Association of Computerized Mammographic Parenchymal Pattern Measure with Breast Cancer Risk: A Pilot Case-Control Study

The proposed mammographic parenchymal pattern measure showed a strong association with breast cancer risk and a low correlation with percentage density and some known risk factors; it may serve as an independent factor in a breast cancer prediction model.

Abstract

Purpose:

To develop a computerized mammographic parenchymal pattern (MPP) measure and investigate its association with breast cancer risk.

Materials and Methods:

A pilot case-control study was conducted by collecting mammograms from 382 subjects retrospectively. The study was institutional review board approved and HIPAA compliant. Informed consent was waived. The cases included the contralateral mammograms of cancer patients (n = 136) obtained at least 1 year before diagnosis. The controls included mammograms of healthy subjects (n = 246) who had cancer-free follow-up for at least 5 years. The data set was historically divided into a training set and an independent test set. An MPP measure was designed to analyze the texture patterns of fibroglandular tissue in the retroareolar region. Odds ratios (ORs) were used to assess the association between breast cancer risk and MPP. To test the trend in ORs, we divided the MPP measure into three categories (C1, C2, and C3) on the basis of its values from low to high, with C1 as the baseline. The confounding factors in this study included patient age, body mass index, first-degree relatives with history of breast cancer, number of previous breast biopsies, and percentage density (PD).

Results:

Among all of the subjects from the training and test data sets, the Pearson product-moment correlation coefficient between MPP and PD was 0.13. With logistic regression to adjust the confounding, the adjusted ORs for C2 and C3 relative to C1 in the test set were 2.82 (P = .041) and 13.89 (P < .001), respectively.

Conclusion:

The proposed MPP measure demonstrated a strong association with breast cancer risk and has the potential to serve as an independent factor for risk prediction.

© RSNA, 2011

Supplemental material: http://radiology.rsna.org/lookup/suppl/doi:10.1148/radiol.11101266/-/DC1

Introduction

Breast cancer remains one of the leading causes of death among women of 40 years and older (1). Screening mammography is the most effective and low-cost method to date for early detection and diagnosis of breast cancer. Previous studies and clinical trials (2–6) have shown that early diagnosis and improved treatment strategies substantially improved the recurrence-free survival and reduced breast cancer mortality. An approach to further reduction of breast cancer death is to identify women at higher risk of developing breast cancer as candidates for preventive treatment (7,8). Wolfe (9) suggested that the radiographic appearance of breast parenchyma provides information for predicting breast cancer risk. Findings in other studies (10–12) suggested that the association of parenchymal pattern with breast cancer risk may be distorted because mammographic density reduces the sensitivity of cancer detection. Several studies (13–15) with large data sets showed that the masking effect of density did exist but it operated in addition to the differences in the risk of breast cancer related to the breast parenchymal patterns. Mammographic density has been shown to be one of the most important independent risk factors for breast cancer (9,16,17), with risk estimates threefold to fivefold greater for women with density in the highest quartile than for women of similar age with density in the lowest quartile (18). Screening mammography has the potential to assist physicians in breast density assessment and, thus, in the prediction of risk for breast cancer. The results of a recent meta-analysis (19) suggested that screening for breast cancer risk in all postmenopausal women should be performed by using risk factors and breast density.

Qualitative descriptions of mammographic density and patterns are implicated in the Wolfe mammographic parenchymal pattern classification for breast cancer risk (9) and in the current Breast Imaging Reporting and Data System density categories (20). At present, the most common approach to breast density estimation is that the radiologist assigns the breast density category following the Breast Imaging Reporting and Data System lexicon by visual judgment of the fibroglandular tissue imaged on mammograms. However, researchers in previous studies found that the assignment of density categories showed various degrees of agreement among radiologists (κ statistic, 0.43–0.76) (21,22). It can be expected that a more accurate and reproducible measure than Breast Imaging Reporting and Data System breast density that does not depend on the subjective impression of human readers may further improve the reliability of breast density estimates for risk prediction.

Although computerized quantitative measurements of breast density are still at an early stage, several studies have demonstrated that computerized techniques can improve the reproducibility and interobserver agreement of breast density measurement (15,22,23). Computerized analysis of the distribution pattern of the breast tissue on mammograms, referred to as the mammographic parenchymal pattern (MPP) measure hereafter, is different from the mammographic density measurements, referred to as percentage density (PD), in which the density is measured as the percentage of dense area estimated from the segmented breast on mammograms. The pattern of breast parenchyma may be changing during the course of breast cancer development before the cancer becomes visible on mammograms. Our purpose was to develop a computerized MPP measure and investigate its association with breast cancer risk.

Materials and Methods

Study Population

This retrospective study was institutional review board approved and Health Insurance Portability and Accountability Act compliant. Informed consent was waived. The study population consisted of women who underwent mammographic screening at the University of Michigan Health System (Ann Arbor, Mich) from 1993 through 2002. Information on breast cancer risk factors, including demographic information, family history of breast cancer, number of breast biopsies, menstrual status, and height and weight, from which we calculated the body mass index (BMI), was routinely collected by means of a self-administered questionnaire for every patient at the time of screening mammography.

Data Set

We historically collected two data sets for development and evaluation of the MPP measure. Our first data set consisted of 217 subjects (81 cases and 136 controls) and was collected for training the MPP measure. Our second independent test set consisted of 165 subjects (55 cases and 110 controls) and was collected after training of the MPP measure.

Mammograms and demographic data of subjects in case groups were consecutively collected from the files of patients who had undergone mammographic evaluation and biopsy at our institution. Eligible subjects had unilateral biopsy-proved breast cancer and normal mammograms dated at least 1 year prior to the diagnosis of breast cancer.

Mammograms and demographic data of control subjects in control groups were also consecutively collected from the files of patients at our institution. Eligible control subjects included women of any age who had normal screening mammograms and had breast cancer–free follow-up for at least 5 years. Excluded were women who had palpable abnormalities, current clinical concerns, prior breast cancer, and breast implants or who underwent follow-up mammography for probably benign findings.

In all subjects, four-view screen-film mammograms, the craniocaudal view and the mediolateral oblique view from each breast, were obtained. In this pilot study, we only used craniocaudal-view mammograms for analysis. For the case subjects, only contralateral mammograms that were confirmed to be free of cancer were used. For the control subjects, we arbitrarily chose to use the craniocaudal-view mammograms from the left breast because their right and left breast parenchyma can be considered symmetrical. All screen-film mammograms were acquired with a screen-film system (MinR; Eastman Kodak, Rochester, NY) and were digitized with a laser film scanner (Lumiscan 85; Lumisys, Los Altos, Calif) at a pixel resolution of 50 × 50 μm and 4096 gray levels. The digitizer was calibrated so that gray-level values were linearly proportional to the optical density in the range of zero to four, with a slope of 0.001 per pixel value. The digitizer output was linearly converted so that a large pixel value corresponded to a low optical density. The full-resolution mammograms were smoothed with a 2 × 2 box filter and were subsampled by a factor of two, resulting in 100 × 100-μm images. The images at a pixel size of 100 × 100 μm were used for the computerized analysis.

Estimation of PD

We used an automated method (Mammographic Density ESTimator) developed previously in our laboratory (24) for segmentation of the dense fibroglandular area on mammograms. Briefly, a dynamic-range compression technique reduces the gray-level range of the breast area. On the basis of the shape of the gray-level histogram, the breast on the mammogram can be classified into four classes, and a threshold level is automatically determined to separate the dense and fatty pixels. The mammographic density is estimated as the percentage of fibroglandular tissue area relative to the total breast area.

Computerized MPP Descriptor

An MPP measure was designed to analyze the texture patterns of fibroglandular tissue in the retroareolar region. Our computerized scheme for MPP analysis is shown in Figure 1 (25). For an input mammogram, the first step was to perform automated breast boundary detection (26) to separate the breast region from other background features such as the directly exposed area, patient identification information, and lead markers. Parenchymal analysis was then performed only in the breast region.

Figure 1:

Diagram of our preliminary computerized scheme for MPP analysis. MPP-ROI = region of interest (ROI) for MPP analysis. CC = craniocaudal.

Nipples were manually identified, and then all subsequent image processing steps were performed automatically. A horizontal line was drawn from the nipple to the chest-wall edge of the image. A 512 × 512-pixel region of interest (MPP-ROI) was then moved from the anterior toward the chest wall, with its center placed along the horizontal line until it fitted in the breast region. To avoid the fatty area around the breast boundary and to cover the main area of the breast parenchyma, the MPP-ROI location was further shifted toward the chest wall, with a maximum shift of 4 mm. However, a minimum of 2 mm of space was kept between the MPP-ROI and the chest wall. The actual number of pixels of the shift varied, depending on the size of the breast. If the 512 × 512-pixel ROI was too large to fit in the breast, the horizontal or vertical side length of the ROI would be automatically reduced so that it is fitted into the breast while satisfying the shift criterion. Overall, 10.1% of ROIs in the training set (22 of 217) were reduced in size, with the smallest ROI having 336 × 352 pixels, and 5.5% of ROIs in the test set (nine of 165) were reduced in size, with the smallest ROI having 384 × 512 pixels.

After the localization of MPP-ROI, 10 run-length statistics and five region-size statistics features were extracted. Details of this process are described in Wu et al (25) and in Appendix E1 (online). Stepwise linear discriminant analysis with simplex optimization was used to search for the best combination of texture features that characterize the parenchymal pattern for breast cancer risk prediction in the training set. The best subset of features was then treated as the input predictor variables to formulate a linear discriminant function, the weights of which were also estimated from the training set. The trained linear discriminant classifier was applied to the independent test set to generate the MPP measure for each subject.

Statistical Analysis

Odds ratios (ORs) were used to assess the association between breast cancer risk and MPP. To test the trend in ORs, we divided the MPP measure into three categories (C1, C2, and C3) on the basis of the histogram of MPP values in the training set from low to high, with C1 as the baseline. The confounding factors included patient age, BMI, first-degree relatives with a history of breast cancer, number of previous breast biopsies, and PD. Details of this analysis are presented in Appendix E2 (online). All statistical analyses were performed by using a software program (PASW Statistics 18.0; IBM, Armonk, NY) and free software for statistical computing (R version 2.12.0; http://www.r-project.org/).

Results

Table 1 summarizes breast cancer risk factors of our population. For the training set, the case subjects had a mean age at screening mammography of 57.4 years ± 12.0 (standard deviation), and the control subjects had a mean age at screening mammography of 53.1 years ± 10.6, with P = .008. For the test set, the case subjects had a mean age at screening mammography of 59.1 years ± 9.6, and the control subjects had a mean age at screening mammography of 56.0 years ± 9.5, with P = .052. In addition, the average PDs were 5.1% (P = .012) and 5.4% (P = .027) greater in the case group than in the control group for the training and test data sets, respectively. First-degree family history of breast cancer and number of breast biopsies (zero and ≥1) were treated as bicategorical variables in the rest of the analysis.

Table 1.

Breast Cancer Risk Factors Analyzed in This Pilot Study

Note.—Numbers in parentheses are percentages. All percentages (except percentages for PD) are relative to the total number of subjects in each group. SD = standard deviation.

^*Student unpaired t test, two sided.

^†Values are for case group versus control group in training set.

^‡Values are for case group in training set versus case group in test set.

^§Values are for case group versus control group in test set.

^‖Values are for control group in training set versus control group in test set.

^#Mann-Whitney U test.

During the design of the MPP descriptor by using the training set, three features including one from the region-size statistics features (large-size emphasis) and two from the run-length statistics features (gray-level nonuniformity in the vertical direction and run-length nonuniformity in the horizontal direction) were selected as the predictor variables for the linear discriminant classifier (Fig 2). The common property of these three texture features was to characterize the spatial variations of density in the ROIs in terms of gray levels, as well as run lengths.

Figure 2:

Box plot of computerized MPP measures.

The correlations of the MPP with confounding factors are summarized in Table 2. The effectiveness for differentiating case subjects from control subjects in the test set in terms of the area under the receiver operating characteristic (ROC) curve (A_z) for age, PD, and the MPP measure were 0.60 ± 0.05, 0.61 ± 0.05, and 0.74 ± 0.05, respectively (Fig 3). The difference in the A_z values between age and PD did not achieve significance (P = .961). The A_z for the MPP measure was significantly higher than that for either age (P = .046) or PD (P = .038).

Table 2.

Correlation of the MPP with Confounding Variables

^*The 95% confidence interval (CI) was calculated on the basis of the Fisher Z transformation.

^†Numbers in parentheses are the Point-Biserial correlation (27), which is a special case of the Pearson product-moment correlation in which one variable is continuous and the other variable is dichotomous.

^‡First-degree family history of breast cancer and numbers of breast biopsies are bicategorical variables.

Figure 3:

Test ROC curves for classification of the case group and the control group according to patient age, PD, and the computerized MPP measure, with A_z values of 0.60 ± 0.05, 0.61 ± 0.05, and 0.74 ± 0.05, respectively.

Although the OR for the MPP (continuous variable) was attenuated after the adjustment for the effects of other variables, the univariate association observed between MPP and breast cancer risk persisted in the multivariate analysis (Table 3). MPP (P < .001), PD (P = .046), and patient age (P = .017) were three factors that contributed significantly to the prediction of breast cancer risk. The trained logistic regression model with all six factors had an A_z value of 0.83 ± 0.03 for the training set and an A_z value of 0.78 ± 0.04 for the test set. The model without patient age and BMI had A_z values of 0.82 ± 0.03 and 0.76 ± 0.04 for the training set and the test set, respectively.

Table 3.

Multivariate Logistic Regression Analysis to Estimate Conditional Effects of MPP Measure on Breast Cancer Risk

^*The univariate analysis was performed to examine the association between MPP and breast cancer risk.

^†The multivariate analysis was performed to examine the independent contribution of breast cancer risk.

^‡First-degree family history of breast cancer and number of breast biopsies are bicategorical variables.

The ORs associated with the MPP measure with and without adjustment of confounding factors are compared in Table 4. The increase in the ORs from the baseline C1 (Table 4) was significant for the C2 group (P = .041) and the C3 group (P < .001). The risk prediction model with MPP as an additional factor is significantly improved in comparison with the model without MPP (P < .001). Image examples from each of the three categories are illustrated in Figure 4.

Table 4.

Crude and Adjusted Breast Cancer ORs Associated with the Computerized MPP Measure

Note.—For the training set, n = 136 for controls and 81 for cases. For the test set, n = 110 for controls and 55 for cases. For the training and test sets combined, n = 246 for controls and 136 for cases. For comparison of the logistic regression models before and after adding MPP, χ² = 42.25, with P < .001.

^*Adjusted with age, BMI, first-degree family history of breast cancer, and number of previous breast biopsies.

^†Adjusted with PD, in addition to the four factors of age, BMI, first-degree family history of breast cancer, and number of previous breast biopsies.

^‡P = .039, two sided.

^§P < .001, two sided.

^‖P = .041, two sided.

^#P < .001, two sided.

^**P = .006, two sided.

^††P < .001, two sided.

Figure 4a:

(a–c) Craniocaudal mammographic images from the test set, with MPP measures of the three categories: (a) C1, (b) C2, and (c) C3. Images a and b were obtained from a 57-year-old woman in the control group, and image c was obtained from a 53-year-old woman in the case group.

Figure 4b:

(a–c) Craniocaudal mammographic images from the test set, with MPP measures of the three categories: (a) C1, (b) C2, and (c) C3. Images a and b were obtained from a 57-year-old woman in the control group, and image c was obtained from a 53-year-old woman in the case group.

Figure 4c:

(a–c) Craniocaudal mammographic images from the test set, with MPP measures of the three categories: (a) C1, (b) C2, and (c) C3. Images a and b were obtained from a 57-year-old woman in the control group, and image c was obtained from a 53-year-old woman in the case group.

Discussion

The appearance of the fibroglandular tissues in the breast may be a biomarker of future breast cancer risk. The association of breast density with breast cancer risk is greater than that for most other established breast cancer risk factors, with the exception of age and mutations in the breast cancer–susceptibility genes BRCA1 and BRCA2 (28). The breast density can be estimated with the percentage of dense area in the breast region of the mammogram, which was one of the commonly used methods for breast density analysis, and/or assessment of the distribution pattern of fibroglandular tissue. The parenchymal pattern has been studied previously by using different methods. Brisson et al (29) observed that the radiologists’ visually estimated PD was strongly correlated with their visual categorization of the Wolfe parenchymal pattern (9) and that addition of the Wolfe parenchymal pattern did not improve the risk prediction. This finding reveals that the visual Wolfe parenchymal pattern (9) tends to be used to assess mainly the amount of density in the breast, likely owing to the fact that the human visual system has difficulty in the discrimination of textural information that is related to higher-order statistics or spectral properties on an image (30,31). Caldwell et al (32) and Byng et al (23) aimed at selecting computerized features that could correlate with the visual Wolfe parenchymal pattern (9) or PD classification and obtained features that were similar but had a weaker relationship with breast cancer risk in comparison with PD. These studies showed that, to use the rich textural information in the breast parenchyma to improve risk prediction, one must design an MPP descriptor that can characterize the complex spatial gray-level relationships of the breast structures while having a low correlation with PD.

Our study found that the computerized MPP descriptor, by using statistical texture analysis in the retroareolar region of the breast, did not show a strong correlation with PD. The Pearson product-moment correlation coefficient between these two density measurements was 0.13, with a 95% CI of 0.03 to 0.22. We also evaluated a set of confounding risk factors including patient age, BMI, first-degree relatives with a history of breast cancer, and the number of previous breast biopsies and showed that the MPP measure had a low correlation with these confounding factors (absolute value ranged from 0.03 to 0.07). These findings suggest that our MPP measure could serve as an additional risk factor that could not be explained by the established breast cancer risk factors.

In our study, we used ROC methods to compare the MPP measure to patient age and PD in terms of their capability in differentiating healthy subjects from subjects who would develop cancer in a future year. Our ROC results indicated that the MPP measure had a significantly higher accuracy for classification of the two groups than did patient age (P = .046) and PD (P = .038). Further studies are needed to investigate whether this finding will translate into higher risk prediction accuracy for the models if the MPP measure is included as an additional factor. The odds of breast cancer increase steadily with increasing MPP measure, with or without adjustment of the confounding risk factors. Our findings indicated that the risk of developing breast cancer in future years is positively associated with the proposed MPP measure that describes the parenchymal pattern on mammograms.

However, as a pilot study, this study had a number of limitations: (a) The sample size was small, and we could not perform stratified analysis among different patient groups; (b) the robustness of the image analysis techniques has yet to be proved with large training and test data sets; (c) the dependence of the association of the MPP measure with breast cancer risk on ROI location and ROI size was not examined; (d) many feature extraction parameters were not optimized; (e) potentially useful analysis, such as multiple-view information fusion from one or both breasts, was not explored; and (f) the association of the proposed MPP with many other risk factors was not investigated. Study is underway to collect a large data set and to improve the parenchymal feature analysis. Finally, although mammography is undergoing transition to full-field digital mammograms, we used digitized screen-film mammograms in this study. In our analysis, we used prior mammograms obtained in patients with at least 5 years of cancer-free follow up. Prior mammograms obtained 5 years or longer ago in our patient files were predominantly screen-film mammograms. Nevertheless, we believe that the computer-vision techniques developed with screen-film mammograms can be easily adapted to raw full-field digital mammograms (for processing) images with proper preprocessing.

In our pilot study, we developed a computerized approach to evaluate the breast parenchymal pattern on mammograms by using statistical texture analysis. This quantitative approach was motivated by the need to alleviate the problem of interobserver and intraobserver variabilities of qualitative estimates performed by human observers. A type of two-dimensional texture features, the region-size statistics features, was designed, and a quantitative MPP descriptor was formulated. This MPP descriptor also provided the advantage of utilizing higher-order statistical textural information in the breast parenchyma that may not be readily perceived by human eyes. The results of our preliminary case-control study indicated that quantitative analysis of MPP was promising for distinguishing patients who would develop breast cancer in a future year from healthy subjects. The proposed MPP measure demonstrated a strong association with breast cancer risk and a low correlation with PD and some known risk factors, suggesting that it has the potential to serve as an independent factor in a breast cancer prediction model.

Advances in Knowledge.

• The computerized mammographic parenchymal pattern (MPP) measure has a low correlation (r = 0.13) with percentage density (PD).

• The computerized MPP measure has a strong association with breast cancer risk.

Implication for Patient Care.

• The computerized MPP measure may provide useful information, in addition to PD, for prediction of breast cancer risk of an individual; it may, therefore, serve as an additional biomarker for breast cancer risk surveillance and patient management, as well as development of methods for preventive and early interventional treatment.

Disclosures of Potential Conflicts of Interest: J.W. No potential conflicts of interest to disclose. H.P.C. No potential conflicts of interest to disclose. Y.T.W. No potential conflicts of interest to disclose. C.Z. No potential conflicts of interest to disclose. M.A.H. No potential conflicts of interest to disclose. A.T. No potential conflicts of interest to disclose. L.M.H. No potential conflicts of interest to disclose. B.S. No potential conflicts of interest to disclose.