Breast Glandularity Distribution and Refining the Mean Glandular Dose Estimates in Digital Mammography

Abstract

Background:

In mammography, radiation dose is typically expressed as the mean glandular dose (MGD), which represents the dose delivered to the glandular tissue of the breast.

Materials and Methods:

This study compares MGD estimates obtained using three different methodologies: (I) MGD-Dance-Laboratory for Individualized Breast Radiodensity Assessment (LIBRA) – Calculated manually for each patient using Dance’s formula, incorporating mammographic breast density values derived from the LIBRA application, thereby replacing Dance’s standard glandularity assumption with image-specific values; (II) MGD-Dance – Calculated using Dance’s formula with the conventional assumption of 50% glandularity; (III) MGD-Displayed – Extracted directly from the Digital Imaging and Communication in Medicine header of each mammogram.

Results:

A total of 688 anonymized mammograms from 172 women undergoing routine screening were analyzed, with complete technical and patient-related data. The mean MGD values obtained by the three methods were: MGD-Dance-LIBRA: 2.97 mGy; MGD-Dance: 2.78 mGy; and MGD-Displayed: 2.81 mGy. The average glandularity across the dataset was estimated at 14%. A strong correlation was observed between MGD-Dance and MGD-Dance-LIBRA values (R² =0.9865). The refined dose estimation using image-specific glandularity from LIBRA consistently produced slightly higher values compared to the standard Dance method, highlighting the impact of the commonly assumed 50% glandularity, which overestimates the true average density.

Conclusions:

Incorporating individualized breast density estimates from the LIBRA application into Dance’s formula provides a more refined and accurate method for calculating MGD in digital mammography.

INTRODUCTION

The global landscape of mammography has witnessed a marked stream in utilization, reflecting its fundamental role in breast cancer screening. Exposing healthy women to ionizing radiation, however, is associated with a risk of inducing breast cancer. Therefore, the dose to the breast must be kept as low as reasonably achievable.[1]

Radiation dose in mammography is usually expressed as the mean glandular dose (MGD) or average glandular dose (AGD).[2,3] Acceptable and achievable dose levels in breast cancer screening, as stipulated in the European guidelines on breast cancer screening, are also expressed as AGDs.[4,5] AGD is also used in dose monitoring and optimization. In modern mammography, the AGD is automatically calculated for each exposure and displayed to the operator, as well as stored within a header of a digital imaging and communication in medicine (DICOM) file. The information may be gathered by dose management systems, allowing further analysis.[6]

The AGD is calculated using conversion factors from air kerma to AGD derived from Monte Carlo simulations.[7,8,9,10,11,12] It depends on X-ray spectra, beam quality, breast thickness, and breast density (glandularity in %). Breast glandularity is defined as the proportion of glandular (fibroglandular) tissue and the whole breast tissue, also including adipose (fatty) tissue. It is an important concept in mammography because glandular tissue is more radiation-sensitive compared to adipose tissue.

Several different methodologies of MGD calculation have been so far utilized in various parts of the world,[13] such as the European guidelines on breast cancer screening[4,5] and the International Atomic Energy Agency recommendations[14] that endorse the methodology described by Dance et al.[7,8,9]

Most MGD (AGD) estimates rely on the assumption that the breast consists of 50% glandular tissue and 50% fatty tissue.[7,10,11,12] However, it is well known that breast composition differs among women, but it also depends on the patient’s age. Therefore, the assumption of 50% glandularity leads to only a rough estimate of the absorbed dose. Some authors have addressed this issue by developing a model to estimate breast glandularity based on the patient’s age and compressed breast thickness (CBT).[8,9] However, even this approach has limitations, as glandularity is not assessed by the actual breast composition. This is a critical point in mammography, since the amount of fibroglandular tissue in the breast is closely related to the risk of radiation-induced breast cancer: women with dense breast tissue have a higher risk of developing radiation-induced breast cancer, compared to women with less dense breasts.[15,16]

Mammographic breast density (MBD) (also called glandularity) is a fraction of the fibroglandular tissue of the breast as seen by the X-ray attenuation patterns on a mammogram. A software named laboratory for individualized breast radiodensity assessment (LIBRA) is an algorithm designed to estimate the MBD. The estimation process uses post-processed images, working either individual or batch processing, allowing a low-cost in man-power alternative. Being an open-source package available in the MATLAB environment (the MathWorks, Inc., Natick, MA, USA),[17] LIBRA has so far been validated for GE and Hologic digital mammography systems.[18]

The present study was conducted with one overarching objective to improve the accuracy of MGD estimation in mammography by addressing the limitations of using a fixed 50% breast glandularity assumption. Herein, we present how the evaluation of the patient’s specific breast glandulatity in %, derived from the LIBRA application into Dance’s formula, improves the accuracy in the calculation of the mean glandular dose, allowing an individualized MGD approach, thereby enhancing patient-specific dose assessment in mammography.

MATERIALS AND METHODS

The study was conducted on the Hologic Selenia Dimensions mammography system, the fully digital X-ray mammography machine installed in 2014 and operated in automatic exposure control mode.

Patient-related information, such as mammographic projections, age, and breast thickness, exposure parameters, was exported from the DICOM image header to MS Excel using the open-source application “Micro DICOM Viewer”[19] such as: age, screening date, CBT, presence of implants, view, laterality, tube voltage (kVp), tube current exposure time product (mAs), target material, filter material, exposure control mode, organ dose, detector ID, and mammographic unit model. Medical physics annual reports were also obtained from participating hospital, as the calculation of MGD requires these data. The data were verified against the internal weekly calibration and the annual quality control tests, executed by the Directorate of Radiation Safety of the Republic of North Macedonia.

The present analysis involved 688 mammograms from 172 women undergoing screening mammography procedures in the year 2024 in a public hospital’s Mammography “Unit H.” The screening program covered women aged 40–64 years. The data were stratified by age (40–49 and 50–64) and the CBT (range 20–110 in 10 mm increments).

In the present work, the MGD will be either extracted from DICOM or calculated using two different methods:

• (1)by the original Dance’s methodology,[8] herein called ’MGD Dance’;
• (2)by modified Dance’s methodology with the glandularity values derived from the LIBRA
• (3)application, herein named “MGD Dance (LIBRA);” from the DICOM header extracted MGD values, which are herein named ’MGD Display’.

The tube output (mGy/mAs) and the half-value layer (HVL) for a range of beam parameters (target/filter/kVp) are measured by the annual technical quality control Directorate of Radiation Safety of the Republic of North Macedonia, and these values have been applied when estimating the MGD.

Calculation of mean glandular dose “MGD dance”

The Danse’s equation used is as follows:

MGD = K∙g∙c∙s      (1)

K is the Incident air kerma at the surface of the breast without backscatter.

g is the conversion factor for a breast model with 50% glandular tissue. This factor is calculated by interpolating HVL and CBT.[8]

c is a factor that corrects for any difference in breast composition diverging from 50% glandularity, defined for two age groups: 40–49 and 50–64 years. This factor is a function of HVL and CBT.[8]

Herein, s is the factor that depends on different target/filter combinations.[8]

These conversion factors are based on Monte Carlo simulations.[7,8,9]

Furthermore, the data collected allowed the estimation of the Incident air kerma (K) at the upper surface of the breast, using the CBT, the focus-detector distance, and the focus support breast distance. The calculation of MGD was made by a previous calculation of K using parameters that define the beam quality and beam output, provided from the State Directorate for Radiation Safety of the Republic of North Macedonia, measured semi-annually. The Hologic mammography system employed a tungsten (W) anode with either a rhodium (Rh) or silver (Ag) filter. The anode/filter combination was selected automatically by the system based on the detected CBT. For CBT values below 70 mm, the W/Rh combination was used, whereas the W/Ag combination was applied for CBT values exceeding 70 mm or during magnification views.

Variations in beam quality and spectral output have a direct influence on the accuracy of MGD estimations. Having in mind that the MGD is calculated from the incident air kerma (K), the X-ray spectrum, and the corresponding glandular dose conversion factors, one should take into account that all these factors depend strongly on the beam quality defined by the anode/filter combination and tube voltage (kVp). As a result, MGD depends on the beam quality.

Refined calculation of mean glandular dose “MGD dance (Laboratory for Individualised Breast Radiodensity Assessment)”

To calculate MGD Dance (LIBRA) the following refinements were undertaken:

• (a)Original c-factor was replaced by the one estimated through the interpolation multivariable fitting process of HVL, CBT, and

The glandularity was estimated with LIBRA algorithm, using the discrete values,[8] and the scipy. Interpolate in Python.

Both the previously elaborated MGD calculation methods “MGD Dance” and “MGD Danse (LIBRA)” take in consideration the output beam factors. These two MGDs differ only the c-factor – the second one taking into calculation the estimated gladularityand the second one taking 50% gladularity for all brests. Hence, the heterohenity in anode/filter selection should not influence the correlation between MGD-Dance and MGD-Dance-LIBRA estimates, presented in our results below.

Estimation of the breast glandularity in %

Breast density (glanularity), expressed as a percentage (%), is the fraction of fibroglandular (dense) tissue from the total breast tissue in %. There are two quantitative methods for breast density estimation, volumetric (three-dimensional) and area-based (two-dimensional [2D]). Area-based calculations use digital mammograms, both the medio-lateral oblique and cranio-caudal projections, to calculate breast density. It involves the determination of the total 2D breast area using the pectoral muscle as a boundary and the air surrounding the breast, and segmenting the breast image to separate regions of dense tissue from the fatty tissue, and measuring the area of the dense parts, where density is mathematically defined using the grey level definition. It is calculated as:

For academic purposes, the area-based segmentation and calculation are often performed using image processing software by ImageJ, MATLAB, and LIBRA.[17,18]

In our calculations, we used LIBRA as an algorithm for fully automated breast percentage density quantification, performed for both raw and processed digital mammography images. The body‐air interface boundary is determined by a threshold based on the grey‐level intensity histogram, independent of any prior assumptions. The boundary between the pectoral muscle and breast tissue areas uses a previously validated algorithm based on a straight line Hough transform.[17] In this study, we used LIBRA version 1.0.4, provided by the Perelman School of Medicine, University of Pennsylvania.[18]

In the present study, the images that included both projections (cranio-caudal and medio-lateral oblique) were analyzed with LIBRA software using the anonymized database, issued by the public hospital with a coded name “Unit H.” The data were downloaded patient by patient from the Picture Archiving and Communication System located at the respective mammography facility. Statistical analyses were performed using the Statistical Package for the Social Sciences (SPSS) IBM (Chicago, USA), 2009 Version 25 for Windows.

RESULTS AND DISCUSSION

Two parameters were used to form the subgroups from the total of 688 mammograms from 172 women: age (40–49 and 50–64), and CBT (20 mm – 110 mm, in 10 mm increments) for each age group.

Figure 1 illustrates the operation of the LIBRA segmentation algorithm. Figure 1a shows one example of a DICOM image from the screening. Figure 1b presents the coloration of the regions with the same level of grey (level of attenuation), with the hottest regions being black. Figure 1c shows the green contours of segmented regions with density corresponding to glandular tissue within the breast.

Figure 1.

Laboratory for Individualized Breast Radiodensity Assessment segmentation algorithm. (a) The mammogram’s original digital imaging and communication in medicine image; (b) a clustering that represents an area of the image with grey-level intensity responding to similar X-ray attenuation qualities; and (c) the final dense-tissues output image after segmentation

The breast density (glandularity) in % was evaluated by LIBRA for each one of the 688 mammography images separately. Figure 2 presents the decrease of the mean glandularity with the increase of CBT for the two age groups (40–49 and 50–64) with the CBT increasing from 20 mm to 110 mm, with an increment of 10 mm. The trends of the curves are modelled with a polynomial fitting, showing decrease of the glandularity with the breast thickness. From Figure 2, it also appears that the age dependence of the gladularity remains inconclusive. The dark blue line shows the resulting mean gladularity for all (the two age groups jointly). From here, it is evident that the breast density for the thinnest breasts (CBT = 20 mm) is about 33% and gradually falls to about 5% for CBT of 90 mm. In other words, as breast thickness increases, the percentage of glandular tissue decreases.

Figure 2.

Mean glandularity (breast density) vs. compressed breast thickness estimated from 688 mammograms using Laboratory for Individualized Breast Radiodensity Assessment[8] for two age groups, (40–49) and (50–64), and for jointly for all ages (dark blue line)

The fitting made by the least squares method shows that the glandularity in % declines with the CBT following a polynomial function:

Glandularity(%)=0.3389(CBT)²-7.2063(CBT)+39.492

In Figure 3, we present the left-skewed distribution of the frequency of occurrence of a MBD within a certain glandularity range with an increment of 1%, derived from the row data from LIBRA. Herein, the data were fitted to a Gaussian function using the least-squares method. The Gaussian peak in the glandularity (fitted maximum) occurs at 8.5%. The median and the mean breast gladulalrity (in %) within the examined population of 172 women (688 images) were evaluated to be 11% and 14%, respectively.

Figure 3.

Frequency of occurrence of certain glandularity for a total of 688 mammography images from 172 women (left-skewed distribution)

In Figure 4, we present the age distribution of the mean glandularity in % in 1-year increments (left), in 5-year increments (right). It appears that the mean gladularity in % is strongly correlated with the age (R² = 0.95). The distribution implies that breast density declines with age, leading to a reduced proportion of glandular tissue and a comparatively higher proportion of fatty tissue. It has to be noted that for very young patients with higher CBT, one needs to have an individualized approach. Furthermore, the “s” and “c” factors may vary depending on the target/filter combination used and glandularity, respectively. For this reason, the error (described with the error bars in Figure 4 (left) may be higher than it appears. While the average density of the women of age 40–44 is about 20%, it falls gradually to about 8% for women of age 60–64. It is also evident that the results for the latter age group show extremely high error (wide confidence intervals). These results were found to be close to the previously reported values (8%–13%).[20]

Figure 4.

Mean mammographic breast density distribution with age: (left) in 1 year age intervals. Error bars represent the 95% confidence intervals; (right) in 5 year age intervals with linear regression fitting line

Table 1 provides a descriptive statistics summary of the data set, including the minimum, maximum, 1^st and 3^rd quartiles, median, mean, and standard deviation for CBT, age, MBD, and the three calculated/extracted values of MGD: (I) MGD Dance, (II) MGD Dance (LIBRA), and MGD Displayed, defined previously. The results demonstrate that the mean MGD Displayed provided slightly higher values than the calculated MGD Dance, whereas it showed lower values compared to MGD Dance (LIBRA).

Table 1.

Statistical description of the included dataset and the three mean glandular dose values calculated from 688 mammograms

Statistic	CBT (mm)	Patient age (year)	MBD (%)	MGDdance (mGy)	MGD dance (LIBRA) (mGy)	MGD displayed (mGy)
Minimum	20	40	1.60	0.24	0.29	0.29
Maximum	104	64	61.90	6.92	7.06	6.62
1st quartile	52	45	6.49	1.90	2.10	1.98
Median	60	47	11.01	2.62	2.86	2.70
Mean	59	48.72	14.38	2.78	2.97	2.81
3rd quartile	68	52	19.89	3.54	3.74	3.46
SD	4.95	1.41	10.60	1.23	1.24	1.13

SD: Standard deviation, CBT: Compressed breast thickness, MBD: Mammographic breast density, MGD: Mean glandular dose, LIBRA: Laboratory for Individualized Breast Radiodensity Assessment

Figure 5 presents a comparative scatter plot of MGD Dance (green marks), MGD Dance (LIBRA) (blue marks), and MGD Displayed (orange marks), versus the CBT, evaluated for all 688 mammography images (both projections). As is evident, the MGD Dance (LIBRA) followed a similar trend to the MGD displayed. Box plots presented on the right side of Figure 5 show the minimum dose (mGy), maximum dose (mGy), median dose (mGy), 25^th percentile, and 75^th percentile of MGD. Both the mean and median appeared slightly higher when calculated by Danse (LIBRA). From here, it follows that the MGD Displayed dose in the DICOM and the MGD calculated by Danse’s algorithm both underestimate the dose absorbed by the glandular tissue in the breasts during mammography.

Figure 5.

Scatter plot of the dependence of mean glandular dose (MGD) dance (laboratory for individualized breast radiodensity assessment), MGD displayed, and MGD dance on the compressed breast thickness. MGD: Mean glandular dose, LIBRA: Laboratory for individualized breast radiodensity assessment

Figure 6 presents the linear regression between MGD Dance (LIBRA) versus MGD Dance (left) and MGD Dance (LIBRA) versus MGD Display, showing a high correlation between them (R² = 0.99), meaning that about 99% of the variation in the calculated MGD can be explained by all the models. The slope in both charts is about 0.94, meaning that the MGD values calculated by Danse (LIBRA) methodology yield greater doses for about 6% (slopes are both about 0.94) compared to the other two. As evident from both the Figure 6, the linear regression becomes divergent for higher MGD doses.

Figure 6.

Linear regression between the mean glandular dose (MGD) values obtained by two different models/methods: (left) MGD Dance versus MGD Dance LIBRA, (right) MGD Display versus MGD Danse LIBRA

As evident from the statistical chart in Figure 5 (right), there are slight differences between the mean values of: (1) calculated MGD Dance (LIBRA) (2.97 mGy), (2) Extracted MGD Displayed (2.81 mGy), and (3) calculated MGD Dance (2.78 mGy). Furthermore, the results of the correlation among the three MGD datasets are presented in the form of Pearson correlation, which indicates that there is a significant, large positive relationship between MGD Dance (LIBRA) and MGD Displayed (r = 0.975, P < 0.001). Furthermore, the results of the Pearson correlation indicated that there is a significant, large positive relationship between MGD Dance (LIBRA) and MGD Dance (r = 0.994, P < 0.001) and between MGD Dance and MGD Displayed (r = 0.978, P < 0.001).

The good agreement between the MGD Dance LIBRA and MGD Displayed from our work also aligns well with previous findings by Borzì et al.,[21] who demonstrated that the MGD values displayed by the systems were generally consistent with those calculated using Dance’s method, where breast glandularity is estimated through LIBRA.

MGD is significantly dependent on the evaluation method of the breast glandularity. It is essential to account for the heterogeneous distribution of glandular tissue when estimating MGD, since areas with varying glandular density can significantly influence the results. However, using the refined Dance’s method with the following refinements: (a) original c factor replaced by the one estimated through the interpolation multivariable fitting process of HVL, CBT, and (b) the breast glandularity (%) estimated with the LIBRA algorithm, yielded higher dose values (MGDs). Hence, the higher dose values of the MGD Danse (LIBRA) and their implementation in the MGD calculations show a step forward toward refining the individual dose calculation, giving higher but more accurate results compared to the other two (MGD Danse and MGD Displayed). The latter two methods obviously slightly underestimate the delivered dose to the glandular tissue by about 6% in comparison to the Dance (LIBRA) method.

The afore-presented Danse’s LIBRA individualized dose approach in mammography is biologically relevant because radiation sensitivity varies between patients due to differences in breast composition, age, and genetic susceptibility. For these reasons, tailoring the radiation dose for each individual minimizes unnecessary exposure of radiosensitive tissues and thus minimizing the biological effects, while maintaining diagnostic image quality.

CONCLUSIONS

Despite the limitations of the present study – namely, the analysis of mammographic images from a single manufacturer and a dataset comprising 688 images from 172 women – the following conclusions can be drawn:

As CBT increases, the proportion of glandular tissue relative to adipose tissue within the breast decreases. In other words, the thicker the breast, the less glandular tissue it contains.

The distribution of glandularity (%) in the study population follows a Gaussian pattern, with a peak (fitted maximum) at 8.5%. The median and mean breast densities across the 172 women (688 images) were 11% and 14%, respectively. These values differ substantially from the 50% glandularity commonly assumed in MGD calculations according to Dance’s method.[8,9]

Small differences were observed between the average MGD values obtained using the following methods: MGD calculated through Dance’s method using LIBRA-derived glandularity – 2.97 mGy; MGD as displayed on the mammography system – 2.81 mGy; MGD calculated through standard Dance’s method (without LIBRA input) – 2.78 mGy.

The slightly higher dose values obtained with the Dance (LIBRA) method indicate a refinement in dose estimation, yielding results that are likely more accurate. In contrast, the standard Dance method and system-displayed MGD underestimates the actual dose delivered to glandular tissue by approximately 6%.

Estimating MGD using Dance’s formula, when incorporating glandularity values derived from the LIBRA application, results in higher and more precise dose estimates. This approach provides a refined method for more accurately assessing individual radiation dose delivered to the glandular breast tissue in mammography procedures.

Conflicts of interest

There are no conflicts of interest.

Funding Statement

Nil.