to prevent effect estimates from being biased by the selective availability of the CC views.

A final limitation is that we only had information about age as an important confounder. Therefore, adjustment of the risk estimates for other confounders was not possible. The inability to adjust for parity and other reproductive factors may have overestimated the risk estimates somewhat both for dense volume and percent dense volume. Not adjusting for BMI is expected to have led to an underestimation of the risk estimate of percentage dense volume in particular, but all in all, the risk estimates for total breast cancer risk are quite comparable to what has been published in other papers on volumetric breast density [15‐18, 37]. Besides that, our study could be seen as an example for other screening programs where often little or no information is available on risk factors other than age.

Strengths of our study are the large cohort, the availability of unprocessed digital data for all its participants and its follow up since 2003 when digital mammography screening was introduced in the Netherlands. Due to linkage with the Netherlands Cancer Registry, which covers all residents of the Netherlands and is more than 95% complete, we have virtually complete information not only on screen-detected, but also on interval breast cancers.

We found that automated breast density measures can be used for breast cancer risk assessment and in particular, for the risk of a tumor that is not detected at the breast cancer screening. Automated measures are especially useful in programs where BI-RADS is not routinely determined or to improve reproducibility.