| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Breast Imaging |
1 From the Departments of Radiology (Z.H., M.L.G., D.E.W., C.E.M., W.Z.) and Hematology and Oncology (O.I.O., S.A.C.), University of Chicago, 5841 S Maryland Ave, Chicago, IL 60637; and Department of Hematology and Oncology, University of Pennsylvania Medical Center, Philadelphia (B.L.W.). From the 1998 and 2000 RSNA scientific assemblies. Received April 27, 2001; revision requested June 8; final revision received April 18, 2002; accepted May 16. Supported in part by grants from the NIH (P20 CA66132 and R21 CA79711), from the U.S. Army Medical Research and Materiel Command (DAMD 17-96-1-6058 and 17-99-1209), and from the Elizabeth Boughton Charitable Fund and the Falk Medical Research Fund. Address correspondence to M.L.G. (e-mail: m-giger@uchicago.edu).
 |
ABSTRACT |
|---|
 TOP ABSTRACT INTRODUCTIONMATERIALS AND METHODS RESULTSDISCUSSIONAPPENDIXREFERENCES |
|---|
MATERIALS AND METHODS: Mammograms from 30 carriers of BRCA1 and BRCA2 mutations and from 142 low-risk women were collected retrospectively and digitized. In addition, 60 of the 142 low-risk women were randomly selected and age matched at 5-year intervals with the 30 mutation carriers. Mammographic features were extracted from the central regions of the breast images to characterize the mammographic density and heterogeneity of dense portions of the breast. These features were then merged into a single value related to the risk of breast cancer by using linear discriminant analysis. The applicability of these computer-extracted features and the output from linear discriminant analysis to differentiate between the carriers of BRCA1 and BRCA2 mutations and the low-risk women in the entire database and in an age-matched group were evaluated by using receiver operating characteristic analysis.
RESULTS: Quantitative analysis of mammograms demonstrated that carriers of BRCA1 and BRCA2 mutations tended to have dense breast tissue, and their mammographic patterns tended to be low in contrast, with a coarse texture. Linear discriminant analysis resulted in values of the areas under the receiver operating characteristic curve of 0.91 and 0.92 in distinguishing between the BRCA1 and BRCA2 mutation carriers and the low-risk women in the entire database and the age-matched group, respectively.
CONCLUSION: The computerized analysis of mammograms suggests that mammographic patterns in carriers of BRCA1 and BRCA2 mutations differ from those of women at low risk for breast cancer. Our computer-extracted features may be useful as radiographic markers for identifying women at high risk for breast cancer.
© RSNA, 2002
Index terms: Breast, diseases, 00.32 • Breast neoplasms, radiography, 00.11 • Computers, diagnostic aid • Genes and genetics • Receiver operating characteristic (ROC) curve
|
INTRODUCTION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODS RESULTSDISCUSSIONAPPENDIXREFERENCES |
|---|
Various studies (4–7) have shown that increased mammographic density is associated with an increased risk of breast cancer. At present, the explanation for this increased risk is unclear. One possibility is that increased density reflects a larger amount of tissue at risk for developing breast cancer. Since most breast cancers develop from the epithelial cells that line the ducts of the breast, having more of this tissue, as reflected by increased mammographic density, may increase one’s chances of developing breast cancer. The histopathologic changes in BRCA1- and BRCA2-associated cancers are often characteristic of the mutant gene and different from sporadic cancers (8). For example, BRCA1-associated breast cancers occur at an earlier average age (44 years) than do sporadic breast cancers (9). In addition, as compared with noncarriers, women with BRCA1-associated breast cancers are characterized by higher-than-expected frequencies of medullary or atypical medullary carcinoma, poorly differentiated (high-grade) tumors, aneuploidy, high S-phase fraction, hormone receptor negativity, and p53 mutation (10,11). BRCA2 tumors are more differentiated, with more tubular and lobular features and a higher rate of hormone receptor positivity. However, it is not known whether differences exist in the nonmalignant breast parenchymal tissue of BRCA1 and BRCA2 mutation carriers when compared with that of noncarriers.
Mammographic parenchymal patterns have been characterized in terms of Wolfe patterns (Dy, P1, P2, and N1) (12) or percentage density (ie, percentage area that appears mammographically dense) with either visual or computerized assessment (13). These assessments were then related to breast cancer risk. Byng et al (14) found that women with dense breast tissue have a higher risk (relative risk 
2.0) of developing breast cancer on the basis of computerized assessments. Individual risk in their study was estimated from the observed breast cancer incidence in the studied population (14). Tahoces et al (15) showed that the mammographic patterns of women with the Dy pattern, which is associated with the highest breast cancer risk, as defined by Wolfe, tend to be coarser than those in women with the other three patterns (ie, P1, P2, and N1). More recently, Huo et al (16) identified computer-extracted features with which to characterize breast density and parenchymal texture patterns from digitized mammograms.
Computerized analysis of digital mammograms is expected to provide more objective and reproducible characterization of mammographic parenchymal patterns than that based on visual assessment. Consequently, the variability introduced by the subjective nature of human observers may be reduced (13). The purpose of our study was to evaluate mammographic patterns, including mammographic density, in women with germ-line mutations in BRCA1 and BRCA2 genes versus that in women at low risk for developing breast cancer, by using computer image analysis.
|
MATERIALS AND METHODS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODS RESULTSDISCUSSIONAPPENDIXREFERENCES |
|---|
In addition, mammograms were obtained consecutively from the screening mammography program of the Department of Radiology (May to December 1996) at the University of Chicago. Information on patient reproductive history, family history of breast cancer, and history of previous breast biopsy was used to calculate the lifetime risk of breast cancer by using the Gail model (17). The Gail model is one of the most common models used to assess individual breast cancer risk and was used in a recently concluded breast cancer prevention trial (18). To be considered at low risk in the study, patients could not have a family history of breast cancer, and the lifetime risk of developing breast cancer, as estimated with the Gail model, had to be less than 10%. Of the patients who supplied the necessary information, 142 were deemed to be at low risk on the basis of the previously mentioned criteria. Tests for BRCA1 and BRCA2 mutation were not offered to these low-risk patients. It is estimated that about three in 1,000 women in the United States today have inherited susceptibility to breast cancer (19). The likely prior probability that the women in the low-risk group in the current study would harbor BRCA1 and BRCA2 mutations was low, since they had no family history of breast cancer that warranted genetic testing and were regarded as low risk without having to undergo genetic testing.
The ages of the 30 BRCA1 and BRCA2 mutation carriers were 33–55 years, with a mean age of 42.7 years and a median age of 41.0 years. The ages of the women in the low-risk group were 35–54 years, with a mean age of 44.7 years and a median age of 45.0 years. To rule out possible bias due to the difference in age distributions of the mutation carriers and the low-risk individuals, 60 low-risk individuals were randomly selected and age matched at 5-year intervals with the 30 BRCA1 and BRCA2 mutation carriers. Our entire database included the 142 low-risk women and the 30 BRCA1 and BRCA2 mutation carriers. The age-matched group included 30 BRCA1 and BRCA2 mutation carriers and 60 of the 142 low-risk women who were age-matched with the 30 mutation carriers. Statistical analysis was performed on both the entire database and the age-matched database to evaluate the applicability of computer-extracted features in differentiating between mammographic patterns in the mutation carriers and low-risk women.
Computer-extracted Features and Classification
Mammograms obtained in all cases were digitized by using a laser scanner (model LD 4500; Konica Medical, Wayne, NJ) at 0.1-mm pixel size and 10-bit gray level. After the mammograms were in digital format, regions of interest (ROIs) of 256 x 256 pixels were manually selected from the central breast region immediately behind the nipple. Figure 1 shows an ROI selected from a breast image. The small ROI size (256 x 256 pixels) was chosen to include small breasts.
|
Four features were selected from stepwise feature selection by using linear discriminant analysis (LDA) on the basis of the analysis of an initial number of cases (15 mutation carriers) from a previous study (16). Details about the four selected features (ie, skewness, balance, contrast, and coarseness) are described in the Appendix. Skewness and balance of a gray-level histogram are used to quantify percentage density, while contrast and coarseness are used to characterize the textural properties of the mammographic patterns. These four selected computer-extracted features were used to evaluate the mammographic characteristics of an extended group (including the previous 15 mutation carriers) of BRCA1 and BRCA2 mutation carriers, in comparison with those of women at low risk in both the entire database and the age-matched database. In addition, we applied LDA to merge the four selected features into a single output value that was related to the risk of breast cancer. LDA is a well-established statistical technique for the classification of two groups (21). The objective of LDA is to generate a discriminant function that is a linear combination of feature variables and outputs a discriminant score for each case. The coefficients in the function are determined such that the separation between the distributions of the discriminant scores of the two groups is maximized.
Although studies (4–7) have shown that increased mammographic density is associated with an increased risk for breast cancer, to our knowledge, few studies have been performed to investigate the mammographic density patterns of women with BRCA1 and BRCA2 mutations in particular (22,23). Mammographic density patterns in these prior studies were characterized on the basis of human visual assessment. In the current study, we also had an experienced radiologist (D.E.W.) rate the mammographic patterns of the cases in our database in terms of percentage density by using a 100-point scale. The computer-calculated skewness feature, as described in the Appendix, was used to estimate percentage density on the basis of gray-level histogram analysis of an ROI.
Evaluation and Statistical Analysis
The usefulness of individual computer-extracted features in differentiation between the BRCA1 and BRCA2 mutation carriers and the women at low risk was evaluated by using receiver operating characteristic (ROC) analysis (24,25). The classification performance of the linear discriminant function in combining the four features to differentiate between the mutation carriers and the low-risk cases was also evaluated by using ROC analysis (24,25).
In ROC analysis, to evaluate the applicability of a decision variable in discriminating between two groups (eg, positive and negative), a set of sensitivity and specificity pairs is calculated by varying the critical value with which the value of the decision variable (eg, an individual computer-extracted feature or the LDA output) from each case is compared. This set of sensitivity and specificity pairs can be plotted as a set of points in a unit square in terms of true-positive fraction (equivalent to sensitivity) versus false-positive fraction (equivalent to 1 - specificity). A smooth ROC curve is then fit to the set of points in the unit square with maximum-likelihood estimation by using the LABROC4 algorithm developed by Metz et al (26). The area under the fitted ROC curve (Az) is used as an index to evaluate the inherent discriminant capacity of a studied variable in differentiating between "positive" and "negative" cases. Note that an Az value can range from 0.5 to 1.0. A higher ROC curve (ie, a larger Az value) indicates a decision variable with greater discrimination performance than does a lower ROC curve (ie, a smaller Az value). With the conventional binormal curve-fitting model, a decision variable that produces an Az value of 0.5 has the same median for true-positive and true-negative cases, whereas a decision variable with an Az value of 1.0 has distributions of true-positive and true-negative cases that are completely separated.
Statistical Z tests (27) were performed to evaluate the significance of the difference between the estimated Az value of a studied decision variable (eg, an individual computer-extracted feature or the LDA output) and an Az value of 0.5. In addition, statistical tests using the CLABROC algorithm (28) for correlated data were performed to evaluate the significance of the differences between the estimated Az values of any two decision variables (eg, computer-extracted features and LDA output) in their usefulness in differentiating between the low-risk and gene mutation cases in our databases.
In the current study, we first performed ROC analysis to evaluate the usefulness of the computer-extracted features and the linear discriminant function in differentiating between the 30 BRCA1 and BRCA2 mutation carriers and the 142 low-risk cases in the entire database. Further, to remove any bias due to patient age, we performed ROC analysis on the 30 BRCA1 and BRCA2 mutation carriers and the 60 age-matched low-risk cases.
|
RESULTS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODS RESULTSDISCUSSIONAPPENDIXREFERENCES |
|---|
|
|
|
|
|
|
|
|
|
|
|
DISCUSSION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODS RESULTSDISCUSSIONAPPENDIXREFERENCES |
|---|
It is interesting to note that the texture feature (coarseness) yielded the highest Az value among all four features (skewness, balance, coarseness, and contrast). Statistical tests in which the CLABROC algorithm (28) was used for correlated data showed that the differences in estimated Az values between coarseness and the other three features were not at a statistically significant level (P > .06). However, in differentiating between mammographic patterns of mutation carriers and low-risk women, the applicability level (Az value) of the four features considered together (with LDA) was significantly better (P < .001) than that of individual features on the basis of statistical analysis with the CLABROC algorithm (28).
In the current study, BRCA1 and BRCA2 gene mutation was the only risk factor considered, although there are many other factors associated with breast cancer risk. Among these, age is the most important, with the risk of developing breast cancer increasing with increasing age. To rule out possible bias due to the difference in age distributions of the mutation carriers and the low-risk cases, evaluation of computer analysis was also performed for an aged-matched group. Although the results obtained from the age-matched group were more relevant to this study, the similar Az values and mammographic characteristics obtained from the cases in both the entire group and the age-matched group suggest that the difference in age distributions between the BRCA1 and BRCA2 mutation carriers and low-risk cases did not have a strong influence on the performance of these individual features for this database. The BRCA1 and BRCA2 mutation carriers tended to have denser breasts than did the low-risk women. In addition, their mammographic patterns tended to be coarser, with lower contrast than those of the low-risk women.
As indicated by the Az values (0.91 and 0.92) obtained with LDA, our computerized classification method reached a high level of performance in differentiating between mammographic patterns of mutation carriers and those of low-risk women in both the entire group and the age-matched group. The lower and upper bounds of 95% CIs for the estimated ROC curve generated from the discriminant scores of the mutation carriers and the low-risk women provided evidence that our computerized method can be used for classification. We demonstrated that computerized analysis of mammographic patterns may enable identification of women at high risk for breast cancer. Identification of high-risk women may provide better opportunity for early detection of breast and/or ovarian cancer in these women.
In the current study, skewness was one of the four features used to characterize mammographic density patterns. We have shown that the use of both density and texture features (coarseness and contrast) significantly improved performance in differentiating between mutation carriers and low-risk cases (P < .001), as compared with the use of skewness alone. In general, breast parenchymal patterns can be described in part by the amount of density and in part by the heterogeneity and/or inhomogeneity of the parenchymal patterns. Thus, we believe that a combination of mammographic features will perform better than a single feature (ie, percentage density) in assessing risk. Misclassification of mutation carriers and low-risk women, as indicated by the overlap between the discriminant scores of the mutation carrier and low-risk groups in Figure 4, may result partially from the fact that some mutation carriers and low-risk women have similar mammographic patterns, as characterized by the computer features (Figs 2, 3) and by a radiologist (Fig 6).
In summary, the results of our evaluation suggest that BRCA1 and BRCA2 mutation carriers tend to have dense breasts, and their mammographic patterns tend to be low in contrast and coarse in texture. Analysis of mammograms by using multiple computer-extracted features yielded significantly better performance in differentiating between low-risk women and mutation carriers than that by using individual features only. Furthermore, our objective computerized method enables characterization of the mammographic patterns associated with breast cancer risk and may aid estimation of an individual’s risk of developing breast cancer. Also, such an objective analysis may allow clinicians to monitor changes in breast parenchyma after interventions to reduce breast cancer risk, which may potentially help in assessing the effectiveness of a chemotherapeutic prevention treatment on an individual basis.
|
APPENDIX |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODS RESULTSDISCUSSIONAPPENDIXREFERENCES |
|---|
|
|
|
Coarseness and contrast were first proposed by Amadasum and King (20) and were used by Tahoces et al (15) to characterize the Wolfe pattern. The mathematic definitions of the two texture features are given subsequently. The coarseness of a texture is defined by the amount of local gray-level variation. The contrast of a texture is defined by the amount of difference among all gray levels in the ROI and by the amount of local variation in gray level presented in the ROI. Notice that the contrast measure is determined with two terms: the gray-level differences in an ROI weighted by the amount of local variation. Thus, ROIs with similar gray-level differences may have different contrasts, depending on local variation in the ROIs. Conversely, ROIs with the same amount of local variation may have different contrast levels, depending on gray-level differences in the ROIs.
Coarseness (local uniformity [20]) is defined as follows:
|
|
|
|
(0,0) to exclude (x, y), W = (2d + 1)2, and d = 1. |
FOOTNOTES |
|---|
Abbreviations: Az = area under the ROC curve, LDA = linear discriminant analysis, ROC = receiver operating characteristic
Author contributions: Guarantors of integrity of entire study, Z.H., M.L.G.; study concepts, Z.H., O.I.O., M.L.G.; study design, Z.H., O.I.O., M.L.G., B.L.W.; literature research, Z.H., O.I.O.; clinical studies, W.Z., S.A.C.; data acquisition, all authors; data analysis/interpretation, Z.H., M.L.G.; statistical analysis, Z.H., C.E.M.; manuscript editing and preparation, Z.H., M.L.G.; manuscript definition of intellectual content, Z.H., M.L.G., O.I.O., C.E.M.; manuscript revision/review and final version approval, all authors.
|
REFERENCES |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODS RESULTSDISCUSSIONAPPENDIXREFERENCES |
|---|
This article has been cited by other articles:
|
|
 |
|
  G. J. R. Porter, A. J. Evans, E. J. Cornford, H. C. Burrell, J. J. James, A. H. S. Lee, and J. Chakrabarti Influence of Mammographic Parenchymal Pattern in Screening-Detected and Interval Invasive Breast Cancers on Pathologic Features, Mammographic Features, and Patient Survival Am. J. Roentgenol., March 1, 2007; 188(3): 676 - 683. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 G. Mitchell, A. C. Antoniou, R. Warren, S. Peock, J. Brown, R. Davies, J. Mattison, M. Cook, I. Warsi, D. G. Evans, et al. Mammographic Density and Breast Cancer Risk in BRCA1 and BRCA2 Mutation Carriers Cancer Res., February 1, 2006; 66(3): 1866 - 1872. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 M. Kriege, C. T.M. Brekelmans, C. Boetes, P. E. Besnard, H. M. Zonderland, I. M. Obdeijn, R. A. Manoliu, T. Kok, H. Peterse, M. M.A. Tilanus-Linthorst, et al. Efficacy of MRI and Mammography for Breast-Cancer Screening in Women with a Familial or Genetic Predisposition N. Engl. J. Med., July 29, 2004; 351(5): 427 - 437. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 J. Cuzick, J. Warwick, E. Pinney, R. M. L. Warren, and S. W. Duffy Tamoxifen and Breast Density in Women at Increased Risk of Breast Cancer J Natl Cancer Inst, April 21, 2004; 96(8): 621 - 628. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 J. A. Harvey and V. E. Bovbjerg Quantitative Assessment of Mammographic Breast Density: Relationship with Breast Cancer Risk Radiology, January 1, 2004; 230(1): 29 - 41. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
| RADIOLOGY | RADIOGRAPHICS | RSNA JOURNALS ONLINE |
) and the BRCA1 and BRCA2 mutation carriers (
) in the entire group (172 cases) in terms of (a) skewness versus coarseness and (b) coarseness versus contrast. The two-dimensional projection line from LDA corresponds to the line that yields the best (or maximum) separation between the BRCA mutation carriers and the low-risk cases in the corresponding two-dimensional feature space. As indicated by the diagonal line, the mutation carriers tended to have more dense breast tissue, with coarser and lower contrast texture patterns on their mammograms, than did the low-risk women.
