Mammography features for early markers of aggressive breast cancer subtypes and tumor characteristics: A population‐based cohort study

Abstract Current breast cancer risk models identify mostly less aggressive tumors, although only women developing fatal breast cancer will greatly benefit from early identification. Here, we evaluated the use of mammography features (microcalcification clusters, computer‐generated Breast Imaging Reporting and Data System [cBIRADS] density and lack of breast density reduction) as early markers of aggressive subtypes and tumor characteristics. Mammograms were retrieved from a population‐based cohort of women that were diagnosed with breast cancer from 2001 to 2008 in Stockholm‐Gotland County, Sweden. Tumor and patient characteristics were obtained from Stockholm Breast Cancer Quality Register and the Swedish Cancer Registry. Multinomial logistic regression was used to individually model each mammographic feature as a function of molecular subtypes, tumor characteristics and detection mode. A total of 4546 women with invasive breast cancer were included in the study. Women with microcalcification clusters in the affected breast were more likely to have human epidermal growth factor receptor 2 subtype (odds ratio [OR] 1.78; 95% confidence interval [CI] 1.24‐2.54) and potentially less likely to have basal subtype (OR 0.54; 0.30‐0.96) compared to Luminal A subtype. High mammographic cBIRADS showed association with larger tumor size and interval vs screen‐detected cancers. Lack of density reduction was associated with interval vs screen‐detected cancers (OR 1.43; 1.11‐1.83) and potentially of Luminal B subtype vs Luminal A subtype (OR 1.76; 1.04‐2.99). In conclusion, microcalcification clusters, cBIRADS density and lack of breast density reduction could serve as early markers of particular subtypes and tumor characteristics of breast cancer. This information has the potential to be integrated into risk models to identify women at risk for developing aggressive breast cancer in need of supplemental screening.


| Study population
Women aged less than 80 years diagnosed with breast cancer from 2001 to 2008 and recorded in the Stockholm-Gotland Regional Breast Cancer quality register (n = 9348) were sent invitations to participate in the LIBRO-1 population cohort study. A total of 5715 women (61%) consented to participate-they provided blood, answered detailed questionnaire on lifestyle including hormonal and reproductive factors and consented to retrieval of mammography images. Detailed information on the cohort has been published previously. [14][15][16][17][18][19] From these 5715 women, 1169 were excluded for the following reasons; one woman was excluded due to missing diagnosis date, 653 women had noninvasive breast cancer or missing invasiveness, and 515 women had multiple (including contralateral) breast cancer. This left 4546 women in our study. The flow chart in Figure 1 describes this selection. All study participants gave informed consent and the study was approved by the ethical committee at Karolinska Institutet.

| Mammographic sources
Mammograms, both analogue and digital, were retrieved from Departments of Radiology and information on mammography screening history were retrieved from the Stockholm-Gotland Regional Cancer Center mammography screening database. 16,24 Mammographic features were evaluated using measures that had the most clinical relevance as described in detail later.

| Microcalcification clusters
We used a method developed in our group for the detection of microcalcification clusters that can be applied on different digital systems and vendors, enabling incorporation of both analogue and digital images for large population studies. 25,26 This method comprises the following steps: (a) image preprocessing, primarily involving denoising, quality improvement and enhancement of small objects, (b) identification of microcalcification candidates, (c) filtering out noise (keeping only objects with shapes, sizes and appearances similar to microcalcifications) and grouping microcalcifications into clusters. 25 Two microcalcifications are defined to be in the same cluster if they are less than 4.1 mm apart and there has to be at least four microcalcifications to form a cluster. 25 This threshold was defined based on our experiment during the development of our algorithm in our earlier manuscript. 25 It is also similar to the threshold value used by the commercial software iCAD, to which we compared our results in our earlier manuscript. 25 For an example of microcalcification cluster detection in digital image, see figure 2 in previously published paper. 25 In our study, we evaluated the presence of microcalcification clusters on the cancerous breast and contralateral unaffected side using mediolateral-oblique (MLO) images closest to diagnosis, defined as 3 years prediagnosis to 3 months postdiagnosis.

| Mammographic density and density change
Percentage mammographic density was calculated using the areabased STRATUS algorithm, which has been developed to analyze a range of image formats, including both analogue and digital images, with automation of density change measurements over time was used. 27 This method has an in-built alignment protocol, which reduces nonbiological variation of breast density changes in women. 27 This measurement was then converted to a categorical variable using cutpoints (2%, 18%, 49%). These cut-points were taken from previously published work to group the percent density into four breast composition categories in line with clinically relevant Breast Imaging Reporting and Data System (BIRADS; American College of Radiology, Reston, VA) score. This computer-generated score is termed and abbreviated as cBIRADS. 28 In the statistical analysis of density, percent density measurements of the contralateral side to the breast cancer were used to ensure that the tumors did not affect image measurements. Examination closest to date of diagnosis was used. After our earlier study that showed increased probability of interval cancer with high mammographic density, 16 we further investigated the clinical relevance of the associations of mammographic density, evaluated using cBIRADS, 20 with subtype, tumor characteristics and detection mode.
For density change analysis, relative annual density area change (RDC) on the contralateral breast was computed by taking the difference in area density between two time points (defined as first and last mammography prediagnosis) per baseline density of each women using the equation 29 where d 1 denotes area density at first mammography t 1 , d 2 denotes area density at last mammography t 2 , and t are times on a yearly unit scale. For classification of molecular subtypes, additional data on ER, PR, HER2 and Ki-67 were obtained from medical and pathology records. 14 A dataset containing RNA-sequenced PAM-50 gene expression was used as training dataset to classify particular molecular subtypes using ER, PR, HER-2, Ki-67 and age at diagnosis as inputs using a random forest algorithm. 14 Full details of the classifier method with robust sensitivity analyses has been published earlier. 14 Screening history from the mammography-screening database at the Stockholm-Gotland Regional Cancer Center was used to determine detection mode. 16

| Microcalcification clusters
Microcalcification clusters were present in the affected breast for 35% of the women and in the unaffected side for 26% of the women.

| Mammographic density (cBIRADS)
Based on cBIRADS categories of the contralateral breast, the percentages of women were 9%, 38%, 47% and 6% in categories A, B, C and D, respectively. No significant associations between molecular subtypes and cBIRADS categories were found. Larger tumors were associated with higher density cBIRADS B, C and D, and interval vs screen-detected cancers were associated with cBIRADS C and D compared to cBIRADS A. Detailed results are presented in Table 3.

| Mammographic density change
36% of the women considered in this analysis showed no reduction in relative density area. Results suggested that the lack of relative density area reduction over time was associated with Luminal B subtype vs Luminal A subtype OR 1.76 (95% CI 1.04-2.99) ( Table 4). In addition, the lack of density area reduction over time also showed increased odds for interval vs screen-detected cancers with OR 1.43 (95% CI 1.11-1.83). However, no significant associations were observed between relative density area reduction and tumor characteristics including tumor size, lymph status and tumor grade (Table 4).

| DISCUSSION
In our study, we found that mammographic features including microcalcification clusters, cBIRADS density and density change have the potential to be useful predictors for particular invasive breast cancer subtypes and tumor characteristics in the early stages of tumorigenesis. Women with microcalcification clusters in the affected breast were more likely to present with the HER-2 subtype but potentially less likely to be of basal subtype compared to Luminal A subtype.
These women had an elevated probability of being screen-detected, rather than interval cancers, compared to women without microcalcification clusters. Reassuringly, these associations were only observed on the affected breast side, but not on the contralateral Microcalcifications in breast tissues are formed through calcium mineralization processes and have been used as indicators of early breast cancer and in particular ductal carcinoma in situ. 10 Our study is the first to suggest that microcalcification clusters in the affected breast are associated with a reduced probability of having a basal subtype vs Luminal A subtype. It is interesting to note that in studies of unaffected women, microcalcification formation has been shown to be positively associated with breastfeeding, 11,26 and that among breast cancer patients a history of breastfeeding has been shown to be protective for the basal subtype. 14,26  38 Therefore, we think that it is problematic to study density in the affected breast, as density in the affected breast would be related to the tumor size, which in turn might be associated with molecular subtype. For a study aimed at testing the hypothesis that density heterogeneity between left and right breasts is associated with molecular subtypes of breast cancer, it would be important to have strong control of the timing of mammograms to ensure that there are no signs of tumors in all included mammograms.

| CONCLUSIONS
Mammographic features (microcalcification clusters, cBIRADS density and density change) could potentially be used as early markers to identify women at increased risk of developing aggressive breast tumors. Current breast cancer risk models identify women who will be at risk for breast cancer, 4-6 but no risk model identifies women at risk for aggressive disease. Future research should evaluate the utility of combining breast cancer risk factors with mammographic features in existing risk models, [4][5][6] to identify women at risk for developing aggressive breast cancer and in need of supplemental screening.

CONFLICT OF INTEREST
PST consulted for AstraZeneca and Duke-NUS. Other authors declare no potential conflicts of interest.

DATA AVAILABILITY STATEMENT
Datasets generated during and/or analyzed in this study are protected under data protection laws in Sweden and could not be made publicly available. Application for data can be made via the Swedish National Board of Health and Welfare and Statistics Sweden. More information is available from https://bestalladata.socialstyrelsen.se/data-forforskning/ and http://www.scb.se/Vara-tjanster/bestalla-mikrodata/

ETHICS STATEMENT
All study participants gave informed consent and the study was approved by the Regional Ethical Review Board in Stockholm, Sweden (Karolinska Institutet, DNR2009/254-31/4).