Breast cancer is among the most prevalent cancers in adult women of any age, but hopefully, it can be screened or diagnosed early in the majority of cases. The importance of early diagnosis has led to several screening guidelines, with mammography as the leading procedure.¹ Reportedly, high mammographic density, due to the high proportion of epithelial-stromal tissue and complex structure of the glands, is an independent risk factor for breast cancer incidence with an anticipated relative risk ranged from 1.8 to 6.0.^2-5 The concealing effect of the dense background causing reduced sensitivity for detection of precancerous and cancerous lesions in mammography makes the owner prone to cancers appearing at screening intervals.^5-7

Since the higher incidence of cancer does not attenuate by time or repeated screening, another possibility is the higher volume of ducts or a more suitable microenvironment for tumor growth where the cancer originates, causing higher cancer incidence.^8-11 The current 5th edition of the Breast Imaging-Reporting and Data System (BI-RADS) criteria for categorizing breast density has replaced estimates of dense area percentage with descriptions about the possibility of obscured lesions¹², which may reflect information related to different breast compositions.

Several computer algorithms have been designed to extract texture features and convert them into higher-dimensional data. For instance, Support Vector Machine (SVM) is a type of automated binary machine learning algorithm that has been used to distinguish and classify radiological patterns of mammograms. Their potential clinical implication is to improve accuracy, reduce intra- and inter-observer variability, and support decision-making by the radiologists.^13,14

The first-order statistical features (FOSFs) of the intensity histogram of mammograms may be considered as radiomic features that describe the distribution of intensity voxels by basic metrics. FOSFs include parameters that provide information beyond the point of human visual perception and are possibly related to cancer incidence.^15-19 Due to lower accuracy of the mammogram and higher susceptibility to cancer incidence in dense breasts, updated screening recommendations endorse early risk assessment and supplemental screening if the risk is considered higher-than-average.²⁰

While basic mammographic features are readily extractable, their potential correlation with pathological or radiological findings offers opportunities for enhancing computer-aided diagnostic and prognostic models. This study aims to investigate the relation between objective (measured) and subjective (radiologist’s interpretation) assessments of breast density, investigate the relationship between mammographic FOSFs and density, and explore the potential of FOSFs as covariates in modeling the association between tumor characteristics and pathology.

In this study, data from 85 patients were analyzed who all had unifocal and unilateral lesions of in situ or invasive carcinoma. The age range of patients was 36 to 79 years. Ninety percent of patients had a history of childbirth, with an average of 3 children; the age at first pregnancy ranged from 15 to 37. A family history of breast cancer in first- or second-degree relatives was present in 21%. A history of regular intake of OCP was reported in 46%. None of the menopausal patients had a history of HRT. Mammogram data for tumor size were available for only 20 patients. Sixteen patients were T1, two were T2, and two others were T3 according to tumor size. Other demographic information is listed in Table 1. There was statistically significant variation between categories of breast density groups in the means of PD after adjusting for age and breast area, determined by one-way ANCOVA. Cases of different density categories were not significantly different for the kurtosis, mean, standard deviation, and minimum measure of their normalized intensity histograms. Other histogram FOSFs, including median, interquartile range, maximum, skewness, and energy, were significantly different in breast density categories (Table 2).

The findings of this study showed some associations between demographic, radiological, and pathological features of the breast composition and the tumor. PD significantly correlated with most FOSFs of the mammogram, and FOSFs of histograms used by the SVM algorithm, besides PD, could discriminate between a or d BI-RADS density groups from others with acceptable accuracy. Also, the spiculated border of the tumor could predict higher PR percentage positivity when the images were adjusted for age, area of the breasts, and FOSFs.

The relationship between age and PD or FOSFs suggests an effect of breast texture change by age. As it has been studied extensively11,21, breast composition is affected by factors such as age and body mass index (BMI), hormone intervention, growth factors, and nutrition. Although genetic factors may explain the majority of the etiology of breast density, only a few genes have been detected so far, whose proliferative activity is linked to breast dense composition and carcinogenesis as well. The acquired factors are estimated to explain 20-30% of the etiology of breast density, and some are shared with cancer etiology,11

Our findings did not suggest any link between the hormone receptor positivity of cancer and dense breasts. Previous studies have inconsistently indicated that high breast density seems to increase the risk of all breast cancer subtypes16,23, but several studies indicate a stronger link between hormone-positive breast cancers and breast density. A number of normal breast epithelial cells express PR and ER, which are involved in growth and proliferation. Blocking hormone receptors can reduce both cancer incidence and breast density. Moreover, there is evidence that at least some microenvironmental factors, such as collagen type 1, small leucine-rich proteoglycans, high expression of DNA damage response genes, and downregulation of CD36, are etiologies shared for breast density and cancer24 many of which are more or less thought to be mediated by hormone receptors.8-10

Patients who had HER2-overexpressing tumors in our study were younger. Accordingly, some epidemiologic studies have shown that HER2-positive tumors, like other aggressive histologies, have a higher prevalence in young women in comparison with older patients25,26, but evidence is limited, in particular, across various ethnicities. Another finding of this study was that HER2 overexpression had a moderate effect on the tumor margin, and tumor spiculation could predict a higher percentage of PR expression. Manifestations of different molecular subtypes of breast tumors have been radiologically described. Our findings from the logistic regression were in agreement with those studies that associate margin spiculation with positivity of hormone receptor markers. Similar studies have evaluated this relation, in which different cut-off values of IHC percentage were used to define the positivity of hormone receptors.27-31 Luminal type tumors are likely to form desmoplastic interactions with the surrounding stroma that cause the stretch of the Cooper’s ligaments, causing margin spiculation.27,32 Conversely, the frequency of spiculated margins is less in HER2 overexpressed32 or triple-negative tumors.28 Still HER2-positive tumors usually present with spiculated margins, while triple-negative tumors are more likely to present with round borders resembling benign masses.33 Abundant PR is extremely rare when ER is negative, so most PR positive tumors express ER as well. PR is a predictive marker, and the absence of PR receptors, even in the presence of ER, is detrimental to survival.34 Most invasive carcinomas express PRA isoform dominancy, which is responsible for invasiveness via cross-signaling with ER.35,36 Therefore, the linkage of tumor features and PR can provide insights into the prognosis of the patient.

We assume that FOSFs extracted by the algorithms are reproducible and operator-independent, and the data of the combined histograms from all four mammographic views primarily represent the composition of the uncancerous breast tissue.11 Mammograms with different textures in the normal population may show different susceptibility to cancer incidence. A study on screening mammograms indicated that a more uniform parenchyma and a higher density percentage were each independently proportional to the incidence of cancer. In a study, an unsupervised clustering could accurately stratify mammograms into four phenotypes based on radiomic complexity scores and density, in which subjects who were diagnosed with breast cancer had higher low-complexity and low- to intermediate- complexity phenotypes and those with very dense breasts were mostly in the minimum complexity group, so complexity was not directly related to density.15,18 Some studies attempted to find specific prognostic models to predict the incidence of ER-positive cancer subtypes, which have been linked to breast dense composition.16,17 A case-control study indicated that a variance measure of mammographic histograms could act as a possible indicator of breast density with a stronger relationship to breast cancer incidence than the calibrated ‘percent glandular’.19 Also, a larger study indicated that manual and automated percentage of breast tissue density, as well as variance, are independently related to breast cancer.37 The evidence may reflect that carcinogenesis is preceded by changes of breast composition parameters.

Except for mean, minimum, and kurtosis, which were supposed to be almost uniform in the standardized mammographic histograms, mean values of interquartile range, skewness, median, maximum, and energy of the images were diverse across the categories of breast density. These variables were not accounted for in the standard calculations of breast density but were representative of texture variability. The FOSFs, besides PD, could almost accurately discriminate BI-RADS density groups, especially for categories a and d, by the SVM algorithm. SVM accuracy of over 80-90% for discrimination of categories a and d of breast density indicates that the FOSFs, besides PD, are generally informative and good at discriminating each of these categories from others. Although these features exhibited strong correlations (Table 5), the rank feature function demonstrated that their contributions to distinguishing between density categories are not equivalent.

For instance, energy, a variable not so important for discriminating categories, correlated negatively with PD, and had a significant negative correlation with the minimum intensity as well, and a significant positive correlation with maximum and standard deviation. The sum of squared pixel raw intensities was used to calculate energy, while PD was calculated from the dense extreme end of the histogram. In other words, most of the energy reflects the fibro-glandular tissue in the mid-range of the histogram (Figure 1).

Our study used the most basic features, none of which accounted for the local variation of the image. This may explain the low accuracy of SVM classification for discriminating categories b or c. Using more complex features, this function was able to accurately classify mammograms according to their breast density using gray level co-occurrence matrices texture features.38 One of the limitations of the current study was the retrospective nature of the data, small sample size, and reliance on existing documented findings.

Given the distribution of pathologic subtypes, studies with higher power in specific ethnic populations are required to enable analyses of subgroups. Despite high concordance of hormone markers in needle and surgical samples39, future studies may assess the reevaluation of negative hormone receptor stains or weak HER2-positive cases. The relationship between the radiological and morphological basis of the breast and tumor and PR positivity can be the subject of future epidemiologic studies.

The linkage between findings of imaging and pathology, although speculative at this stage, could potentially have implications for understanding the pathogenesis of breast cancer. Moreover, it may inform future studies, particularly about dense breasts, in which the masking effect and higher risk of cancer may need enhanced filtered images or customized prediction models.

The study protocol was reviewed and approved by the Ethics Committee of Iran University of Medical Sciences. The approval was granted under the ethics code IR.IUMS.FMD.REC.1400.112, in accordance with institutional guidelines and national ethical standards for research involving human participants.