Mammographic tumour appearance is related to clinicopathological factors and surrogate molecular breast cancer subtype

Mammographic tumour appearance may provide prognostic useful information. For example, spiculation indicates invasiveness, but also better survival compared to tumours with other appearances. We aimed to study the relationship between mammographic tumour appearance and established clinicopathological factors, including surrogate molecular breast cancer subtypes, in the large Malmö Diet and Cancer Study. A total of 1116 women with invasive breast cancer, diagnosed between 1991 and 2014, were included. Mammographic tumour appearance in relation to status for oestrogen receptor (ER), progesterone receptor (PR), human epidermal growth factor receptor 2, histological grade, Ki67 and molecular subtype was analysed using various regression models. All models were adjusted for relevant confounders, including breast density, which can affect mammographic appearance. The results consistently showed that spiculated tumours are indicative of favourable characteristics, as they are more likely to be ER and PR positive, and more often exhibit lower histological grade and lower Ki67 expression. Furthermore, spiculated tumours tend to be of luminal A-like subtype, which is associated with a good prognosis. The establishment of associations between mammographic tumour appearance and clinicopathological factors may aid in characterizing breast cancer at an earlier stage. This could contribute to more individualized breast cancer treatment in the future.

to invasiveness, tumour size and axillary lymph node involvement within a smaller subset of the same cohort used in this study, and found an association between spiculation and invasiveness, regardless of breast density 5 .
In addition to clinicopathological factors, surrogate molecular subtypes aid in the prediction of prognosis and the choice of appropriate treatment 8 . It has been suggested in the literature that the presence of calcifications on mammography is more common in the human epidermal growth factor receptor 2 amplified (HER2+) subtype 9,10 , while triple-negative breast cancer (TNBC) is more likely to present as a mass 10 . Few previous studies have, however, investigated the combination of the information available from mammographic appearance, clinicopathological factors, including molecular subtypes, and breast density. The aim of the present study was thus to investigate the relationship between mammographic tumour appearance and clinicopathological factors, including surrogate molecular subtypes, in the large cohort of incident breast cancer cases within the Malmö Diet and Cancer Study.

Methods and material
Study population. Data were collected from the Malmö Diet and Cancer Study (MDCS), a large population-based prospective cohort study, in which 28,098 individuals were enrolled from 1991 to 1996, of which 17,035 were women 11 . Inhabitants aged 44-74 years were recruited from the city of Malmö in the south of Sweden. The original intent of the study was to investigate possible associations between diet and cancer. At the baseline examination, anthropometric variables (blood pressure, height, weight, lean body mass and body fat mass) and blood samples were collected, and a comprehensive questionnaire was filled in 12 . The MDCS is updated regularly with information on new cancer cases through data collection from national registers: the Swedish Cause of Death Register, the Swedish Cancer Register, and the Regional Tumour Register for Southern Sweden 5 . Women with prevalent breast cancer at baseline or a history of breast cancer (n = 572) were excluded. A total of 1242 women in the cohort were diagnosed with incident breast cancer between 1991 and 2014. After the exclusion of carcinoma in situ (n = 105) and bilateral breast tumours (n = 21), 1116 women with invasive breast cancer remained eligible for the present study. Informed consent was obtained at the baseline examination. The Ethics Committee at Lund University approved this study (Official Records Nos. 652/2005 and 166/2007). The study was carried out in accordance with the declaration of Helsinki.

Mammography.
A protocol was set up to extract information from the original mammography report at the time of cancer diagnosis. This protocol had several variables, including mammographic tumour appearance, breast density and mode of cancer detection 5 . All screening mammograms were double read by two breast radiologists, but the diagnostic imaging during follow-up assessment was performed by one breast radiologist. Mammograms of clinically detected cancers were single read by one breast radiologist. When any of the information was lacking in the original mammography report, the mammograms were re-read by an experienced breast radiologist (SZ and/or HS).
Mammographic tumour appearance. Information on the most dominant mammographic appearance of the tumour was obtained retrospectively. It should be noted that more than one appearance could have been readily visible and described in the original report, but only the most dominant appearance was recorded in the study protocol. Tumours were classified into the following comprehensive categories, based on the work by Luck et al. 13 : well-defined mass, partly ill-defined mass, ill-defined/diffuse mass, spiculated mass, comedo-type microcalcifications, non-specific calcifications, architectural distortion and asymmetrical density. For statistical analysis, these categories were converted into five larger categories: distinct mass (including well-defined and partly ill-defined tumours), ill-defined mass, spiculated mass, calcifications (including comedo-type and nonspecific calcifications) and tissue abnormality (including the less frequent features architectural distortion and asymmetrical density). Some of the appearances are illustrated in Fig. 1.
Breast density. In the clinical setting at the Department of Breast Radiology in Malmö, breast density is divided into three groups: fat involuted, moderately dense and dense (Fig. 2). These three groups were employed in the present study. Fat involuted corresponds to Breast Imaging Reporting and Data System (BI-RADS) 4 14 density score 1, moderately dense to BI-RADS 4 density score 2-3, and dense to BI-RADS 4 density score 4.
Mode of cancer detection. Method of cancer detection was divided into screening-detected or clinically detected, i.e. either detection via the breast cancer screening programme that started in Malmö in 1990, or detected clinically, due to a lump in the breast or other symptoms that caused the woman to seek medical attention. Thus, clinically detected tumours also include interval cancers, i.e. cancers diagnosed between screening episodes.    15 .
Histological grade and tumour type. Breast cancers diagnosed from 1991 to 2004 were included in the first TMA and the histological grade (also Nottingham grade or Elston grade 17 ) and histological type according to the World Health Organization classification was re-assessed by an experienced breast pathologist. For the period 2005-2014, this information was extracted from medical records 18 .
Ki67. Information on IHC proliferation marker Ki67 expression was collected during three time periods, from TMA assessment 1991-2004 and 2005-2007, and from medical records 2008-2014. Estimates of Ki67 expression can vary due to inter-and intra-observer variability, and due to variability in staining between pathology laboratories 19,20 . Ki67 status in the MDCS cohort varies during the long follow-up period of 13 years. Thus, using the same cut-off values for all three time periods could be misleading. To overcome this risk in the MDCS cohort, the values were divided into three equally sized groups for each time period 18 . The group with the lowest scores is denoted "low", middle scores "intermediate", and highest scores "high".
Tumour size. Data on tumour size were collected from medical records during the entire follow-up period and were based on histopathological measurements. In multifocal breast cancer, the largest focus was used. Statistical methods. Descriptive statistics, logistic regression, ordinal regression and multinomial logistic regression were carried out. The relation between mammographic appearance and ER, PR, and HER2 status was analysed with logistic regression, generating odds ratios (OR) and 95% confidence intervals (CI). To study the distribution of tumour appearance within the three categories of histological grade and Ki67, ordinal regression models were used, generating OR and 95% CI. The proportional odds assumption was assessed for each model, and if it holds, the OR:s can be interpreted as odds of being category 3 compared to the combined odds of being category 1 and 2. Potential associations between mammographic tumour appearance and surrogate molecular breast cancer subtypes were assessed through multinomial logistic regression, where distinct mass was set as a reference for the tumour appearances, and luminal A-like as a reference for the subtypes, yielding relative risk ratios (RRR) with 95% CI. All statistical calculations were adjusted for breast density, mode of cancer detection, age at diagnosis and tumour size. Breast density was divided into three groups, and mode of detection into screening or clinical, as described in the mammography section above. Age at diagnosis was divided into quartiles. Tumour size was dichotomised into ≤ 20 mm and > 20 mm. In all the analyses, a p-value of less than 0.05 was considered statistically significant. Statistical analyses were performed with Stata version SE 14.2. Table 1 provides information on the study population, including patient and tumour characteristics, in relation to mammographic appearance.

Results
Tumour appearance in relation to clinicopathological factors. Ill-defined tumours, spiculated tumours and tumours presenting as tissue abnormality were more likely to be ER+ than ER−, as compared to a distinct mass, according to logistic regression analysis, with an adjusted OR (OR adj) of 2.0 (CI 1.1-3.6), 6.0 (CI 3.2-11.2) and 4.4 (CI 1.0-19.6), respectively (Table 2). Furthermore, spiculated tumours were more likely to be PR+ than PR− compared to a distinct mass, with an OR adj of 1.7 (CI 1.2-2.5) ( Table 2). We found no statistical evidence for an association between mammographic appearance and HER2 status (Table 2). However, ill-defined masses and tumours presenting as calcifications were slightly more often HER2+ than were the other tumour features in terms of distribution. According to ordinal regression analysis, the odds of spiculated tumours, compared to distinct masses, to be histological grade 3, compared to the combined categories of grade 1 and 2, were 50% lower, OR adj 0.5 (CI 0.4-0.7), given that all the other variables in the model were kept constant ( www.nature.com/scientificreports/ Tumours presenting as tissue abnormalities were also more likely to be of grade 1 or 2, rather than grade 3, compared to distinct masses, with an OR adj of 0.3 (CI 0.1-0.6). Spiculated tumours were more likely to exhibit lower Ki67 expression compared to distinct masses, with an OR adj of high Ki67 compared to the combined categories  Tumour appearance in relation to surrogate molecular subtypes. The frequencies of the mammographic appearances within the four breast cancer subtypes are given in Table 4. It was found to be more likely that an ill-defined mass would be of luminal A-like subtype than TNBC, than would a distinct mass, with an RRR adj of 0.5 (CI 0.2-0.9), according to multinomial logistic regression analysis ( Table 5). The relative risk of a spiculated tumour, compared to a distinct mass, being luminal B-like, HER2+ subtype or TNBC, compared to luminal A-like was lower: RRR adj 0.6 (CI 0.4-1.0), 0.4 (CI 0.2-0.8) and 0.1 (0.1-0.3), respectively. Or expressed more generally, it is more likely that a spiculated tumour will be luminal A-like than luminal B-like, HER2+ subtype or TNBC. No evidence was found of any association between molecular subtype and tumours presenting as calcifications. Moreover, tumours presenting as tissue abnormality were less likely to be TNBC than luminal A-like, with an RRR adj of 0.2 (CI 0.0-0.8) ( Table 5).

Discussion
In this large study on 1116 cases of incident breast cancer, we found strong evidence supporting that established clinicopathological factors and surrogate molecular subtypes differ in regard to mammographic tumour appearance. A particularly interesting result was that all the methods of statistical analysis indicated an association between spiculation and favourable tumour characteristics. The finding that spiculated tumours are more often ER+ and PR+ is in agreement with several previous studies [23][24][25][26][27] . However, the cut-off values for hormone receptor positivity and tumour appearance categorization differ between some studies 23,26 . In agreement with several other studies, we found spiculation to be associated with lower histological grade [28][29][30] and lower Ki67 values 23,26 . A previous study 31 , in which a different categorization of mammographic appearance was used, revealed an association between higher histological grade and spiculation with calcifications. However, in that study, the association with higher histological grade was even  Table 5. Mammographic appearance in relation to molecular subtypes. *Adjusted for age (categorical), tumour size, mode of detection and breast density. www.nature.com/scientificreports/ stronger in non-spiculated tumours, both with and without calcifications. Despite some methodological differences between our study and previous studies, it is clear that spiculation is indicative of favourable tumour characteristics. On the cellular level, the spicules of malignant tumours can represent tumour infiltration, a desmoplastic response in the adjacent stroma or periductal fibrosis 32 . However, the link between these features and the favourable characteristics of spiculated tumours is not clear. Future studies on spiculation and the molecular mammary microenvironment may increase this knowledge. According to our findings, ill-defined masses were also more likely to be ER+, however, we could find no evidence in the literature supporting or contradicting this finding. This could be because the category ill-defined mass is seldom used, and when it is, often with other outcomes 13,29,33 than those considered in the present study. Regarding HER2, we found no evidence of any association with tumour appearance. Previous studies have shown an association between spiculation and HER2− 23,26 as well as calcifications and HER2+ 9,[34][35][36] . In terms of frequency, the majority of spiculated tumours in our study were HER2− (94.5%), and calcifications were more often HER2+ (15.5%). Hence, our results point in the same direction, although with weak statistical support. Tumours presenting as tissue abnormality are more likely to be ER+, of lower histologic grade and lower Ki67 expression. The number of observations in this group is however small, making it difficult to draw any reliable conclusions.
Information on surrogate molecular subtypes is essential in breast cancer management. These were therefore included in this study to refine our analyses and to make the results more clinically comprehensible. Spiculated tumours were more commonly luminal A-like subtype, which is in line with the findings of three previous studies 26,33,37 , one of which 37 highlighted women under the age of 40. Breast density, which is higher in younger women, is an interesting factor that can affect tumour appearance to some degree. By adjusting for breast density, we were able to show that the association between spiculation and luminal A-like subtype persists, regardless of density. Tamaki et al. 38 found that masses classified as having indistinct margins on mammography, according to BI-RADS, were more often HER2+ subtype or TNBC than luminal type cancers. This is in contrast to our finding that ill-defined masses are more often luminal A-like than TNBC. The reason for this discrepancy is not known, but could be due to differences in mammographic categorization and population sampling. In the present study, the appearance was extracted from the original radiology report, and BI-RADS is not used in clinical practice at our department. In addition, their population was younger, with a median age of 50 years (range 27-89 years), compared to a median age of 66 years (range 45-91 years) in the present study. Associations between various types of calcifications and HER+ subtype have been described previously 36,39,40 . However, we found no evidence of this in our study. In order to increase group sizes, we combined all types of calcifications into one group, and it was therefore not possible to study differences between various types of calcifications. Several other studies have reported that TNBC is associated with a mammographic mass and can be mistaken for a benign breast lesion 10,[41][42][43][44] . It is therefore of the utmost importance to investigate these lesions carefully. In terms of frequency, a large proportion of the distinct masses in our study constituted TNBC (17.3%), as opposed to spiculated tumours, of which only a few (3.1%) were TNBC (Table 4). However, a potential association between TNBC and distinct mass in our material cannot be directly identified due to the setup of statistical analysis.
Some issues require consideration. Firstly, this was a retrospective observational study performed at a single department. Secondly, the women with breast cancer in the MDCS tend to be of ethnic Swedish descent, and with a higher level of education than the average population 45 , which could limit the representativeness of the findings. However, the clinicopathological factors studied were distributed as expected in routine clinical practice. Hence, we believe there is a low risk of selection bias, and that the internal comparisons should not be affected. Thirdly, the use of TMA for evaluation of clinicopathological parameters is worth considering as there is always a risk that the cores obtained for TMA do not reflect the original tumour and its potential heterogeneity correctly. However, this risk is reduced by obtaining two cores from different areas of each tumour, an approach that has been shown to be highly representative in breast cancer 46 . Furthermore, agreement between TMA assessment and clinical records has been shown to be high 47 . Fourthly, digital mammography was implemented in 2004, and therefore both analogue and digital images were included in this study. However, this has been shown not to influence screening performance 48 . Finally, a wide range of mammographic appearances is used in the literature 6,10,26 , which leads to difficulties when comparing studies. Also, we only considered the most dominant appearance, while others have considered combinations of appearances. Nevertheless, we believe our results to be generalizable, as spiculated appearance is often treated as a separate entity in the literature. This study confirms the clinical notion that breast tumours with a spiculated appearance on mammography have more favourable characteristics and hence the prognosis is potentially better. Survival was not an endpoint in this study, but is planned for a future study. Overdiagnosis of slow-growing tumours that would perhaps not have led to breast cancer morbidity or mortality is a well-known issue in breast cancer screening. In this study, spiculated tumours were more frequent among screening-detected than clinically detected tumours, which may indicate that some of these tumours were so-called over-diagnosed cancers.
In conclusion, this study provides strong statistical evidence of several associations between mammographic tumour appearance and clinicopathological factors, including molecular subtypes. In particular, the results consistently indicate favourable characteristics of spiculated tumours. Defining associations between the mammographic tumour appearance and the clinicopathological outcome may aid in characterizing breast cancer already from the initial mammogram, which could potentially contribute to more individualized breast cancer treatment in the future.

Data availability
The data supporting the findings of this study are available from the Malmö Cohorts. However, restrictions apply to the availability of the data, which were used under license for the current study, and are not publicly available. For more information visit the Malmö Cohorts webpage: https ://www.malmo -kohor ter.lu.se/Engli sh.