The relationship of anthropometric measures to radiological features of the breast in premenopausal women.

We studied 273 premenopausal women recruited from mammography units who had different degrees of density of the breast parenchyma on mammography, in whom we measured height, weight and skinfold thicknesses. Mammograms were digitized to high spatial resolution by a scanning densitometer and images analysed to measure the area of dense tissue and the total area of the breast. Per cent density and the area of non-dense tissue were calculated from these measurements. We found that the mammographic measures had different associations with body size. Weight and the Quetelet index of obesity were strongly and positively associated with the area of non-dense tissue and with the total area of the breast, but less strongly and negatively correlated with the area of dense tissue. We also found a strong inverse relationship between the areas of radiologically dense and non-dense breast tissue. Statistical models containing anthropometric variables explained up to 8% of the variance in dense area, but explained up to 49% of the variance in non-dense area and 43% of variance in total area. These results suggest that aetiological studies in breast cancer that use mammographic density should consider dense and non-dense tissues separately. In addition to per cent density, methods should be examined that combine information from these two tissues. ImagesFigure 1

The radiographic appearance of the female breast -aries bet%veen individuals. owing to variations in the relative amounts of fat. connective tissue and epithelial tissue (Ingleby and Gerson-Cohen. 1960). Fat is radiological1v lucent. whereas connectixe and epithelial tissue are radiologically dense. These -ariations in the mammographic density of breast tissue are referred to as the parenchymal pattern of the breast.
An association between the mammographic parenchvmal pattern of the breast and risk of breast cancer has been reported by Wolfe (1 976a-c) and others. and has been the subject of reviews (Bovd et al. 1984: Oza andBovd. 1993): carefully conducted studies support an association between parenchymal pattems and breast cancer risk.
In studies that have classified mammographic densities quantitativelv. A omen with many extensive areas of density have consistentlyv been found to have an approximately fourto sixfold increased risk of breast cancer compared with women with no densie areas (Boyd et al. 1982. 1995a: Brisson et al. 1982. 1989: Wolfe et al. 1987: Saftlas et al. 1991: Byrne et al. 1995. The increased risk of breast cancer in women with manv areas of dense breast tissue is at least as large as. or larger than. is associated with most other known risk factors for the disease. Previous studies that hasve examined the relationship between mammographic densities and other breast cancer risk factors have consistently found an inverse association between body Aieight and per cent of the breast area occupied by radiologically dense Recerved 6 November 1997Revised 12 March 1998Accepted 18 March 1998 Correspondence to: NF Boyd. Division of Epidemiology and Statistics, Ontano Cancer Institute. 610 University Avenue. Toronto M5G 2M9. Canada breast tissue (Grove et al. 1979: Brisson et al. 1982: Graselle et al. 1982: Janzon et al. 1982: Wrhitehead et al. 1985: Boyd and McGuire. 1990: Bartovv et al. 1995: Boyd et al. 1995b).
The purpose of the present paper is to examine this association further. The subjects of this report are premenopausal %voomen. a group in which we have presviously noted an association betv een the Quetelet index and per cent density (Boyd et al. 1 995b).

PATIENTS AND METHODS General method
Information about risk factors for breast cancer was collected by questionnaire from w omen without breast cancer but w-ith different degrees of mammographic density. as assessed by radiologists as a percentage of the breast area on a fi e-point scale. The anthropometric variables of height. weight and skinfold thickness were measured at the time of intersviewi.
Method of sampling and classification of breast density Source of subjects The goal of the sampling procedure w-as to assemble premenopausal women w ith a wiide spectrum of mammographic densities. Subjects aged between 29 and 51 sears were identified between 1990 and 1992 from the mammographic units of St. Michael's and Mount Sinai Hospital in Toronto. Women wvere referred to these units for a variety of reasons. including suspicion of breast disease. the presence of risk factors such as a familx historx of breast cancer. or for routine examination. Breast density. as provisionally assessed by the radiologist w as definitivelx classified by quantitatis e methods that are described below-.
Method of recruitment Subjects identified in the manner described above were contacted by a letter that explained the goals and procedures of the study. This was followed by a phone call. during which their eligibility was determined. Subjects were eligible for the study if they were menstruating regularly. were not pregnant or breast feeding. had no previous history of cancer. had not had a hysterectomy or oopherectomy and were not scheduled to have breast surgery. All subjects taking any type of exogenous hormone preparation were excluded. Of the subjects contacted: 65% were eligible for inclusion in the study and. of these. 95% agreed to take part. The rate of participation did not vary according to extent of mammographic density. Subjects who agreed to enter the study were then visited in their homes by the study research assistant during the luteal phase of their menstrual cycle (between days 20 and 24). and the following measurements were made. Subjects were recruited into the present study only after mammograms had been taken. but the phase of the menstrual cycle during which mammograms were obtained was not recorded.

Anthropometnc measures
Each subject was weighed on a balance scale and measured for height. Skinfold thickness in the triceps. subscapular and iliac crest areas was measured using Lange calipers by a research assistant trained and certified by the Department of Athletics and Recreation. Universitv of Toronto. Canada.
Definitive classification of breast parenchymal pattem The measurements in the following analysis were made using a randomly selected. craniocaudal (viewing from above. down) mammographic view of one breast from each subject.
Mammograms were digitized and presented for analysis as an array of 675 x 925 pixels (0.0676 mm' per pixel). The manipulation of images and all calculations of the parameters to be described were performed on a Sun 4/260 workstation (Sun Microsystems. Mountain View. CA.USA). A Megavision 1024xm image processor/display (Megavision. Goleta. CA. USA) was used to present the images to the observer. An interactive density thresholding technique was used with a graphics overlay. in which an observer interactively highlighted a selected pixel value in colour by manipulation of a trackball. The process of measurement is illustrated in Figure 1.
The observer first selected a grey value as a threshold to separate the image of the breast from the background and determined the breast size. A second threshold was then selected to identify the edges of region(s) which are representative of radiographically dense tissue in the image. the sum of which gives the area of density in the breast. The proportion of the total area occupied by the radiographically dense tissue was calculated as the percentage of the entire projected area of the breast. expressed as per cent density. All thresholds were selected by one observer (NFB) who was unaware of anv of the characteristics of the subjects. Further details of this method are described elsewhere (Byng et al. 1994). We have found high levels of intraand inter-reader reliability with the measurement. In the present study. duplicates of a subset of the images were included as a check on reliability. which was found to be hiah with a test-retest correlation of 0.9 or greater.

Statistical analysis
Data analysis was carried out using the SAS statistical softs-are package (SAS Institute. 1989). Data were inspected for skewness  before analysis and. when necessarn. a transformation from the power family was applied. Details of the transformations used are given in the footnotes of the tables of results. The associations between anthropometric variables and mammographic measures were examined using Pearson's correlation coefficients. Multiple linear regression analysis and partial correlations were used to examine the relationship between each of the four measurements obtained from the mammogram and anthropometric variables after adjustment for other anthropometric variables. In addition. all models were controlled for age. age at menarche. parity and a family history of breast cancer. P-values < 0.05 were considered to be statistically significant.

Characteristics of subjects
Two hundred and seventy-three subjects were studied. All were premenopausal. with a median age of 43 years (range 29-51).  Distribution of mammographic features Figure 2 shows the distribution of the mammographic features included in the analyses that follow. The median area of the breast in the mammographic image was 101.0 cm' (IQR 63). the median area of dense tissue was 39.5 cm' (IQR 37.9) and the median area of non-dense tissue was 51.2 cm' (IQR 69.9).
Relationship between measures of mammographic features Table 1 shows the Pearson correlation coefficients between the measurements of areas of dense and non-dense tissues. total area and the per cent of the total area of the mammogram occupied by dense tissue. The total area of the breast was strongly correlated with the area of non-dense tissue (r = 0.801: P = 0.0001) and less strongly with the area of dense tissue (r = 0.1 17: P = 0.05). The per cent of the total area occupied by dense tissue was strongly correlated with both dense and non-dense areas. although in opposite directions (r = 0.753 and -0.886 respectively: P = 0.0001 for each). and the areas of dense and non-dense tissue were correlated inversely with each other (r = -0.456: P = 0.0001 ).  Height was not significantly associated with any of the mammographic measures. Weight and the Quetelet index were both strongly (positively) associated with the area of non-dense tissue r= 0.601: P = 0.0001: and r = 0.661: P = 0.0001 respectively) and with the total area of the breast (r = 0.584: P = 0.0001. and r = 0.623: P = 0.0001). but had a much weaker and negative corre- The association of skinfold thickness with mammographic measures. in general. resembled that of weight. All skinfold measures. and their sum. were strongly and positively correlated with total area and area of non-dense tissue. and negatively with the per cent area of dense tissue. These variables were also negatively. although less strongly. associated with the area of dense tissue.

Regression analysis and partial correlations of mammographic and anthropometric measures
Because height. weight and the Quetelet index are highly correlated. we examined their separate influences in a series of linear regression analyses (results are shown in Table 3). The partial correlations are given to show the magnitude of the association between each mammographic feature and anthropometric vanrable. after adjustment for the other variables in the model. All models were controlled for age. age at menarche. parity and family history of breast cancer. The R' is the variance in the mammographic measure explained by each regression model. The independent variables in each model were a subset of the anthropometric variables shown in the table. plus the variables we controlled for. and each measure obtained from the mammogram was the dependent variable. Because the skinfold thicknesses were all highly correlated with each other. we used only the sum of the three measures.
The anthropometric measures height (negatively). weight (positively) and the Quetelet index (positively) were all independently associated with the area of non-dense tissue. Body fat was statistically significant (positively) only in the model with height and weight. Similar associations were found with the total area of the breast. except that this measures was not independently associated with body fat.
Regression analysis showed that the anthropometric measures. w ith the variables for which we controlled and depending upon the model. accounted for between 46.6% and 48.4% of the variance in area of non-dense tissue and 42.5-42.7% of the variance in total breast area.
Weight and the Quetelet index were both negatively and significantly associated k ith the area of dense breast tissue. except w hen body fat (which is highly correlated with other indices of bodv size) w-as included in the model. although the correlation coefficients. in absolute value. were much smaller than was found between these variables and the other mammographic measures. The models containing anthropometric variables and the controlling variables accounted for between 7.0% and 7.9% of the variance in the area of mammographically dense tissue.
Per cent density was significantly and independently associated with height (positis ely). and with weight. body fat and the Quetelet index (negatively). Models containing these variables and the controlling variables accounted for between 30.5% and 33.2% of the Xariance in per cent density.
Although height was not significantly correlated with anv of the breast measurements in univariate analysis (Table 2). after controllinc for weight. as well as age. agye at menarche. parity and family history of breast cancer. it became significantly associated in multivariate analysis Awith total area. area of non-dense tissue (both negatively) and per cent density (positively).
Because the areas of dense and non-dense tissue were inxerselv correlated with each other. we next examined their influence on the regression analvsis of each other with anthropometric measures. The area of dense tissue was included in models given in Table 3 as an additional independent variable in the analysis. with area of non-dense tissue as a dependent variable. SimilarIl. the non-dense area was included among the independent variables. with area of dense tissue as the dependent variable. The results of these analyses are summarized in Table 4. When the non-dense area was included in the four regression models w ith dense area as the dependent variable it was highlI significant (P << 0.001). with partial correlations of -0.449 to -0.458. In both of the models that contained it. the Quetelet index w-as significant. Weight w-as also significant in both models in which it was included. As in the models shown in Table 3. neither height nor body fat was statisticallv significant in the models shown in Table 4. The X%ariance in the dense area explained by the regression model increased from approximately 7%7 to 26%7. after the inclusion of the area of nondense tissue as an independent variable.
Br;r,sC Jo,1na ov Carce 19980 78(9 1233-1238 W.Z. : -: .'Ca-ce, Pesea-c-Carroaq-1-3-98 e We also found a negative relationship between the dense and non-dense areas in the mammogram. Because the total breast area is composed of only dense and non-dense tissue, there must be a relationship between the per cent of the total area occupied by these types of tissue. Howvever. the actual areas of dense and nondense tissue might vary independently. and there is no reason to suppose they should be correlated. However. variations in the nondense area accounted for 21%7c of the variance in the area of dense tissue. and the negative association of the areas of these two tissue types suggests a common underlying mechanism related to their formation. As both the dense and non-dense area measurements involved the same measurement process. we explored the possibility that the observed dense/non-dense correlation was explained by correlated measurement errors for the two variables. A sample of 30 subjects had replicate measurements of these variables from four different observers. These data were used to model the measurement error. When measurement error was adjusted for in this sample. the observed dense/non-dense correlation changed only by 0.01. As this model for measurement error probably overestimates the error for a specific observer. we conclude that correlated errors cannot account for our results. Radiologically dense breast tissue is composed of fibrous stroma and epithelium. and non-dense tissue is composed of mainly fat. Several potential mechanisms exist to explain a quantitative relationship between the tissues responsible for the dense and non-dense radiological components of the breast. Adipocytes in the breast develop from preadipocytes that are part of the breast stroma and have the morphology of fibroblasts (Ailhaud et al. 1992). This terminal differentiation. which is associated with the accumulation of fat in adipocytes. may be associated with a reduction in the area of radiologically dense tissue in the mammogram and an increase in the area of radiolucent tissue. A number of interactions have been described between adipocytes and mammary epithelium. In vitro experiments show that adipocytes exert an influence on mammary epithelial cell proliferation. probably through an effect on extracellular components. and also promote epithelial cell differentiation (Roncari andHamilton. 1993: Xu andBjomtorp. 1987). Adipocytes. as well as epithelial and other stromal cells in the breast. are influenced by sex hormones. For example. the activity of lipoprotein lipase in adipocytes is controlled by progesterone. which also is thought to play a role in proliferation of epithelial cells in the breast (Xu and Bjorntorp. 1987).
These results indicate that the relationship of dense and nondense tissue areas in the mammogram should be examined separately in relation to other risk factors for breast cancer. and their associations with risk of the disease determined. Combining dense and non-dense areas into a single index of per cent dense tissue. as has been done to date in studies of breast cancer risk. may not be the optimal way of treating this information and alternatives should be examined.