Calibrated Breast Density Measurements

Abstract

Rationale and Objectives:

Mammographic density is an important risk factor for breast cancer, but translation to the clinic requires assurance that prior work based on mammography is applicable to current technologies. The purpose of this work is to evaluate whether a calibration methodology developed previously produces breast density metrics predictive of breast cancer risk when applied to a case-control study.

Materials and Methods:

A matched case control study (n = 319 pairs) was used to evaluate two calibrated measures of breast density. Two-dimensional (2D) mammograms were acquired from six Hologic mammography units: three conventional Selenia 2D full field digital mammography systems and three Dimensions digital breast tomosynthesis systems. We evaluated the capability of two calibrated breast density measures to quantify breast cancer risk: the mean (PG_m) and standard deviation (PG_sd) of the calibrated pixels. Matching variables included age, hormone replacement therapy usage/duration, screening history, and mammography unit. Calibrated measures were compared with the percentage of breast density (PD) determined with the operator-assisted Cumulus method. Conditional logistic regression was used to generate odds ratios (ORs) from continuous and quartile (Q) models with 95% confidence intervals (CIs). The area under the receiver operating characteristic curve (Az) was used as a comparison metric. Both univariate model and models adjusted for body mass index and ethnicity were evaluated.

Results:

In adjusted models, both PG_sd and PD were statistically significantly associated with breast cancer with similar Az of 0.61 to 0.62. The corresponding ORs and CIs were also similar. For PG_sd, the OR was 1.34 (1.09, 1.66) for the continuous measure and 1.83 (1.11, 3.02), 2.19 (1.28, 3.73), and 2.20 (1.26, 3.85) for Q2-Q4. For PD, the OR was 1.43 (1.16, 1.76) for the continuous measure and 0.84 (0.52, 1.38), 1.96 (1.19, 3.23), and 2.27 (1.29, 4.00) for Q2-Q4. The results for PG_m were slightly attenuated and not statistically significant. The OR was 1.22 (0.99, 1.51) with Az = 0.60 for the continuous measure and 1.24 (0.78, 1.97), 0.98 (0.60, 1.61) and 1.26, (0.77, 2.07) for Q2-Q4 with Az = 0.60.

Conclusion:

The calibrated PG_sd measure provided significant associations with breast cancer comparable to those given by PD. The calibrated PG_m performed slightly worse. These findings indicate that the calibration approach developed previously replicates under more general conditions.

1. Introduction

A high level of breast density assessed from two-dimensional (2D) mammograms is a strong breast cancer risk factor (1–6). There are various methods of measuring breast density (3, 7–11) yielding qualitatively similar findings. An accurate and reliable breast density measurement that has validated association with breast cancer risk is a necessary first step to translate the research findings to date to point of care clinical management and breast cancer risk prediction.

According to the American Cancer Society, over 300,000 women will be diagnosed with breast cancer in the US this year resulting in approximately 40,000 deaths. Benefits of using breast cancer risk assessment as a guide for personalized screening strategies are well documented in the research setting (12). In both the US and western Europe, breast cancer cure rates have improved (around 75%), but treatment costs are considerable indicating there is not only a need to predict who will develop breast cancer, but also emphasize the importance of lifestyle changes and drug interventions to prevent this disease (13). Maximizing the benefits of breast screening, while minimizing the risks, requires abandoning the one-size-fits all approach and moving to personalized breast screening regimens (14). Unfortunately, risk models used clinically are not accurate at the individual level (15), breast density is not routinely considered by healthcare providers as a risk factor (16), and is primarily confined to risk-related research studies. To date, there is not a recognized standard for breast density determination for risk prediction outside of the research environment (17).

Translating a breast density measure to point of care requires a metric that is uniform and consistent across imaging platforms both spatially and temporary. To achieve these ends, our focus in this report is evaluating calibrated measures of breast density. The calibration technique used in this report relies on establishing references by imaging breast tissue equivalent (BTE) phantoms. Calibration compensates for the x-ray image acquisition technique influences on the pixel representation, producing inter-image normalization. To date, findings from phantom based calibration with two-dimensional (2D) mammography have yielded mixed associations with breast cancer (7, 8, 10, 18–20).

The calibration technique used in this report is an outgrowth of earlier work by Kaufhold et al. (21). Heine et al extended this calibration work initially with phantom studies using a specific indirect x-ray detection full field digital mammography unit (22–25) that produced several calibrated breast density metrics predictive of breast cancer (18, 19, 26, 27). Subsequently, this calibration approach was modified to operate on direct x-ray conversion FFDM units and evaluated with phantom studies (26, 28, 29); this approach was extended to patient mammograms for risk prediction purposes in this report using a matched case-control study.

2. Materials and Methods

2.1. Population and Imaging

The study consists of 319 individually matched pairs of adult women that attended the breast clinics within this Cancer Center. Cases are either (i) women attending the breast clinics at this Center diagnosed with first time unilateral breast cancer (referred to as internal patients), or (ii) attendees from surrounding area clinics recently diagnosed with first time unilateral breast cancer, sent to this Center for treatment or further evaluation (referred to as external patients). All cases have pathology verified breast cancer. Controls were attendees of this Center without a history of breast cancer. Controls were individually matched to cases on age (± 2 years), hormone replacement therapy (HRT), screening history; and mammography unit. The HRT match was based on status of current users or non-users. Non-users included women that have not taken HRT for at least two years. If a case was a current HRT user, the control was matched on duration (± 2 years). Controls were matched by screening history using a three-category classification. Group 1 included women with prior screening history by any means; the duration between the last screening and the study image date must be no more 30 months apart. Group 2 included women with a screening history that do not fit within Group 1 or Group 3. Group 3 included women with no screening history. The population breakdown is provided in Table 1.

Table 1a.

Population Characteristics: This table describes the population characteristics by either distribution mean for a given measure or percentages of the population. Where applicable, the standard deviation of the respective distribution is provided parenthetically.

Measure	p-values	Case n	Case Mean (standard deviation) or relative frequency	Control n	Control Mean (standard deviation) or relative frequency	Total n	Total Mean (standard deviation) or relative frequency
Age	0.13	319	58.8 (11.3)	319	58.7 (11.3)	638	58.8 (11.3)
Race
Caucasian	0.91	273	85.6%	271	85.0%	544	85.3%
African-American	0.24	26	8.2%	36	11.3%	62	9.7%
Asian	1.00	8	2.5%	8	2.5%	16	2.5%
More than One	0.63	3	1.0%	1	0.3%	4	0.6%
Other	N/A	3	1.0%	0	0.0%	3	0.5%
Unknown	0.51	6	1.9%	3	0.9%	9	1.4%
Ethnicity
Non-Hispanic	<0.001	287	90.0%	258	80.9%	545	85.4%
Hispanic	<0.001	29	9.1%	59	18.5%	88	13.8%
Unknown	1.00	3	0.9%	2	0.6%	5	0.8%
BMI	0.07	319	28.7 (6.0)	314	27.8 (6.5)	632	28.3 (6.3)
Screening Group	N/A
Group 1		224	70.2%	224	70.22%	448	70.22%
Group 2		58	18.9%	58	18.2%	116	18.2%
Group 3		37	11.6%	37	11.6%	74	11.6%
HRT Usage	N/A
Current		18	5.6%	18	5.6%	36	5.6%
Not Currently		301	94.4%	301	94.4%	602	94.4%
MS	0.46
Pre-Menopausal		79	24.8%	73	22.9%	152	23.8%
Menopausal		240	75.2%	246	77.1%	486	76.9%
PD	0.05	319	25.3 (11.6)	319	23.6 (11.8)	638	24.5 (11.7)
PGM	0.66	319	25.0 (16.7)	319	24.5 (17.3)	638	24.7 (17.0)
PGSD	0.05	319	9.9 (5.1)	319	9.2 (5.0)	638	9.51 (5.1)

Two-dimensional (2D) study mammograms were acquired from one of six Hologic mammography units: three conventional 2D Selenia FFDM units (H units) and three Dimensions digital breast tomosynthesis (DBT) units (D units). This Center phased out conventional 2D units over the past three years and now performs all breast screening examinations with DBT units operating in the tomosynthesis combination-HD (combo-HD) mode. The 2D FFDM components from the combo-HD mode were used to supplement the dataset derived from the 2D FFDM units. These units have an Automatic Self-adjusting Tilt compression paddle referred to as a FAST paddle^™ (Hologic, Inc., Bedford, MA). In the screening environment, two types of FAST paddles are normally used that differ in size. The choice of paddle is determined by the x-ray technician’s judgment as to which is appropriate for the size of the patient’s breast. The detector field of view (FOV) adjusts automatically when a given paddle is attached: 24cm × 29cm (3328 × 4096 pixels) for the large paddle; and 18 cm × 24 cm (2560 × 3328 pixels) for the small paddle. The detector has a 70μm pitch. Raw data are in monochrome 1 format with 14 bit per pixel dynamic range. Two-dimensional units have either W targets with Rh and Ag filter options (n = 2) or Mo targets with Mo and Rh filter options (n = 1). DBT units operating in the 2D mode have W targets and Rh and Ag filter options (n = 3). Each unit has its own set of calibration curves. We use raw mammograms in cranial caudal (CC) view as study images for calibration purposes. Breast density was determined from the unaffected breast for cases. A control’s study image was matched laterally to its case. For cases, images were acquired before cancer treatment. For the Cumulus operation, we used the clinical display (for presentation) images (12 bit per pixel dynamic range).

2.2. Analyses

We applied a phantom based calibration technique developed by Heine et al. (22, 24–26, 28, 29) to each mammogram producing the percent glandular (PG) representation. The PG representation is normalized for the acquisition technique differences, resulting in an intensity scale that ranges from 0 to 100. In this current calibration procedure, a serial updating scheme was included to ensure that the baseline BTE phantom calibration reference data collected in the past is serially accurate and if not, the baseline is updated (29). This updating methodology requires minimal BTE phantom acquisitions bi-weekly for monitoring accuracy. Two previously validated calibrated breast density metrics (18, 26) were considered: the average (PG_m) and standard deviation (PG_sd) of the pixel values within the usable breast region. Before calibrating a mammogram, an erosion operation was applied in a radial fashion to eliminate a portion of the breast area where the compressed breast thickness is unknown (26). If we consider the outline of the breast background region as a semicircular border, we erode the breast area inward radially by 25% as shown in Figure 1, defining the usable region. This procedure is intended to eliminate (approximate) the portion of the breast area corresponding to where the compressed breast thickness is unknown.

Figure 1.

PG_sd Illustration: The top row shows 4 mammograms in the processed (for presentation) representation used for viewing purposes. The bottom row shows the corresponding calibrated images after erosion. The ordering of the images corresponds with PG_sd quartiles (QRTs) with QRT 1 on the left and QRT 4 on the on the right. The PG_sd values from left to right are provided: 4.1 (PG_m = 20.3); 5.5(PG_m= 17.2); 8.0 (PG_m= 21.0); and 12.3 (PG_m= 52.10). For reference, the corresponding PD values are also provided: 31.4 (quartile 4); 13.2 (quartile 1); 35.0 (quartile 4); and 38.2 (quartile 4).

The Cumulus software (University of Toronto, Toronto Canada) was used to calculate the percentage of breast density (PD), which was implemented by XXX. Cumulus software 3 was used in the batch mode to label clinical display images. This is an operator assisted methodology with a graphical interface. The operator sets the breast border and breast density thresholds. The operator can adjust image contrast as well if needed. Cumulus calculates the percentage of dense pixels. In the batch mode of operation, the user is blinded to final PD result. Cases and controls were randomly mixed and XXX was blinded to all patient information and case-control status to ensure objectivity. We use PD as a standard for comparisons.

Conditional logistic regression modeling was used to estimate odds ratios (ORs) and 95% confidence intervals (CIs) for continuous measures and quartile models for both the calibrated breast density measures and the standard PD measure. Breast density distributions were log-transformed to reduce distribution skewness Continuous ORs are presented as per standard deviation (SD) increase determined from the respective breast density measurement distribution. The area under the receiver operating characteristic curve (Az) was used to compare the breast density measures in their ability to discriminate between cases and controls. We also performed a subgroup analysis by dichotomizing the population by women with H and D unit images to evaluate the consistency of the calibration approach. Although these units are from the same manufacturer, there are design differences that could influence the calibration. The compression paddles attach differently, which could influence the compressed breast thickness correction and erosion process parameters. Moreover, for the same target/filter combination, calibration curves also differ across the H and D systems (see Supplemental Tables). Models will be presented that are unadjusted and adjusted for body mass index (BMI) and ethnicity. For comparing proportions, the McNemar’s (exact) test was used when considering the entire case-control population and the binomial proportional test to compare the internal-external case groups. To compare continuous measures of breast density, the t-test was used. We also performed regression analyses on the scatter plots between each calibrated measure of breast density with PD. Both regression slopes (m) and correlation coefficients (R) are provided with 95% CIs. Image processing was performed in the IDL environment (Version 8.6, Exelis Visual Information Solutions, Inc., Jersey City, NJ) using version 8.6 and regression analyses in the SAS environment (SAS Institute Inc., Cary, NC) using version 9.4.

3. Results

Patient characteristics for the entire population are provided in Table 1a. Both Caucasian (p = 0.91) and African American (p = 0.24) women are represented similarly across case and control groups, but the proportion of Hispanic women was higher among controls (18.5%) than cases (9%) [p< 0.001]. Differences in BMI (p=0.07) and menopausal status (p = 0.46) were not statistically significant. Similarities in the breast density measures were mixed across cases and controls: both PD (p = 0.05) and PG_sd (p = 0.05) showed similar differences across these groups whereas PG_m was similar (p = 0.66). Table 1b shows the breakdown for the case population stratified by internal and external status. For the most part, race and ethnicity were similar across these groups: Caucasian (p = 0.73); African American (p = 0.54); Hispanic (p = 0.24); and non-Hispanic (p = 0.20). In contrast, screening group were similar: group 1 (p = 0.07); group 2 (p = 0.22); and group 3 (p = 0.25). Breast density, BMI (p = 0.86) and menopausal status (p = 0.53) were similar across these groups. PD (p = 0.17), PG_m (p = 0.34) and PG_sd (p = 0.72) were also similar. Internal and external cases were similar in most factors; differences within the screening categories are accounted for in the matching. Differences in both BMI and the number of Hispanic women between cases and controls (Table 1a) indicate the importance of controlling for these factors in the modeling.

Table 1b.

Internal and External Population Characteristics: This table describes the internal-external case population characteristics by either the distribution mean for a given measure or percentages of the population. When applicable the standard deviation of the respective distribution is provided parenthetically.

Measure	p-values	Internal n	Internal Mean (standard deviation) or relative frequency	External n	External Mean (standard deviation) or relative frequency	Total n	Total Mean (standard deviation) or relative frequency
Age	0.01	166	60.3 (11.3)	153	57.2 (11.1)	319	58.8 (11.3)
Race
Caucasian	0.73	141	84.9%	132	86.3%	273	85.6%
African-American	0.54	15	9.0%	11	7.2%	26	8.2%
Asian	0.91	4	2.4%	4	2.6%	8	2.5%
More than One	0.49	1	0.6%	2	1.3%	3	0.9%
Other	0.59	2	1.2%	1	0.7%	3	0.9%
Unknown	0.92	3	1.8%	3	2.0%	6	1.9%
Ethnicity
Non-Hispanic	0.20	146	88.0%	141	92.2%	287	90.0%
Hispanic	0.24	18	10.8%	11	7.2%	29	9.1%
Unknown	0.59	2	1.2%	1	0.7%	3	0.9%
BMI	0.86	166	28.7 (5.5)	153	28.8 (6.4)	319	28.7 (6.0)
Screening Group
Group 1	0.07	124	74.7%	100	65.4%	224	70.2%
Group 2	0.22	26	15.7%	32	20.9%	58	18.2%
Group 3	0.25	16	9.6%	21	13.7%	37	11.6%
HRT Usage	0.11
Current		13	7.8%	5	3.3%	18	5.6%
Not Currently		153	92.2%	148	96.7%	301	94.4%
Menopausal Status	0.58
Pre-Menopausal		39	23.5%	40	26.1%	79	24.8%
Menopausal		127	76.5%	113	73.9%	240	75.2%
PD	0.17	166	24.5 (11.4)	153	26.3 (11.8)	319	25.3 (11.6)
PGM	0.34	166	25.9 (17.0)	153	24.1 (16.4)	319	25.0 (16.7)
PGSD	0.72	166	9.8 (5.2)	153	10.0 (5.0)	319	9.9 (5.1)

Breast cancer associations are provided in Table 2. ORs for PD were statistically significant in the third quartile [OR = 1.60 (1.02, 2.53)] in the unadjusted model, upper two quartiles in the adjusted model [OR = 1.96 (1.19, 3.23) and 2.27 (1.29, 4.00)], and in the continuous models [OR = 1.21(1.02, 1.43) and 1.43 (1.16, 1.76)]. ORs for PG_sd were statistically significant in the upper two quartiles [OR = 1.86 (1.13, 3.0 and 1.71 (1.03, 2.86)] in the unadjusted model, upper three quartiles [OR = 1.83 (1.11, 3.02), 2.19 (1.28, 3.73), and 2.20 (1.26, 3.85)] in the adjusted model, and in continuous models [OR = 1.25 (1.03, 1.50) and 1.34(1.09, 1.66)]. In contrast, PG_m was not statistically significantly associated with breast cancer, although similar in direction for the other measures with similar Az (see Table 2). For reference, Figure 1 shows four clinical display mammograms and the corresponding calibrated images. In the continuous models both PD and PG_sd provided similar Az: 0.56 to 0.58 (unadjusted) and 0.62–0.61 (adjusted). PD and PG_sd also provided similar Az in the quartile models: 0.55 to 0.54 (unadjusted) and 0.65 to 0.64 (adjusted). In contrast, the Az for PG_m was marginally attenuated.

Table 2.

Breast Cancer Associations: This table gives the breast cancer associations for the percentage of breast density using the Cumulus method (PD), and the calibrated standard deviation (PG_sd) and average (PG_m) measures. Odds ratios (ORs) are provided with 95% confidence intervals parenthetically. The standard deviation (SD) for each breast density distribution and area under the receiver operating characteristic curve are provided.

PD					PGsd					PGm
QRT	Case n=319	unadjusted OR (95% CI)	BMI and ethnicity adjusted OR (95% CI)		QRT	Case n=319	unadjusted OR (95% CI)	BMI and ethnicity adjusted OR (95% CI)		QRT	Case n=319	unadjusted OR (95% CI)	BMI and ethnicity adjusted OR (95% CI)
	67	1.00 (Ref.)	1.00 (Ref.)			57	1.00 (Ref.)	1.00 (Ref.)			72	1.00 (Ref.)	1.00 (Ref.)
2	56	0.84 (0.53, 1.34)	0.84 (0.52, 1.38)		2	82	1.53 (0.95, 2.47)	1.83 (1.11, 3.02)		2	89	1.24 (0.78, 1.97)	1.39 (0.85, 2.27)
3	102	1.60 (1.02, 2.53)	1.96 (1.19, 3.23)		3	95	1.86 (1.13, 3.07)	2.19 (1.28, 3.73)		3	70	0.98 (0.60, 1.61)	1.11 (0.66, 1.89)
4	94	1.53 (0.97, 2.43)	2.27 (1.29, 4.00)		4	86	1.71 (1.03, 2.86)	2.20 (1.26, 3.85)		4	88	1.26 (0.77, 2.07)	1.69 (0.94, 3.04)
Az		0.55 (0.50, 0.61)	0.65 (0.59, 0.70)		Az		0.54 (0.48, 059)	0.64 (0.59, 0.69)		Az		0.53 (0.48, 0.58)	0.60 (0.54, 0.65)
Log Con	SD	unadjusted OR (95% CI)	BMI and ethnicity adjusted OR (95% CI)		Log Con	SD	unadjusted OR (95% CI)	BMI and ethnicity adjusted OR (95% CI)		Log Con	SD	unadjusted OR (95% CI)	BMI and ethnicity adjusted OR (95% CI)
	0.491	1.21 (1.02, 1.43)	1.43 (1.16, 1.76)			0.659	1.25 (1.03, 1.50)	1.34 (1.09, 1.66)			0.819	1.12 (0.93, 1.34)	1.22 (0.99, 1.51)
Az		0.56 (0.51,0.62)	0.62 (0.56, 0.67)		Az		0.58 (0.52, 0.63)	0.61 (0.55, 0.66)		Az		0.53 (0.48, 0.59)	0.60 (0.54, 0.65)

Regression analysis was performed to study the correlation between PD and the calibrated measures. Figure 2 shows the corresponding scatter plots: R = 0.67 (0.63, 0.72) for PD vs. PG_m; and R = 0.60 (0.55, 0.65) for PD vs. PG_sd (regression quantities are provided in the caption). The influence of the CIs for each slope is superimposed on the respective plot (dashed line). Figure 1 illustrates the correspondence between PG_sd and PD. The clinical display mammograms shown in the top row are aligned according quartiles of PG_sd. The bottom row shows the associated calibrated images. The corresponding quartiles for PD for the images in the top row are the fourth, first, fourth and fourth. This lack of correspondence parallels the moderate correlation between these measures illustrated in Figure 2. In contrast with PD, it is more difficult to discern by observation images with increased PG_sd. As noted in the top row, images that appear more spatial homogeneous with lower density tend to have lower PG_sd values. We stratified the associations by 2D and DBT units shown in Table 3 because the calibration was applied across different imaging platforms. PD gave similar associations compared with the findings from the 2D units; the central values for the D units are similar but without statistical significance. The PG_sd measure followed a similar trend as PD. PG_m was similar across H and D units with loss of statistical significance in the D units. To explore whether the erosion process influenced the findings, we performed the association analyses for 35%, 30%, 20%, and 15% erosion depths for the D units, and the findings (not shown) did not differ from the respective quantities provided in Table 3. Table 4 provides the intra-measurement comparison across the image platforms separated by case-control status. The PG_m measure shows a significant difference for the case group across platforms.

Figure 2.

Correlation Analyses: This shows the scatter plots for PD (vertical axis) with PG_m (left pane) and PG_sd (right pane) plotted with solid points. The respective regression lines (solid) are given by: PD = 0.47×PG_m + 12.0 (left); and PD = 1.39× PG_sd + 11.2 (right) with R = 0.67 and 0.60 respectively. The 95% confidence intervals for each slope are superposed on respective regression lines (dashed lines): (0.44, 0.50) PD vs. PG_m and (1.25, 1.54) for PD vs. PG_sd.

Table 3.

Breast Cancer Associations dichotomized by two-dimensional (2D) and DBT units: This provides the breast cancer associations for the percentage of breast density using the Cumulus method (PD), and the calibrated standard deviation (PG_sd) and average (PG_m) measures dichotomized by 2D (H) and DBT units (D). Odds ratios (ORs) are provided with 95% confidence intervals parenthetically. The standard deviation (SD) and area under the receiver operating characteristic curve (Az) are provided for each model.

PD H Units					PGsd H Units					PGm H Units
QRT	Case n=185	unadjusted OR (95% CI)	BMI and ethnicity adjusted OR (95% CI)		QRT	Case n=185	unadjusted OR (95% CI)	BMI and ethnicity adjusted OR (95% CI)		QRT	Case n=185	unadjusted OR (95% CI)	BMI and ethnicity adjusted OR (95% CI)
	37	1.00 (Ref.)	1.00 (Ref.)			34	1.00 (Ref.)	1.00 (Ref.)			37	1.00 (Ref.)	1.00 (Ref.)
2	27	0.80 (0.43, 1.48)	0.82 (0.43, 1.55)		2	40	1.24 (0.64, 2.42)	1.50 (0.74, 3.03)		2	61	1.66 (0.90, 3.04)	1.85 (0.98, 3.50)
3	62	1.89 (1.02, 3.51)	2.23 (1.15, 4.35)		3	59	1.92 (1.01, 3.63)	2.36 (1.19, 4.66)		3	38	1.09 (0.55, 2.16)	1.24 (0.59, 2.58)
4	59	1.75 (0.95, 3.23)	2.19 (1.06, 4.52)		4	52	1.67 (0.87, 3.18)	2.02 (1.01, 4.04)		4	49	1.38 (0.72, 2.64)	1.63 (0.76, 3.49)
Az		0.59 (0.52, 0.66)	0.68 (0.61, 0.75)		Az		0.54 (0.47, 0.62)	0.64 (0.57, 0.71)		Az		0.56 (0.49, 0.63)	0.63 (0.56, 0.70)
Log Con	SD	unadjusted OR (95% CI)	BMI and ethnicity adjusted OR (95% CI)		Log Con	SD	unadjusted OR (95% CI)	BMI and ethnicity adjusted OR (95% CI)		Log con	SD	unadjusted OR (95% CI)	BMI and ethnicity adjusted OR (95% CI)
	0.482	1.30 (1.04, 1.62)	1.45 (1.11, 1.89)			0.642	1.32 (1.03, 1.69)	1.42 (1.07, 1.88)			0.721	1.14 (0.91, 1.42)	1.25 (0.96, 1.63)
Az		0.58 (0.51, 0.65)	0.64 (0.57, 0.71)		Az		0.61 (0.54, 0.68)	0.64 (0.57, 0.71)		Az		0.55 (0.47, 0.62)	0.60 (0.53, 0.67)
PD D Units					PGsd D Units					PGm D Units
QRT	Case n=134	unadjusted OR (95% CI)	BMI and ethnicity adjusted OR (95% CI)		QRT	Case n=134	unadjusted OR (95% CI)	BMI and ethnicity adjusted OR (95% CI)		QRT	Case n=134	unadjusted OR (95% CI)	BMI and ethnicity adjusted OR (95% CI)
	30	1.00 (Ref.)	1.00 (Ref.)			57	1.00 (Ref.)	1.00 (Ref.)			40	1.00 (Ref.)	1.00 (Ref.)
2	30	0.98 (0.49, 1.96)	0.97 (0.46, 2.06)		2	82	1.15 (0.57, 2.37)	1.60 (0.73, 3.52)		2	20	0.46 (0.21, 1.00)	0.42 (0.18, 0.96)
3	40	1.34 (0.68, 2.66)	1.75 (0.81, 3.80)		3	95	1.28 (0.62, 2.65)	1.33 (0.61, 2.89)		3	41	0.96 (0.48, 1.92)	1.06 (0.50, 2.24)
4	34	1.17 (0.58, 2.37)	2.17 (0.86, 5.49)		4	86	1.22 (0.56, 2.68)	1.90 (0.75, 4.79)		4	33	0.72 (0.31, 1.65)	1.03 (0.38, 2.75)
Az		0.50 (0.42, 0.58)	0.63 (0.54, 0.71)		Az		0.52 (0.43, 0.60)	0.59 (0.51, 0.67)		Az		0.57 (0.48, 0.65)	0.63 (0.55, 0.72)
Log Con	SD	unadjusted OR (95% CI)	BMI and ethnicity adjusted OR (95% CI)		Log Con	SD	unadjusted OR (95% CI)	BMI and ethnicity adjusted OR (95% CI)		Log con	SD	unadjusted OR (95% CI)	BMI and ethnicity adjusted OR (95% CI)
	0.504	1.11 (0.86, 1.42)	1.42 (1.01, 2.00)			0.685	1.15 (0.86, 1.52)	1.24 (0.91, 1.71)			0.935	1.08 (0.81, 1.45)	1.18 (0.85, 1.64)
Az		0.53 (0.45, 0.61)	0.628		Az		0.53 (0.45, 0.61)	0.58 (0.50, 0.67)		Az		0.51 (0.43, 0.60)	0.57 (0.49, 0.66)

Table 4.

Measurement Comparisons across FFDM and DBT units: This table provides p-values from the intra-measurement comparisons (t-test). Patients were dichotomized according to the FFDM technology used for their imaging. The analysis was further separated by case and control status.

	Case					Control
Breast Density Measure	H unit mean	D unit mean	H unit standard deviation	D unit standard deviation	p-value	H unit mean	D unit mean	H units standard deviation	D unit standard deviation	p-value
PD	25.89	24.59	11.32	11.95	0.3214	23.51	23.83	11.68	12.02	0.8141
PGsd	9.91	9.78	4.94	5.39	0.8194	9.05	9.30	4.83	5.15	0.6519
PGm	23.43	27.15	15.48	18.18	0.0563	22.28	27.52	15.52	19.08	0.0095

4. Discussion

The current work was based on data from six mammography units including three conventional FFDM units and three DBT units. As noted, there are differences between these units that could influence calibration accuracy. The calibrated PG_sd measure yielded estimates of breast cancer risk comparable to those based on PD in this study for both the main and subgroup analyses. In the subgroup analyses for the D units, both measures lost significance in the quartile models, whereas PG_sd also lost significance in the continuous model as well; this loss may be due to the decreased number of samples (i.e. n = 134). The PG_sd measure was also comparable with PD previously when evaluating calibrated measures from indirect x-ray conversion FFDM (30). Although the Az quantities for both PD and PG_sd in this work indicate that these measures are not strong discriminators, these quantities are within range of for both PD and other density metrics previously reported (5, 31, 32). It follows that the calibrated PG_sd measure generalizes across FFDM technologies. In contrast, the PG_m measure did not produce significant findings in this report. In a previous comparison study (18), the PG_m measure produced significant findings although weaker than those provided by PG_sd. The current association findings for PG_sd are consistent with previous comparisons and with other commercially available standardized metrics that capture volumetric mammographic density (5).

There are a variety of methods for measuring breast density (11, 33, 34) and it can be difficult to make comparison across different density metrics due to scaling variation, and differences in study design and study populations. Astley et al compared five methods of measuring breast density including both automated and operator assisted methods (33). These measures capture mammographic density (i.e. the degree of dense tissue) summarized with planar percentages, area, and volumetric measures. Destounis et al compared qualitative and quantitative breast density measures (11). Most of these measures also capture dense tissue with similar metrics with the exception of Tabar system (35), which assesses structure with a defined five point operator rating. The PG_sd metric captures variation in mammographic density, which is a different characteristic compared to the degree of dense tissue. We speculate that PG_sd may be capturing elements characterized by the Tabar scale. In this report, PD serves as a side-by-side reference standard. We note our findings for continuous PD in this report are in agreement with a recent meta-analysis that examined percentage of breast density measures including PD (4). Our comparison with PD provides context for the merits of the PG_sd metric.

There are several limitations with our study. We used cases from two sources. Both the study design and similarities between the two case groups indicate this has little impact on the findings. Calibration requires an accurate estimate of the compressed breast thickness. As such, this methodology does not include the entire breast due to the erosion process, required to eliminate breast area where the thickness was uncertain. Thus, some breast tissue was not accounted for in the density estimation. We believe this is a nominal limitation with respect to the variation measure due to our current and previous findings. In this study, we used images from a specific manufacture. This limitation is partially mitigated when considering that the H and D units have different calibration data and the compression paddles differ. Moreover, this calibration approach was developed and evaluated previously with a case-control study using images from a different FFDM technology.

5. Conclusion

There is a critical need to measure breast density accurately as it is the foundation for appropriate clinical application. The BI-RADS (American College of Radiology) tissue composition descriptors are used to indicate when mammography might be ineffective and are part of the standardized clinical reporting. Evidence indicates that many women will experience shifts in their composition classification due to operator variability (17). Decades of research have clearly demonstrated that breast density is a strong and well-established breast cancer risk factor. Although there are many studies assessing various breast density methods for risk prediction, there is no standard or accepted measure (11), and it is not included routinely in clinical risk assessments (16). Accurate risk assessments can be applied for screening and prevention programs. Current absolute risk models are useful for individual counseling and evaluating risk associated with interventions, but are not accurate enough to inform high-risk prevention strategies or deciding who should be screened (36). Moreover, there are no guidelines for supplemental screening of high density women, although evidence shows that supplemental screening detects more cancers for this group but could increase the recall rate (17). In sum, an accepted standardized measure of breast density is required to fully actualize the benefits of personalized breast screening and interventions. The current study represents an important step forward in realizing these anticipated benefits.

Abbreviations

2D

Two-Dimensional

Ag

Silver

Az

Area under the receiver operating characteristic curve

BI-RADS

Breast Imaging Reporting and Data System (American College of Radiology)

BMI

Body Mass Index

BTE

Breast Tissue Equivalent

CC

Cranial Caudal

CIs

Confidence Intervals

Comb-HD

Combined 2D and DBT acquisition including the C-view synthetic 2D image

D

Collection of DBT units

DBT

Digital Breast Tomosynthesis

FAST

Automated Self-Adjusting Tilt Compression Paddle^™ (Hologic, Inc., Bedford, MA)

FFDM

Full Field Digital Mammography

FOV

Field of View

H

Collection of 2D units

HRT

Hormone Replacement Therapy

IDL

Interactive Data Language (Exelis Visual Information Solutions, Boulder, CO)

MS

Menopausal Status

n

Number

OR(s)

Odds Ratio(s)

PD

Percentage of Breast Density

PG_m

Calibrated Mean Breast Density Measure

PG_sd

Calibrated Standard Deviation Breast Density Measure

Q

Quartile

QRT

Quartile

R

Linear Correlation Coefficient

Rh

Rhodium

SAS

(SAS Institute Inc., Cary, NC)

SD

Standard Deviation

W

Tungsten

XXX

Blinded Author Name