Comparison of percent density from raw and processed full-field digital mammography data

Increased mammographic breast density is an established breast cancer risk factor [1‐3]. Irrespective of the method of measurement, the majority of studies have found a three- to sixfold increased risk of breast cancer in the highest vs. lowest density categories [2]. The majority of these studies estimated density from digitized film mammograms. Currently, more than 85 percent of the Mammography Quality Standards Act certified facilities operate FFDM units [4]. If breast density is to be used in the clinic for risk assessment and patient management, it is important to evaluate the performance of density estimated on full-field digital mammography (FFDM).

Film mammography and FFDM have similar diagnostic accuracy in breast cancer screening [5]. However, images acquired from these two forms of mammography may not appear similar on display [6]. FFDM systems can produce images in both the raw and clinical display representations, and these display formats differ across manufacturers. Clinical display images are processed (that is, enhanced) with algorithms developed by the respective manufacturers for improved diagnostic capability. Due to the size of images and storage considerations, the raw data is often discarded leaving only this processed data available for examination. Display variations may influence operator-assisted measurements of breast density. Currently, it is not known which FFDM representation is appropriate for percent density estimation.

In this report, we investigated the intra-measure reproducibility and the inter-measure correlation of percent density measures (PD) between raw and processed FFDM data representations. We also estimated and compared the associations of PD from both the raw and processed representations with breast cancer.

Cases and matched controls were of similar age and postmenopausal hormone use, as expected due to the matched design (Table 1). Cases had a higher BMI (26.6 vs. 25.3) and were more likely postmenopausal (79% vs. 73%) than controls. The mean PD was higher for cases than controls for both FFDM representations (P < 0.01 for the raw data and P < 0.03 for the processed data from the paired t test). Mean PD quantities were greater for the processed images than the raw images within cases (P < 0.11) and controls (P < 0.001), although differences were not significant among the cases (Table 1).

The intra-operator reproducibility for PD was high as evident from the concordance correlation of r_c = 0.92 for the raw images, and r_c = 0.87 for the processed images. The regression analyses from the repeated PD measurements also showed the degree of reproducibility with m = 0.96 ± 0.05, b = 0.76 and r = 0.92 for the raw images shown in Figure 1, and m = 0.81 ± 0.05, b = 1.3, and r = 0.89 for the processed images shown in Figure 2. The similarity between the respective r_c and r indicates that the first and second PD reproducibility measurement distributions have close agreement in the mean and variance for a given image representation. However, the deviation in the processed PD from the ideal regression parameter values (that is, m = 1 and b = 0) indicates that PD is better reproduced from the raw images.

The PD measurements from the processed clinical display images are linearly associated with the PD measurements from the raw images with r = 0.82. As shown in Figure 3, the processed PD is slightly elevated relative to the raw PD; the slope is significantly different from unity with m = 0.74 ± 0.03, and the intercept is significantly different from zero with b = 6.76. When categorizing each of the PD distributions into their respective quartiles (see Table 2 for cutoff values and ranges), the relationship between the two measurements shows moderate agreement, with a weighted kappa (κ) of 0.64 (95% CI: 0.59 to 0.70) [9]. The joint frequency quartile distribution for the two PD measures is provided and compared in Table 2. Of the 360 subjects, 217 (60.3%) were classified into the same quartiles by assessments made on their raw and processed images. Of the 143 with discrepant classifications, 131 (91.6%) differed by a single quartile (when applying the published Boyd categories [10], results were similar with 66% of the patients classified in the same category and 32.5% within one category shift, see Additional file 1).

The PD and breast cancer associations for both FFDM representations are summarized in Table 3 (raw) and Table 4 (processed). The ORs and AUCs for the associations of breast cancer with categorical (quartiles of PD based on controls, provided in Table 2) and continuous measures of PD (per standard deviation) are provided. Because our interest is in the performance of the Cumulus method applied to FFDM image formats, we focus on the simple BMI-adjusted models for the primary comparison. The associations and discriminatory accuracy of quartiles of PD from raw (OR: 1.0 (ref.), 2.6 (1.2 to 5.4), 3.1 (1.4 to 6.8), 4.7 (2.1 to 10.6); AUC = 0.63) and processed images (OR: 1.0 (ref.), 2.2 (1.1 to 4.1), 2.2 (1.1 to 4.4), 3.1 (1.5 to 6.6); AUC = 0.64) with breast cancer were similar. The associations of continuous PD and breast cancer for the raw (OR = 1.8 (1.3 to 2.4); AUC = 0.64) and processed (OR = 1.6 (1.2 to 2.2); AUC = 0.63) images were also similar. In all of the BMI-adjusted models, the ORs corresponding to the PD and breast cancer associations were significant (Tables 3 and 4). The BMI variable had the largest influence on the breast density associations, as expected [11]. The addition of breast area and menopausal status together strengthened (marginally) the PD associations for both representations. In the fully adjusted continuous PD models, the association of breast cancer with percent density from raw (OR = 1.8 (1.3 to 2.4)) and processed images (OR = 1.7 (1.2 to 2.3)) showed little difference from the BMI-adjusted models.

Secondarily, we compared the discriminatory accuracy (or AUC) of the ability of various models to separate the cases from the controls (Tables 3 and 4). Within either representation for both quartile and continuous unadjusted models, the AUC increase due to the inclusion of BMI was significant (P < 0.003), whereas the addition of the other covariates with BMI individually, or simultaneously, produced marginal AUC increases (P > 0.20). However, there was no statistical significant difference in AUC for PD from raw vs. processed images with breast cancer for either the unadjusted or adjusted models (P > 0.30).

Several of our findings merit particular comment specific to this FFDM technology. First, PD was reproducible whether assessed from raw or processed images but reproducibility was slightly higher for the raw representation. Second, PD assessed from the two FFDM representations was highly correlated, although PD measured from the processed images was slightly greater than from the raw images. And, earlier analysis showed that PD estimated from processed images (same type of FFDM unit used here) was less than that estimated from film [12]. These findings suggest that when merging mammograms from film and FFDM systems for breast density analyses, the choice of both film vs. FFDM as well as FFDM representation could impact the study. Much of the discrepancy between the two image representations of PD was accounted for by a one quartile shift. However, a small proportion of samples, approximately 4% (that is 12/360), showed a large discrepancy of two quartile shifts (Table 2). Further, our inter-FFDM measure agreement is in the range of inter-rater PD agreement using various FFDM models [13]. We demonstrated high intra-rater reproducibility of both the raw and processed FFDM PD, which is consistent with findings by other groups using different types of FFDM units [14]. Finally, this work showed that PD estimated from either representation was associated with breast cancer.

The inter-FFDM representation differences in our findings may be in part due to the pixel dynamic range compression (14 to 12 bit) or the two-stage mapping (that is, the mapping applied by the manufacturer) used to form the processed images from the raw images [15]. The detector response is linear in X-ray exposure for this system over a wide range of energies [16], which may also influence the raw image display. These differences may also be due to operator preference or natural variation (that is, each image requires its own window level and threshold adjustments). Because there are relatively few FFDM reports evaluating PD from FFDM, the generality of our findings will require further investigations.

The present results should be interpreted in light of certain limitations and qualifications. Limitations of the study include the relatively limited number of cases and controls, emphasizing the importance of replicating these findings in larger studies. Also, the use of mammograms at the time of diagnosis limits our ability to draw additional conclusions regarding temporality and risk. And, we were not able to examine or adjust for the perimenopausal category within menopausal status, due to the limited nature of the question; however, this should not influence the comparison between the two density measures. These findings apply to one specific FFDM design and to a Cumulus operator with considerable experience and preferences. Therefore, it will be necessary to conduct similar research on other FFDM technologies with multiple operators and larger datasets to ensure our results generalize. Similarly, the operator was not a clinician, and our results strictly apply in the research environment and not to the clinical setting per se. Although given an equally skilled operator, we would expect the findings to theoretically translate to the clinic. The detector size for the FFDM design (that is, the earliest FFDM technology) used in our study precludes its use on women with large breasts with one exposure [6], limiting the range of breast sizes for the cases and controls included in the study [7] and consequently our generalizability to women with large breasts. Our findings will require replication on newer FFDM models that do not have the size limitation. We observed slightly higher reproducibility in the analysis of the raw data than processed data, but this may reflect that the operator has more experience labeling raw data than processed data rather than any true difference. The reproducibility analysis was restricted to control images, which should not limit generalization because there are not large differences in the unadjusted case-control PD distributions. Moreover from our operator's experience, the very low-density images, which are less prevalent in cases, are more difficult to label (that is, in these situations the operator may perceive a range of acceptable control settings before the density estimation is performed) than high-density images. The reproducibility analysis was performed over a short time period, which could result in higher reproducibility compared to a longer span between evaluations. Finally, we did not have corresponding mammograms from film mammography units to serve as comparison to the raw and processed FFDM representations.

The Cumulus software was developed for digitized film data applications, not for FFDM. The Cumulus3 operating instructions indicate that the input data scale should parallel that of the display representation images (that is, larger pixel values should correspond with radiographically dense or bright image regions). It should be noted, the pixel scale is reversed for our raw images (that is, larger values corresponded with fatty area). This suggests that the raw images from this specific mammography unit are in the incorrect format and will not display properly within the Cumulus3 interface. To the contrary, we did not experience any difficulty reading the raw images without additional preparation. This raw image pixel scale was automatically inverted (within Cumulus) so that adipose regions were dark and glandular regions bright when reading the raw images into the Cumulus environment. Because we are uncertain as to why the raw images displayed properly within the Cumulus3 interface, in conflict with operating instructions, this work should be replicated on both similar and different FFDM designs as well. Notwithstanding these artifacts, the comparisons and findings presented in this report are internally valid.

In summary, PD from the two FFDM representations was reproducible and correlated. Further, the breast cancer associations were similar across the data representations and agree with those reported previously [2]. However, the raw data representation provided slightly better reproducibility. Thus, the raw data may be preferable for PD applications when possible. Although these results are encouraging, additional evaluation of prospective and larger study populations, operators, and system designs is required to confirm our findings.