Different measures of mammographic images or density patterns 12345 have been used as surrogate endpoints for breast cancer in etiologic research 6789, in treatment 10111213 and in prevention trials 1415161718. Mammographic density is also important in the way in which it affects mammographic screening sensitivity 19. It has been suggested that women with high-risk density patterns should be screened more frequently and/or with additional views per breast 202122 or that they should be prescribed chemoprevention 23.

The first qualitative classification of mammographic density patterns was described by Wolfe in 1976 1. A modification of this method was proposed by Tabár 3. Wolfe also measured the percentage of the breast containing radiographic densities on a continuous scale with the use of a polar planimeter 24. A modification of the latter method is the BI-RADS system used in clinical radiology practice in the USA 2. The computer-assisted technique of measuring percentage mammographic densities developed by Boyd and colleagues 4 as well as that developed at the University of Southern California 5 are other methods of quantitative assessment.

The percentage density methods and the Wolfe patterns have consistently been shown to be strongly related to breast cancer risk in different populations 67152526272829, whereas the Tabár classification has so far only been shown to be related to breast cancer risk among Chinese women 30. Several risk factors for breast cancer have been shown to be associated with the different methods of density measurement 678930313233.

The Mammography and Breast Cancer Study is a research project using mammographic density patterns as surrogate endpoints for breast cancer among postmenopausal women attending the Norwegian Breast Cancer Screening Program in Tromsø. The purpose of this study was to classify mammograms according to four methods and to examine their agreement and their relationship to selected risk factors for breast cancer.

Table 1 shows the basic characteristics of the total study population and the women with mammograms in the high-risk categories. Reader GM classified mammograms with 28.3% density or more to be in the upper quartile; the corresponding figure for reader GU was 19.0% density. The value for the upper quartile was in between these, namely 21% density, when the dense area from reader GU and breast area for reader GM was used. In all, 47% and 24% of the mammograms were classified as high-risk according to the Wolfe (P2, DY) and Tabár (IV, V) classifications, respectively. Among the mammograms in the upper reader-specific quartiles, more than 95% were also classified as high-risk according to the Wolfe method, whereas the corresponding proportion for the Tabár method was less than 70% (Table 1).

The Pearson correlation coefficient was 0.93 and 0.86 for the repeated quantitative readings conducted 3 months (GM) and 18 months (GU) apart, respectively. The intra-rater agreement for the upper reader-specific quartile versus the three lower quartiles was moderate for reader GM (Kappa = 0.59; 95% CI 0.29 to 0.90) and reader GU (Kappa = 0.59; 95% CI 0.29 to 0.90). For reader NB the intra-reader agreement was good for the Wolfe classification (Kappa = 0.61; 95% CI 0.34 to 0.89) and very good for the Tabár classification (Kappa = 0.89; 95% CI 0.69 to 1.00). The Pearson correlation coefficient was 0.86 for the two original percentage density readings and 0.93 for the subset of 189 mammograms. Both the agreements between the reader-specific upper quartiles (crude Kappa = 0.69; 95% CI 0.64 to 0.74) and between all four quartiles (weighted Kappa = 0.71; 95% CI 0.68 to 0.74) were good. The agreement between high-risk and low-risk Wolfe and Tabár patterns was moderate (crude Kappa = 0.51; 95% CI 0.46 to 0.56).

Table 2 shows the distribution of mammograms according to the four Wolfe and five Tabár patterns with the corresponding median percentage density assessment for the two readers. According to reader GM the median percentage density for Wolfe high-risk patterns was 28.6% and for Tabár 34.9%. The corresponding figures for reader GU were 19.4% and 25.2%.

Table 3 shows that there was a significant inverse relation between age at screening, number of children and BMI and unfavorable density patterns according to all four classifications. Ages at menarche and menopause showed no association with any of the four outcome variables. Women who had their first child after their 25th birthday were more likely to have high-risk density patterns according to all four readings. A trend test achieved borderline significance for two of the assessments (GU (P = 0.07) and the Wolfe classification (P = 0.05)).

None of the associations displayed in the table between risk factors and the two quantitative readings changed when adjusted for the Wolfe or Tabár classification.

After adjustment for percentage density as assessed by either reader, age at entry (P = 0.14), number of children (P = 0.54) and BMI (P = 0.57) were no longer associated with the Wolfe classification. The association between age (P = 1.0) and the Tabár classification disappeared, whereas the inverse association with number of children (P < 0.05) and BMI (P < 0.03) remained statistically significant after adjustment for percentage density.

Our study finds that although the four methods vary depending on which mammograms are classified as unfavorable, all four methods seem to capture the same overall associations with risk factors for breast cancer. There was a high correlation for both inter-reader and intra-reader reliability between the two quantitative methods. Furthermore, mammograms classified as high-risk according to the Wolfe and Tabár classifications had the highest median percentage density. Age at screening, number of children and BMI were inversely associated with all four assessments of breast density. After adjustment for percentage density, there was no longer any association with these risk factors and the Wolfe pattern, whereas the association between number of children and BMI remained for the Tabár pattern. Controlling for the pattern classification did not change the associations between risk factors and percentage density.

All four methods of mammographic assessment have a subjective component. The percentage of dense area measurement is a ratio of two measures (dense area : total area). The mammograms in this study were rather dark, and differences in the appearance of the mammographic images due to the properties of the scanners did lead to variations in defining both areas. The two readers have both trained with Dr NF Boyd in Canada and have also worked together on previous studies 3536. Nevertheless, some of the discrepancy was due to differences in judgement as to what constitutes a dense area. However, the relative assessment between what constituted high-density or low-density mammograms was highly correlated. This supports the comparability of the readings.

The Kappa values for inter-rater agreement in our study are of the same magnitude as or better than those found in several other studies examining different kinds of mammographic reading. Venta and colleagues found a Kappa value of 0.46 for density measurements on X-ray and digital mammograms recorded by two radiologists 41. In a previous study including more than 3,500 premenopausal and postmenopausal women, we found the overall agreement between high-risk and low-risk for the Wolfe and Tabár classifications to be poor (Kappa = 0.22) 3 in comparison with the moderate agreement (Kappa = 0.51) in the present study. The two classifications are not strictly independent because one reader performed both assessments. We attribute the latter higher Kappa value to the fact that all women were postmenopausal, resulting in more low-density patterns, which are easier to assess.

Our results are in agreement with a recent study describing that the Wolfe pattern classification was redundant when percentage density was available as a measure for breast cancer risk 42. In contrast to this, the association between parity and BMI with the Tabár classification remained after adjustment for density, suggesting that this classification captures something more than just density assessments. However, we do not know whether this additional information from the Tabár classification is related to breast cancer risk. The association between age, parity, BMI and mammographic density is similar, as described in the literature 67; this attests to the validity of the mammographic measurements. The results also underscore the importance of adjusting for these factors when other associations are being explored. The positive association between age at first birth and unfavorable patterns indicated in the present study has also been found in some previous reports 382943444546, but not in others 47.

The present results are in agreement with studies finding no associations between ages at menarche 44 and menopause 47 and unfavorable density patterns, but not with others 845. We found a positive association between age at menarche and unfavorable density patterns among premenopausal women and an inverse association among postmenopausal women 8. In two other studies, a positive overall association was revealed 4547. In the study by El-Bastawissi and colleagues, age at menopause was positively related to unfavorable patterns 45. In our previous study the same relationship was found to be of borderline significance 8.

Our study has several strengths. It was a part of a population-based screening project with a high attendance rate. The three readers were experienced and blinded to the risk factors. A study from the Finnish screening program using the Wolfe classification showed that at younger ages there was a greater probability of misclassification from low-risk to high-risk and at older ages there was a greater probability of misclassification from high-risk to low-risk 48. Because the misclassification seems to be age dependent we also consider it a strength that all our women were postmenopausal. Our study is cross-sectional and does not have information on the temporal relationship between the risk factors examined and the mammographic density patterns. Several studies have shown that different types of hormone use will temporarily change mammographic density 495051 but that there will be a reversal of the hormone-induced changes on cessation of treatment 505152.

The high-risk classification of mammograms varied to some extent according to the four assessment methods commonly used previously. Our results indicate that the quantitative methods convey additional information over the qualitative methods. Quantitative measures should therefore be preferred when high-risk density patterns are used as surrogate endpoints in etiologic research, clinical or preventive trials. The difference between Tabár and Wolfe categories in relation to breast cancer risk and quantitative density assessment needs to be investigated further. Once breast cancer screening programs start to use digital mammograms, a quantitative method may become usable also in large-scale screening programs.

BMI = body mass index; CI = confidence interval; OR = odds ratio.

The authors declare that they have no competing interests.

ITG conceived the study, had full access to all of the data in the study, and takes responsibility for the integrity of the data and the accuracy of the data analyses. Classification of mammograms was performed by NB, GM and GU. Analysis and interpretation of data were performed by ITG, YB, GU, GM and EL. ITG and YB drafted the manuscript. ITG, YB, NB, GU, GM and EL revised the manuscript critically for important intellectual content. All authors read and approved the final manuscript.