Cancer Causes Control (2014) 25:1037–1043 DOI 10.1007/s10552-014-0404-4
ORIGINAL PAPER
Assessment of a fully automated, high-throughput mammographic density measurement tool for use with processed digital mammograms A. M. Couwenberg • H. M. Verkooijen • J. Li • R. M. Pijnappel • K. R. Charaghvandi M. Hartman • C. H. van Gils
•
Received: 24 January 2014 / Accepted: 20 May 2014 / Published online: 25 June 2014 Ó Springer International Publishing Switzerland 2014
Abstract Purpose The ImageJ model is a recently developed automated breast density measurement tool based on analysis of Cumulus outcomes. It has been validated on digitized film-screen mammograms. In this study, the ImageJ model was assessed on processed full-field digital mammograms and correlated with the Breast Imaging Reporting and Data System (BI-RADS) density classification. Also, the association with breast cancer risk factors is observed. Methods Women with mammographies between 2001 and 2011 at the University Medical Center Utrecht, The Netherlands were included. We composed a training set, read with Cumulus, for building the ImageJ model [n = 100 women,
A. M. Couwenberg University Medical Center Utrecht, Heidelberglaan 100, 3584 CX Utrecht, The Netherlands e-mail: [email protected] H. M. Verkooijen R. M. Pijnappel Department of Radiology, University Medical Center Utrecht, Utrecht, The Netherlands J. Li Human Genetics Lab, Genome Institute of Singapore, Singapore, Singapore K. R. Charaghvandi Department of Radiotherapy, University Medical Center Utrecht, Utrecht, The Netherlands M. Hartman Department of Surgery, National University of Singapore and National University Health System, Singapore, Singapore C. H. van Gils (&) Julius Center for Health Sciences and Primary Care, University Medical Centre Utrecht, Utrecht, The Netherlands e-mail: [email protected]
331 images; craniocaudal (CC) and mediolateral oblique (MLO) views, left and right] and a validation set for model assessment and correlation with the BI-RADS classification [n = 530 women, 1,977 images; average of available CC and MLO views, left and right]. Pearson product-moment correlation coefficient was used to compare Cumulus with ImageJ, Spearman correlation coefficient for ImageJ with BI-RADS density, and generalized linear models for association with breast cancer risk factors. Results The correlation between ImageJ and Cumulus in the training set was 0.90 [95 % confidence interval (CI) 0.86–0.93]. After application to the validation set, we observed a high correlation between ImageJ and the BIRADS readings (Spearman r = 0.86, 95 % CI 0.84–0.88). Women with higher density were significantly younger, more often premenopausal, had lower parity, more often a benign breast lesion or family history of breast cancer. Conclusions The ImageJ model can be used on processed digital mammograms. The measurements strongly correlate with Cumulus, the BI-RADS density classification, and breast cancer risk factors. Keywords Breast cancer Breast density Mammography Breast cancer risk factors
Introduction High mammographic density is a strong breast cancer risk factor [1, 2]. In addition, high density makes detection of cancer more difficult because of a masking effect [1, 3]. This knowledge has led to increasing interest in understanding and measuring mammographic density in the last few decades. In most countries, however, information on breast density is not routinely used in clinical practice or screening
123
1038
programs. There is no clear evidence yet on what the most effective strategy would be in women with high breast density for early detection of breast cancer [4–7]. Also, there is no gold-standard method for measuring density. In routine clinical care, BI-RADS density [8], also known as the American College of Radiology (ACR) classification, is commonly used by radiologists. It is a qualitative, fourcategory density classification, ranging from almost entirely fatty (BI-RADS 1) to extremely fibroglandular dense (BI-RADS 4) breasts [9]. The BI-RADS classification is a well-recognized method, though with modest interreader reproducibility (kappa statistic = 0.56) [10]. In research, quantitative methods are used to measure mammographic density. Semiautomated thresholding methods, such as Cumulus [11], have been most frequently used. These methods count the number of pixels in dense areas of the breast, as predefined by an observer. Afterwards, the percent density, absolute dense area, and nondense area are calculated. This is a reproducible method with an interclass correlation coefficient for trained readers of more than 0.9 [10] and correlates strongly with breast cancer risk [1]. Since these semiautomated thresholding methods are highly time consuming, and therefore not feasible for breast density measurements in clinical practice or screening programs, several automated density measurement tools have been developed [12–20]. Roughly, two groups of automated methods can be distinguished: thresholding-based tools [17, 18], which measure the size of the dense and fatty areas in the breast, and volumetric tools, which measure the volumes of the dense and nondense tissues, making use of information from extra parameters such as energy spectra for different anode target/filter materials [15, 19]. This information is stored automatically in the header of ‘‘raw’’ or unprocessed full-field digital mammograms (FFDMs) [15]. All automated methods correlate modestly or well with Cumulus and/or with breast cancer risk but either have not been validated for processed FFDM images [14, 17, 18, 20] or require unprocessed FFDM images [15]. Nowadays FFDM is widely available, but as unprocessed images are often not stored, there is a need for a reproducible, automatic density measurement method that can be used for commonly available processed images. In this study we assess an automated, high-throughput thresholding density measurement tool, named the ImageJ model (or ImageJ-based method), originally developed on digitized film-screen images in 2012 by a research group from Karolinska Institute, Sweden [17]. They used algorithms from ImageJ [21] public-domain Java-based image processing software (developed by the National Institutes of Health) to develop a model which mimics Cumulus density estimations. The ImageJ model showed high correlation with Cumulus in an external validation set in
123
Cancer Causes Control (2014) 25:1037–1043
postmenopausal women (r = 0.88, 95 % CI 0.87–0.89). Moreover, the association between breast density and breast cancer risk was equally strong when using this automated tool or using Cumulus density measurements [17]. Recently, the ImageJ model has been validated with Cumulus in a cohort of film-screen images of mainly premenopausal women and shown to be a strong predictor of risk when taking the left and right MLO readings together [22]. However, the ImageJ model density measurement tool has not been evaluated on processed FFDM images. The aim of this study is to assess the automated density ImageJ model with FFDM images of Dutch women and correlate these density estimations with the widely used BIRADS density classification. In addition, we investigate the association between established breast cancer risk factors and the automated density measurements.
Methods Study population The ImageJ model was tested/validated on images of women who visited the University Medical Center (UMC) Utrecht in The Netherlands between 2001 and 2011 for mammographic examination. Training and validation of the model were done on separate patient groups. Training set The training set included 100 women for whom 331 processed FFDM mammograms were available, including left and right, craniocaudal (CC) and mediolateral oblique (MLO) views. Most of the images (60.0 %) were taken at UMC Utrecht using a Hologic Selenia mammography system. The remaining mammograms were taken in a screening center or other hospital. The images were anonymized, with only information on age at mammography available (age range 26–86 years, 8 % missing age). Validation set The validation set was used for independent assessment of the ImageJ model by comparing its results with the BIRADS density classification. This consisted of 4,429 processed FFDM digital images taken during diagnostic workup and follow-up of 530 patients with benign or malignant breast lesions at the Radiology Department of UMC Utrecht, The Netherlands. All contralateral (MLO and CC) images taken at time of diagnosis (baseline) and follow-up were included. Information on patient characteristics was obtained from medical files: age ([50 years), reproductive characteristics, previous benign or malignant breast
Cancer Causes Control (2014) 25:1037–1043
disease, medical history, medication, and history of malignancy in family members. In case of breast cancer diagnosis following baseline mammography, tumor characteristics, performed treatment, and recurrence and survival rates were collected. Previously reported benign breast disease included mastopathy, cysts, fibroadenomas, granulomas, sclerosing adenosis, ductal hyperplasia, mastitis, and benign lesions of unknown type. Density readings User-assisted thresholding method (Cumulus) Mammographic density in the training set was first assessed using Cumulus software (University of Toronto, Canada) [11]. All mammograms were read by one trained reader (M.L.). Two thresholds were set by the reader: one to distinguish the breast area from the background, and a second to separate the dense from the nondense area. The dense area and total breast area were computed by the application, and the nondense area was computed by subtracting the dense area from the total breast area. Their sizes (cm2) were calculated by multiplying the number of pixels by the size of pixels in the particular area. The percent density was defined as the dense area divided by the breast area multiplied by 100. Batch files were created with randomly assigned patients. The intraclass correlation between the batches (n = 11) was 0.96, 0.94, and 1.00 for percent density, dense area, and breast area, respectively. BI-RADS classification Mammographic density in the validation set was first assessed by an expert breast radiologist (R.M.P.) using the BI-RADS density classification. According to this qualitative classification, mammographic density was divided into four categories: (1) almost entirely fat breasts, (2) scattered fibroglandular densities, (3) heterogeneously dense, and (4) extremely dense breasts [9]. BI-RADS was classified for baseline images (n = 530) and follow-up mammograms (n = 1,447). All patients had at least one follow-up mammogram from 6 months from baseline on, ranging to a maximum of nine performed mammograms. ImageJ-based density measurement method The information derived from the ImageJ automated image processing tool was used for making a statistical model to estimate the density. The training set assessed by Cumulus was used for building the model. Technical details have been described elsewhere [17]. In short, using the ImageJ tool, we first applied a preprocessing and automated thresholding plugin to the mammographic images.
1039
Preprocessing was done by removing the mammography view tags and extracting the breast region from the rest of the image. Next, automated thresholding methods were applied to separate the areas of dense tissue from the remaining area. A total of 15 thresholding methods were applied to the preprocessed image before and after subtracting the background. Areas identified as dense tissue were subdivided into smaller objects for measurement. During the second step in ImageJ, objects in the thresholded images were counted and measured. For each thresholding method, the measurements were obtained for the total breast as well as for the dense parts. Eventually, the output of ImageJ includes 1,008 measurements/variables. Not all of these variables are informative. Analysis is limited to 993 variables with fewer than 200 not a number (NaN) values. The remaining NaN values were replaced by a numerical value of zero, and principal component analysis was carried out on all 993 variables. In a third step, a model is built using R statistical software [22]. The model is based on the Cumulus percent density (PD) and uses the Cumulus estimations as functions of the principal components. The PDs were square-root transformed to create a normal distribution. The automated density measurements were correlated to the Cumulus estimations for validation of the model. Finally, a learning set was created within the training set for model selection. This step was used to evaluate the performance of the model before it was retrained on the full training set. To determine the high-throughput feature, we time-tracked the ImageJ method while converting and analyzing one mammogram. Statistical analysis R statistical software and SPSS (version 20) were used for statistical analysis. The square-root-transformed Cumulus estimations from the training set were plotted against the square-root-transformed ImageJ model estimations, and Pearson’s correlation coefficient (r) with accompanying 95 % confidence interval was computed for percent density, dense area, and total breast area. Using a Bland–Altman plot we present the agreement for percent density between the two methods by plotting the difference between the two measurements against the mean of the two with the measurements back-transformed to the original values. To compare the median percent density measures from Cumulus and the ImageJ model we used the Wilcoxon signed-rank test. To examine the reliability of the ImageJ model, the model was trained 10 times with the same training set and compared with the Cumulus results using the mean Pearson correlation coefficient. We also performed sensitivity analyses, retraining the model separately for CC and MLO images and Hologic and images
123
1040
Cancer Causes Control (2014) 25:1037–1043
Fig. 2 Bland–Altman plot of the difference between Cumulus and ImageJ breast density measurements in the training set. PD percent density (%) Fig. 1 Line of concordance between ImageJ outcomes of squareroot-transformed PD and Cumulus outcomes in the training set. CCC concordance correlation coefficient. PCC Pearson correlation coefficient
from unknown systems, and examined whether this led to different correlations between percent density measurements done with Cumulus or the ImageJ model. All FFDM images (n = 4,429) collected in the validation set were assessed using the ImageJ tool. As the BI-RADS classification as reported by the radiologist was based on all available views, the average ImageJ percent density was also computed for all these views. The correlation between the results from the ImageJ model and BI-RADS density classification was estimated by Spearman rho’s coefficient (q) with accompanying 95 % confidence interval. This was done separately for the baseline mammograms and the follow-up images. In addition, this comparison is visualized using box plots. Spearman correlations were also computed for the baseline images of pre- and postmenopausal women separately. For further validation, we investigated the association between breast cancer risk factors and the ImageJ model density measurements for the patients in our validation set, using generalized linear models. We examined this association with percent density divided in quintiles (Q1–Q5). We used a test for linearity for normally distributed data, a Jonckheere–Terpstra test for skewed continuous data, and chi-squared linear test for proportions.
Results ImageJ converted and analyzed each mammogram of the training set (331 FFDM images) within 20 s.
123
All these analyses were used to make a model. High concordance was observed for percent density measurements between Cumulus and the ImageJ model (Pearson correlation coefficient 0.90, 95 % CI 0.86–0.93) (Fig. 1). The Bland–Altman plot (Fig. 2) shows a mean difference in percent density of 0.1 % (95 % CI -0.8 to 1.1 %) with 95 % limits of agreement ranging from -10.1 to 10.3 %. There is a small absolute difference between the ImageJ model and Cumulus estimations with median of 2.30 % and interquartile range of 1.1–4.4 %. The median of the ImageJ model measured percent density (17.3 %) was practically similar to the Cumulus percent density median (17.7 %) (Wilcoxon signed-rank test z = -0.21, p = 0.98), as were the ranges of the ImageJ model (2.8–45.4 %) and the range of Cumulus (2.0–51.7 %). Regarding the variance of both methods, the ImageJ model was likely to give a lower density estimation than Cumulus (91.5 versus 119.8, respectively). High concordance between the measurements with Cumulus and the ImageJ model was also observed for the dense area (Pearson correlation coefficient 0.89, 95 % CI 0.85–0.92) and the total breast area (Pearson correlation coefficient 0.98, 95 % CI 0.97–0.98). To examine the reproducibility of the ImageJ model when assessing percent density, we trained the model 10 times with the same training set (n = 331). The newly trained models generated a mean Pearson correlation coefficient of 0.89, with a minimum of 0.86 and a maximum of 0.93. Model analysis was done for CC and MLO view, Hologic and unknown system, and age separately (Table 1). The PD correlation did not appear different
Cancer Causes Control (2014) 25:1037–1043
1041
Table 1 Model analysis: Pearson product-moment correlation coefficient of ImageJ and Cumulus Model (number of images)
Density
Pearson correlation (95 % CI)
All images (n = 331)
PD
0.90 (95 % CI 0.86–0.90)
Dense area
0.89 (95 % CI 0.85–0.90)
Breast area PD
0.98 (95 % CI 0.97–0.98) 0.87 (95 % CI 0.79–0.93)
Dense area
0.73 (95 % CI 0.57–0.84)
CC view (n = 149)
Breast area
0.97 (95 % CI 0.94–0.98)
PD
0.87 (95 % CI 0.78–0.92)
Dense area
0.83 (95 % CI 0.72–0.90)
Breast area
0.91 (95 % CI 0.84–0.95)
PD
0.90 (95 % CI 0.83–0.94)
Dense area
0.81 (95 % CI 0.69–0.89)
Breast area
0.97 (95 % CI 0.95–0.98)
Unknown systems (n = 133)
PD
0.86 (95 % CI 0.75–0.92)
Dense area
0.70 (95 % CI 0.49–0.80)
Breast area
0.96 (95 % CI 0.92–0.98)
Age, B50 years (n = 117)
PD
0.86 (95 % CI 0.75–0.92)
Table 2 Baseline characteristics of women in the validation set
Dense area
0.83 (95 % CI 0.70–0.91)
Characteristic
Breast area PD
0.95 (95 % CI 0.91–0.97) 0.86 (95 % CI 0.77–0.92)
Dense area
0.82 (95 % CI 0.71–0.89)
Breast area
0.97 (95 % CI 0.95–0.98)
MLO view (n = 169)
Hologic system (n = 198)
Age, [50 years (n = 187)
PD percent density
between CC and MLO models, or between younger and older women. The PD correlation for Hologic images was slightly higher than for mammograms from unknown manufacturers. For the dense area, correlations were slightly lower for the CC models and the models trained on mammograms from unknown manufacturers. In the validation set, the median BI-RADS classification was 2, corresponding to 28.3 % of the women, and the median PD of the ImageJ model was 14.0 % with a range of 1.0–53.5 % (Table 2, baseline characteristics). The correlation between the BI-RADS density classification and the ImageJ model PD for the baseline mammogram was high, with a Spearman’s q of 0.86 (95 % CI 0.84–0.88). The estimates from the mammograms obtained during the first follow-up (530 patients) and those obtained during the second follow-up (422 patients) showed very good correlation as well, with Spearman’s q = 0.85 (95 % CI 0.83–0.87) and q = 0.86 (95 % CI 0.83–0.88), respectively. The box plot (Fig. 3) shows a clear trend of increasing median percent density for higher BI-RADS categories: 6.4 % for BI-RADS 1, 13.8 % for BI-RADS 2, 21.7 % for BI-RADS 3, and 28.1 % for BI-RADS 4. Spearman correlations were not different in pre- versus postmenopausal women: 0.87 (95 % CI 0.80–0.91) and 0.84 (95 % CI 0.81–0.87), respectively.
Fig. 3 Box plot of BI-RADS density categories and ImageJ percent density (PD) in the validation set
Age at mammographya (years) Postmenopausal Number of children
Cases (n = 530) (% of total)
Missing data
59 (50–93) 80.6 % (353) 2 (0–11)
17.4 % (92) 17 % (90)
Parous Nulliparous
18.4 % (81)
1 child
14.1 % (62)
2 children 3? children
40.0 % (176) 27.5 % (121)
Breast cancer in first-degree relative
31.8 % (157)
History of benign breast disease
17.1 % (90)
7 % (37)
BI-RADS density 1
33.6 % (178)
2
28.3 % (150)
3 4
26.2 % (139) 11.9 % (63)
a
Median (range) for continuous variables; % (n) for categorical variables
Table 3 presents the associations between established breast cancer risk factors and the percent density measured by the ImageJ model. Of the women in the cohort, 80.6 % were postmenopausal. We were not able to distinguish between pre- and perimenopausal status. High breast density was significantly associated with younger age, fewer children or nulliparity, breast cancer in a first-degree relative, or previous history of benign breast lesions. As there was a large amount of missing data for some established
123
1042
Cancer Causes Control (2014) 25:1037–1043
Table 3 Association of breast cancer risk factors and breast density assessed by the ImageJ model Q1 (4.2 %)a n = 105
Q2 (8.6 %)a n = 107
Q3 (14 %)a n = 107
Q4 (20.4 %)a n = 106þ
Q5 (28.4 %)a n = 105
p trend
Mean (SD) Postmenopausal Nulliparous First-degree family history of BC History of benign breast lesion
80.0 % (84)
72.9 % (78)
78.5 % (84)
58.5 % (62)
41.9 % (44)
\0.001
7.6 % (8)
10.3 % (11)
15.0 % (16)
16.0 % (17)
27.6 % (29)
\0.001
29.5 % (31)
21.5 % (23)
23.4 % (25)
34.9 % (37)
39.1 % (41)
0.008
7.6 % (8)
10.3 % (11)
17.8 % (19)
23.6 % (25)
25.7 % (27)
0.001
Median (Q1–Q3) Age at mammography (years)
65 (57–71)
64 (55–71)
60 (55–68)
57 (51–64)
53 (51–59.5) \0.001
Number of children
2.0 (2.0–3.0)
2.0 (1.0–2.25)
2.0 (1.0–3.0)
2.0 (1.0–2.0)
1.0 (0–2.0)
\0.001
BC breast cancer, SD standard deviation a
Quintile (mean PD)
breast cancer risk factors such as young age at menopause, hormonal replacement therapy, and contraceptive pill use, these variables were excluded from analysis (Table 3).
Discussion In this study we showed that the ImageJ model, an automated density measurement tool recently developed based on Cumulus estimations, can be used on processed digital images. In addition, the ImageJ model measurements correlate well with Cumulus, the BI-RADS breast density classification, and established breast cancer risk factors. Other automated measurement tools have been developed recently [12–20]. All of them correlate modestly to well with Cumulus or the BI-RADS density classification, and most of them show a clear relationship with breast cancer risk. The majority of these methods have only been evaluated for use with film-screen mammograms [12, 14, 18–20, 22]. Methods that have been evaluated for use with FFDM images specifically need unprocessed mammograms, and the ones that have been validated for digital images make use of unprocessed mammograms, which are often not available in many centers [15]. Our study is the first to evaluate an automated tool on processed FFDM images. We were able to compare the ImageJ model PD estimations with the BI-RADS density classification and to study them in relation to breast cancer risk factors. No evaluation of breast cancer risk could be assessed because our data were collected from a hospital population consisting of women who underwent biopsy due to breast complaints or were referred after positive screening mammography. Women with high breast density have higher breast cancer risk but also higher chance of falsepositive outcome of screening mammography [23].
123
Therefore, comparing breast density between women with malignant versus benign biopsy outcome does not give a valid estimate of the density–breast cancer risk relationship. A study in an unselected screening population with FFDM and sufficient follow-up would provide the ultimate proof for the use of ImageJ model PD for predicting breast cancer risk using processed FFDM mammograms. Another limitation of this study concerns the limited information on breast cancer risk factors available from the medical charts of this hospital population. Nevertheless, we were able to show statistically significant associations with a number of established breast density or breast cancer risk factors such as age, menopausal status, parity, family history of breast cancer, and benign breast lesions in the past. Body mass index [BMI (kg/m2)], described to be inversely associated with percent mammographic density in the literature [24], was not routinely registered in our patients’ charts, therefore we could not examine this relationship in our study. The correlation between ImageJ model PD and Cumulus PD seemed to be independent of mammographic view or age category. Models trained on a mixture of mammograms from unknown manufacturers performed less well than models trained on mammograms from one known manufacturer. The correlation between the dense area measures seemed slightly lower for the CC than for the MLO mammograms. We do not have a good explanation for this. The variation in correlation between these models could possibly be influenced by number of images. The ImageJ model is based on a learning model affected by the total amount of trained images. We do not, however, see this same effect when studying dense area in the other subgroups of similar or smaller size, or when studying percent density. A strength of this study is the use of independent training and validation sets with large sample size. The training set was read by a single experienced Cumulus
Cancer Causes Control (2014) 25:1037–1043
reader (M.L.), and the BI-RADS density classification of the validation set was performed by one experienced breast radiologist (R.M.P.). The use of a single reader/radiologist prevented noise due to interobserver disagreement within the sets. Our results give a good indication that the ImageJ model can be used for processed digital images. Besides this finding, the automated method is high-throughput with an image conversion and analysis time of 20 s, making it convenient for measurement of large image sets. Furthermore, no extensive training is needed and only publicdomain software is required. There are important opportunities to use the ImageJ model for both research and clinical purposes. Acknowledgments This work was supported by the Agency for Science, Technology, and Research (A-STAR), Singapore under the 2nd Joint Council Office (JCO) Career Development Grant (13302EG065). We thank Mariette Lokate (M.L.) for providing a cohort of images with Cumulus density estimations for this study.
References 1. Boyd NF, Guo H, Martin LJ, Sun L, Stone J, Fishell E et al (2007) Mammographic density and the risk and detection of breast cancer. N Engl J Med 356(3):227–236 2. McCormack Va, dos Santos, Silva I (2006) Breast density and parenchymal patterns as markers of breast cancer risk: a metaanalysis. Cancer Epidemiol Biomark Prev 15(6):1159–1169 3. Kerlikowske K (2007) The mammogram that cried Wolfe. N Engl J Med 356(3):297–300 4. Nothacker M, Duda V, Hahn M, Warm M, Degenhardt F, Madjar H, Weinbrenner SAU (2009) Early detection of breast cancer: benefits and risks of supplemental breast ultrasound in asymptomatic women with mammographically dense breast tissue. A systematic review. BMC Cancer 20(9):335 5. Kavanagh AM, Byrnes GB, Nickson C, Cawson JN, Giles GG, Hopper JL, Gertig DMED (2008) Using mammographic density to improve breast cancer screening outcomes. Cancer Epidemiol Biomark 10:2818–2824 6. Benndorf M, Baltzer PA, Vag T, Gajda M, Runnebaum IBKW (2010) Breast MRI as an adjunct to mammography: does it really suffer from low specificity? A retrospective analysis stratified by mammographic BI-RADS classes. Acta Radiol 51:715–721 7. Berg WA, Zhang ZLD et al (2012) Detection of breast cancer with addition of annual screening ultrasound or a single screening MRI to mammography in women with elevated breast cancer risk. JAMA 307(13):1394–1404
1043 8. Reston V (1993) American College of Radiology: breast imaging reporting and data system (BIRADS). Am Coll Radiol 9. Yaffe MJ (2008) Mammographic density. Measurement of mammographic density. Breast Cancer Res 10(3):209 10. Boyd NFML (2011) Mammographic density and breast cancer risk: current understanding and future prospects. Breast Cancer Res 13:223 11. Byng JW, Boyd NF, Fishell E, Jong Ra, Yaffe MJ (1994) The quantitative analysis of mammographic densities. Phys Med Biol 39(10):1629–1638 12. Shepherd JA, Herve L, Landau J, Fan B, Kerlikowske KCS (2005) Novel use of single X-ray absorptiometry for measuring breast density. Technol Cancer Res Treat 4(2):173–182 13. Shepherd JA, Kerlikowske K, Ma L, Duewer F, Fan B, Wang J, Malkov S, Vittinghoff ECS (2011) Volume of mammographic density and risk of breast cancer. Cancer Epidemiol Biomark 20(7):1473–1482 14. Kallenberg MGJ (2011) Automatic breast density segmentation: an integration of different approaches. Phys Med Biol 56:2715–2729 15. Highnam R, Brady M, Yaffe MJKN (2010) Robust breast composition measurement—Volpara (TM). Digit Mammogr 6136:342–349 16. Van Engeland S, Snoeren PR, Huisman H, Boetes CKN (2006) Volumetric breast density estimation from full-field digital mammograms. IEEE Trans Med Imaging 25(3):273–282 17. Li J, Szekely L, Eriksson L, Heddson B, Sundbom A, Czene K et al (2012) High-throughput mammographic-density measurement: a tool for risk prediction of breast cancer. Breast Cancer Res 14(4):R114 18. Heine JJ, Carston MJ, Scott CG, Brandt KR, Wu F, Pankratz VS et al (2008) An automated approach for estimation of breast density. Cancer Epidemiol Biomark Prev 17:3090–3097 19. Pawluczyk O, Augustine BJ, Yaffe MJ, Rico D, Yang J, Mawdsley GEBN (2003) A volumetric method for estimation of breast density on digitized screen-film mammograms. Med Phys 30(3):352–364 20. Nickson C, Arzhaeva Y, Aitken Z, Elgindy T, Buckley M, Li M et al (2013) AutoDensity: an automated method to measure mammographic breast density that predicts breast cancer risk and screening outcomes. Breast Cancer Res 15(5):R80 21. ImageJ, U.S. National Institutes of Health, Bethesda, MD, USA. http://imagej.nih.gov/ij/ 22. Sovio U, Li J, Aitken Z, Humphreys K, Czene K, Moss S et al (2014) Comparison of fully and semi-automated area-based methods for measuring mammographic density and predicting breast cancer risk. Br J Cancer 20:1–9 23. Ihaka R, Gentlemen RR (1997) The R Project for Statistical Computing [Internet]. Statistics Department of the University of Auckland. http://www.r-project.org/ 24. Carney PA, Miglioretti DL, Yankaskas BC, Kerlikowske K, Rosenberg R, Rutter CM et al (2003) Individual and combined effects of age, breast density and hormone replacement therapy use on the accuracy of screening mammography. Ann Intern Med 138:168–175
123
ORIGINAL PAPER
Assessment of a fully automated, high-throughput mammographic density measurement tool for use with processed digital mammograms A. M. Couwenberg • H. M. Verkooijen • J. Li • R. M. Pijnappel • K. R. Charaghvandi M. Hartman • C. H. van Gils
•
Received: 24 January 2014 / Accepted: 20 May 2014 / Published online: 25 June 2014 Ó Springer International Publishing Switzerland 2014
Abstract Purpose The ImageJ model is a recently developed automated breast density measurement tool based on analysis of Cumulus outcomes. It has been validated on digitized film-screen mammograms. In this study, the ImageJ model was assessed on processed full-field digital mammograms and correlated with the Breast Imaging Reporting and Data System (BI-RADS) density classification. Also, the association with breast cancer risk factors is observed. Methods Women with mammographies between 2001 and 2011 at the University Medical Center Utrecht, The Netherlands were included. We composed a training set, read with Cumulus, for building the ImageJ model [n = 100 women,
A. M. Couwenberg University Medical Center Utrecht, Heidelberglaan 100, 3584 CX Utrecht, The Netherlands e-mail: [email protected] H. M. Verkooijen R. M. Pijnappel Department of Radiology, University Medical Center Utrecht, Utrecht, The Netherlands J. Li Human Genetics Lab, Genome Institute of Singapore, Singapore, Singapore K. R. Charaghvandi Department of Radiotherapy, University Medical Center Utrecht, Utrecht, The Netherlands M. Hartman Department of Surgery, National University of Singapore and National University Health System, Singapore, Singapore C. H. van Gils (&) Julius Center for Health Sciences and Primary Care, University Medical Centre Utrecht, Utrecht, The Netherlands e-mail: [email protected]
331 images; craniocaudal (CC) and mediolateral oblique (MLO) views, left and right] and a validation set for model assessment and correlation with the BI-RADS classification [n = 530 women, 1,977 images; average of available CC and MLO views, left and right]. Pearson product-moment correlation coefficient was used to compare Cumulus with ImageJ, Spearman correlation coefficient for ImageJ with BI-RADS density, and generalized linear models for association with breast cancer risk factors. Results The correlation between ImageJ and Cumulus in the training set was 0.90 [95 % confidence interval (CI) 0.86–0.93]. After application to the validation set, we observed a high correlation between ImageJ and the BIRADS readings (Spearman r = 0.86, 95 % CI 0.84–0.88). Women with higher density were significantly younger, more often premenopausal, had lower parity, more often a benign breast lesion or family history of breast cancer. Conclusions The ImageJ model can be used on processed digital mammograms. The measurements strongly correlate with Cumulus, the BI-RADS density classification, and breast cancer risk factors. Keywords Breast cancer Breast density Mammography Breast cancer risk factors
Introduction High mammographic density is a strong breast cancer risk factor [1, 2]. In addition, high density makes detection of cancer more difficult because of a masking effect [1, 3]. This knowledge has led to increasing interest in understanding and measuring mammographic density in the last few decades. In most countries, however, information on breast density is not routinely used in clinical practice or screening
123
1038
programs. There is no clear evidence yet on what the most effective strategy would be in women with high breast density for early detection of breast cancer [4–7]. Also, there is no gold-standard method for measuring density. In routine clinical care, BI-RADS density [8], also known as the American College of Radiology (ACR) classification, is commonly used by radiologists. It is a qualitative, fourcategory density classification, ranging from almost entirely fatty (BI-RADS 1) to extremely fibroglandular dense (BI-RADS 4) breasts [9]. The BI-RADS classification is a well-recognized method, though with modest interreader reproducibility (kappa statistic = 0.56) [10]. In research, quantitative methods are used to measure mammographic density. Semiautomated thresholding methods, such as Cumulus [11], have been most frequently used. These methods count the number of pixels in dense areas of the breast, as predefined by an observer. Afterwards, the percent density, absolute dense area, and nondense area are calculated. This is a reproducible method with an interclass correlation coefficient for trained readers of more than 0.9 [10] and correlates strongly with breast cancer risk [1]. Since these semiautomated thresholding methods are highly time consuming, and therefore not feasible for breast density measurements in clinical practice or screening programs, several automated density measurement tools have been developed [12–20]. Roughly, two groups of automated methods can be distinguished: thresholding-based tools [17, 18], which measure the size of the dense and fatty areas in the breast, and volumetric tools, which measure the volumes of the dense and nondense tissues, making use of information from extra parameters such as energy spectra for different anode target/filter materials [15, 19]. This information is stored automatically in the header of ‘‘raw’’ or unprocessed full-field digital mammograms (FFDMs) [15]. All automated methods correlate modestly or well with Cumulus and/or with breast cancer risk but either have not been validated for processed FFDM images [14, 17, 18, 20] or require unprocessed FFDM images [15]. Nowadays FFDM is widely available, but as unprocessed images are often not stored, there is a need for a reproducible, automatic density measurement method that can be used for commonly available processed images. In this study we assess an automated, high-throughput thresholding density measurement tool, named the ImageJ model (or ImageJ-based method), originally developed on digitized film-screen images in 2012 by a research group from Karolinska Institute, Sweden [17]. They used algorithms from ImageJ [21] public-domain Java-based image processing software (developed by the National Institutes of Health) to develop a model which mimics Cumulus density estimations. The ImageJ model showed high correlation with Cumulus in an external validation set in
123
Cancer Causes Control (2014) 25:1037–1043
postmenopausal women (r = 0.88, 95 % CI 0.87–0.89). Moreover, the association between breast density and breast cancer risk was equally strong when using this automated tool or using Cumulus density measurements [17]. Recently, the ImageJ model has been validated with Cumulus in a cohort of film-screen images of mainly premenopausal women and shown to be a strong predictor of risk when taking the left and right MLO readings together [22]. However, the ImageJ model density measurement tool has not been evaluated on processed FFDM images. The aim of this study is to assess the automated density ImageJ model with FFDM images of Dutch women and correlate these density estimations with the widely used BIRADS density classification. In addition, we investigate the association between established breast cancer risk factors and the automated density measurements.
Methods Study population The ImageJ model was tested/validated on images of women who visited the University Medical Center (UMC) Utrecht in The Netherlands between 2001 and 2011 for mammographic examination. Training and validation of the model were done on separate patient groups. Training set The training set included 100 women for whom 331 processed FFDM mammograms were available, including left and right, craniocaudal (CC) and mediolateral oblique (MLO) views. Most of the images (60.0 %) were taken at UMC Utrecht using a Hologic Selenia mammography system. The remaining mammograms were taken in a screening center or other hospital. The images were anonymized, with only information on age at mammography available (age range 26–86 years, 8 % missing age). Validation set The validation set was used for independent assessment of the ImageJ model by comparing its results with the BIRADS density classification. This consisted of 4,429 processed FFDM digital images taken during diagnostic workup and follow-up of 530 patients with benign or malignant breast lesions at the Radiology Department of UMC Utrecht, The Netherlands. All contralateral (MLO and CC) images taken at time of diagnosis (baseline) and follow-up were included. Information on patient characteristics was obtained from medical files: age ([50 years), reproductive characteristics, previous benign or malignant breast
Cancer Causes Control (2014) 25:1037–1043
disease, medical history, medication, and history of malignancy in family members. In case of breast cancer diagnosis following baseline mammography, tumor characteristics, performed treatment, and recurrence and survival rates were collected. Previously reported benign breast disease included mastopathy, cysts, fibroadenomas, granulomas, sclerosing adenosis, ductal hyperplasia, mastitis, and benign lesions of unknown type. Density readings User-assisted thresholding method (Cumulus) Mammographic density in the training set was first assessed using Cumulus software (University of Toronto, Canada) [11]. All mammograms were read by one trained reader (M.L.). Two thresholds were set by the reader: one to distinguish the breast area from the background, and a second to separate the dense from the nondense area. The dense area and total breast area were computed by the application, and the nondense area was computed by subtracting the dense area from the total breast area. Their sizes (cm2) were calculated by multiplying the number of pixels by the size of pixels in the particular area. The percent density was defined as the dense area divided by the breast area multiplied by 100. Batch files were created with randomly assigned patients. The intraclass correlation between the batches (n = 11) was 0.96, 0.94, and 1.00 for percent density, dense area, and breast area, respectively. BI-RADS classification Mammographic density in the validation set was first assessed by an expert breast radiologist (R.M.P.) using the BI-RADS density classification. According to this qualitative classification, mammographic density was divided into four categories: (1) almost entirely fat breasts, (2) scattered fibroglandular densities, (3) heterogeneously dense, and (4) extremely dense breasts [9]. BI-RADS was classified for baseline images (n = 530) and follow-up mammograms (n = 1,447). All patients had at least one follow-up mammogram from 6 months from baseline on, ranging to a maximum of nine performed mammograms. ImageJ-based density measurement method The information derived from the ImageJ automated image processing tool was used for making a statistical model to estimate the density. The training set assessed by Cumulus was used for building the model. Technical details have been described elsewhere [17]. In short, using the ImageJ tool, we first applied a preprocessing and automated thresholding plugin to the mammographic images.
1039
Preprocessing was done by removing the mammography view tags and extracting the breast region from the rest of the image. Next, automated thresholding methods were applied to separate the areas of dense tissue from the remaining area. A total of 15 thresholding methods were applied to the preprocessed image before and after subtracting the background. Areas identified as dense tissue were subdivided into smaller objects for measurement. During the second step in ImageJ, objects in the thresholded images were counted and measured. For each thresholding method, the measurements were obtained for the total breast as well as for the dense parts. Eventually, the output of ImageJ includes 1,008 measurements/variables. Not all of these variables are informative. Analysis is limited to 993 variables with fewer than 200 not a number (NaN) values. The remaining NaN values were replaced by a numerical value of zero, and principal component analysis was carried out on all 993 variables. In a third step, a model is built using R statistical software [22]. The model is based on the Cumulus percent density (PD) and uses the Cumulus estimations as functions of the principal components. The PDs were square-root transformed to create a normal distribution. The automated density measurements were correlated to the Cumulus estimations for validation of the model. Finally, a learning set was created within the training set for model selection. This step was used to evaluate the performance of the model before it was retrained on the full training set. To determine the high-throughput feature, we time-tracked the ImageJ method while converting and analyzing one mammogram. Statistical analysis R statistical software and SPSS (version 20) were used for statistical analysis. The square-root-transformed Cumulus estimations from the training set were plotted against the square-root-transformed ImageJ model estimations, and Pearson’s correlation coefficient (r) with accompanying 95 % confidence interval was computed for percent density, dense area, and total breast area. Using a Bland–Altman plot we present the agreement for percent density between the two methods by plotting the difference between the two measurements against the mean of the two with the measurements back-transformed to the original values. To compare the median percent density measures from Cumulus and the ImageJ model we used the Wilcoxon signed-rank test. To examine the reliability of the ImageJ model, the model was trained 10 times with the same training set and compared with the Cumulus results using the mean Pearson correlation coefficient. We also performed sensitivity analyses, retraining the model separately for CC and MLO images and Hologic and images
123
1040
Cancer Causes Control (2014) 25:1037–1043
Fig. 2 Bland–Altman plot of the difference between Cumulus and ImageJ breast density measurements in the training set. PD percent density (%) Fig. 1 Line of concordance between ImageJ outcomes of squareroot-transformed PD and Cumulus outcomes in the training set. CCC concordance correlation coefficient. PCC Pearson correlation coefficient
from unknown systems, and examined whether this led to different correlations between percent density measurements done with Cumulus or the ImageJ model. All FFDM images (n = 4,429) collected in the validation set were assessed using the ImageJ tool. As the BI-RADS classification as reported by the radiologist was based on all available views, the average ImageJ percent density was also computed for all these views. The correlation between the results from the ImageJ model and BI-RADS density classification was estimated by Spearman rho’s coefficient (q) with accompanying 95 % confidence interval. This was done separately for the baseline mammograms and the follow-up images. In addition, this comparison is visualized using box plots. Spearman correlations were also computed for the baseline images of pre- and postmenopausal women separately. For further validation, we investigated the association between breast cancer risk factors and the ImageJ model density measurements for the patients in our validation set, using generalized linear models. We examined this association with percent density divided in quintiles (Q1–Q5). We used a test for linearity for normally distributed data, a Jonckheere–Terpstra test for skewed continuous data, and chi-squared linear test for proportions.
Results ImageJ converted and analyzed each mammogram of the training set (331 FFDM images) within 20 s.
123
All these analyses were used to make a model. High concordance was observed for percent density measurements between Cumulus and the ImageJ model (Pearson correlation coefficient 0.90, 95 % CI 0.86–0.93) (Fig. 1). The Bland–Altman plot (Fig. 2) shows a mean difference in percent density of 0.1 % (95 % CI -0.8 to 1.1 %) with 95 % limits of agreement ranging from -10.1 to 10.3 %. There is a small absolute difference between the ImageJ model and Cumulus estimations with median of 2.30 % and interquartile range of 1.1–4.4 %. The median of the ImageJ model measured percent density (17.3 %) was practically similar to the Cumulus percent density median (17.7 %) (Wilcoxon signed-rank test z = -0.21, p = 0.98), as were the ranges of the ImageJ model (2.8–45.4 %) and the range of Cumulus (2.0–51.7 %). Regarding the variance of both methods, the ImageJ model was likely to give a lower density estimation than Cumulus (91.5 versus 119.8, respectively). High concordance between the measurements with Cumulus and the ImageJ model was also observed for the dense area (Pearson correlation coefficient 0.89, 95 % CI 0.85–0.92) and the total breast area (Pearson correlation coefficient 0.98, 95 % CI 0.97–0.98). To examine the reproducibility of the ImageJ model when assessing percent density, we trained the model 10 times with the same training set (n = 331). The newly trained models generated a mean Pearson correlation coefficient of 0.89, with a minimum of 0.86 and a maximum of 0.93. Model analysis was done for CC and MLO view, Hologic and unknown system, and age separately (Table 1). The PD correlation did not appear different
Cancer Causes Control (2014) 25:1037–1043
1041
Table 1 Model analysis: Pearson product-moment correlation coefficient of ImageJ and Cumulus Model (number of images)
Density
Pearson correlation (95 % CI)
All images (n = 331)
PD
0.90 (95 % CI 0.86–0.90)
Dense area
0.89 (95 % CI 0.85–0.90)
Breast area PD
0.98 (95 % CI 0.97–0.98) 0.87 (95 % CI 0.79–0.93)
Dense area
0.73 (95 % CI 0.57–0.84)
CC view (n = 149)
Breast area
0.97 (95 % CI 0.94–0.98)
PD
0.87 (95 % CI 0.78–0.92)
Dense area
0.83 (95 % CI 0.72–0.90)
Breast area
0.91 (95 % CI 0.84–0.95)
PD
0.90 (95 % CI 0.83–0.94)
Dense area
0.81 (95 % CI 0.69–0.89)
Breast area
0.97 (95 % CI 0.95–0.98)
Unknown systems (n = 133)
PD
0.86 (95 % CI 0.75–0.92)
Dense area
0.70 (95 % CI 0.49–0.80)
Breast area
0.96 (95 % CI 0.92–0.98)
Age, B50 years (n = 117)
PD
0.86 (95 % CI 0.75–0.92)
Table 2 Baseline characteristics of women in the validation set
Dense area
0.83 (95 % CI 0.70–0.91)
Characteristic
Breast area PD
0.95 (95 % CI 0.91–0.97) 0.86 (95 % CI 0.77–0.92)
Dense area
0.82 (95 % CI 0.71–0.89)
Breast area
0.97 (95 % CI 0.95–0.98)
MLO view (n = 169)
Hologic system (n = 198)
Age, [50 years (n = 187)
PD percent density
between CC and MLO models, or between younger and older women. The PD correlation for Hologic images was slightly higher than for mammograms from unknown manufacturers. For the dense area, correlations were slightly lower for the CC models and the models trained on mammograms from unknown manufacturers. In the validation set, the median BI-RADS classification was 2, corresponding to 28.3 % of the women, and the median PD of the ImageJ model was 14.0 % with a range of 1.0–53.5 % (Table 2, baseline characteristics). The correlation between the BI-RADS density classification and the ImageJ model PD for the baseline mammogram was high, with a Spearman’s q of 0.86 (95 % CI 0.84–0.88). The estimates from the mammograms obtained during the first follow-up (530 patients) and those obtained during the second follow-up (422 patients) showed very good correlation as well, with Spearman’s q = 0.85 (95 % CI 0.83–0.87) and q = 0.86 (95 % CI 0.83–0.88), respectively. The box plot (Fig. 3) shows a clear trend of increasing median percent density for higher BI-RADS categories: 6.4 % for BI-RADS 1, 13.8 % for BI-RADS 2, 21.7 % for BI-RADS 3, and 28.1 % for BI-RADS 4. Spearman correlations were not different in pre- versus postmenopausal women: 0.87 (95 % CI 0.80–0.91) and 0.84 (95 % CI 0.81–0.87), respectively.
Fig. 3 Box plot of BI-RADS density categories and ImageJ percent density (PD) in the validation set
Age at mammographya (years) Postmenopausal Number of children
Cases (n = 530) (% of total)
Missing data
59 (50–93) 80.6 % (353) 2 (0–11)
17.4 % (92) 17 % (90)
Parous Nulliparous
18.4 % (81)
1 child
14.1 % (62)
2 children 3? children
40.0 % (176) 27.5 % (121)
Breast cancer in first-degree relative
31.8 % (157)
History of benign breast disease
17.1 % (90)
7 % (37)
BI-RADS density 1
33.6 % (178)
2
28.3 % (150)
3 4
26.2 % (139) 11.9 % (63)
a
Median (range) for continuous variables; % (n) for categorical variables
Table 3 presents the associations between established breast cancer risk factors and the percent density measured by the ImageJ model. Of the women in the cohort, 80.6 % were postmenopausal. We were not able to distinguish between pre- and perimenopausal status. High breast density was significantly associated with younger age, fewer children or nulliparity, breast cancer in a first-degree relative, or previous history of benign breast lesions. As there was a large amount of missing data for some established
123
1042
Cancer Causes Control (2014) 25:1037–1043
Table 3 Association of breast cancer risk factors and breast density assessed by the ImageJ model Q1 (4.2 %)a n = 105
Q2 (8.6 %)a n = 107
Q3 (14 %)a n = 107
Q4 (20.4 %)a n = 106þ
Q5 (28.4 %)a n = 105
p trend
Mean (SD) Postmenopausal Nulliparous First-degree family history of BC History of benign breast lesion
80.0 % (84)
72.9 % (78)
78.5 % (84)
58.5 % (62)
41.9 % (44)
\0.001
7.6 % (8)
10.3 % (11)
15.0 % (16)
16.0 % (17)
27.6 % (29)
\0.001
29.5 % (31)
21.5 % (23)
23.4 % (25)
34.9 % (37)
39.1 % (41)
0.008
7.6 % (8)
10.3 % (11)
17.8 % (19)
23.6 % (25)
25.7 % (27)
0.001
Median (Q1–Q3) Age at mammography (years)
65 (57–71)
64 (55–71)
60 (55–68)
57 (51–64)
53 (51–59.5) \0.001
Number of children
2.0 (2.0–3.0)
2.0 (1.0–2.25)
2.0 (1.0–3.0)
2.0 (1.0–2.0)
1.0 (0–2.0)
\0.001
BC breast cancer, SD standard deviation a
Quintile (mean PD)
breast cancer risk factors such as young age at menopause, hormonal replacement therapy, and contraceptive pill use, these variables were excluded from analysis (Table 3).
Discussion In this study we showed that the ImageJ model, an automated density measurement tool recently developed based on Cumulus estimations, can be used on processed digital images. In addition, the ImageJ model measurements correlate well with Cumulus, the BI-RADS breast density classification, and established breast cancer risk factors. Other automated measurement tools have been developed recently [12–20]. All of them correlate modestly to well with Cumulus or the BI-RADS density classification, and most of them show a clear relationship with breast cancer risk. The majority of these methods have only been evaluated for use with film-screen mammograms [12, 14, 18–20, 22]. Methods that have been evaluated for use with FFDM images specifically need unprocessed mammograms, and the ones that have been validated for digital images make use of unprocessed mammograms, which are often not available in many centers [15]. Our study is the first to evaluate an automated tool on processed FFDM images. We were able to compare the ImageJ model PD estimations with the BI-RADS density classification and to study them in relation to breast cancer risk factors. No evaluation of breast cancer risk could be assessed because our data were collected from a hospital population consisting of women who underwent biopsy due to breast complaints or were referred after positive screening mammography. Women with high breast density have higher breast cancer risk but also higher chance of falsepositive outcome of screening mammography [23].
123
Therefore, comparing breast density between women with malignant versus benign biopsy outcome does not give a valid estimate of the density–breast cancer risk relationship. A study in an unselected screening population with FFDM and sufficient follow-up would provide the ultimate proof for the use of ImageJ model PD for predicting breast cancer risk using processed FFDM mammograms. Another limitation of this study concerns the limited information on breast cancer risk factors available from the medical charts of this hospital population. Nevertheless, we were able to show statistically significant associations with a number of established breast density or breast cancer risk factors such as age, menopausal status, parity, family history of breast cancer, and benign breast lesions in the past. Body mass index [BMI (kg/m2)], described to be inversely associated with percent mammographic density in the literature [24], was not routinely registered in our patients’ charts, therefore we could not examine this relationship in our study. The correlation between ImageJ model PD and Cumulus PD seemed to be independent of mammographic view or age category. Models trained on a mixture of mammograms from unknown manufacturers performed less well than models trained on mammograms from one known manufacturer. The correlation between the dense area measures seemed slightly lower for the CC than for the MLO mammograms. We do not have a good explanation for this. The variation in correlation between these models could possibly be influenced by number of images. The ImageJ model is based on a learning model affected by the total amount of trained images. We do not, however, see this same effect when studying dense area in the other subgroups of similar or smaller size, or when studying percent density. A strength of this study is the use of independent training and validation sets with large sample size. The training set was read by a single experienced Cumulus
Cancer Causes Control (2014) 25:1037–1043
reader (M.L.), and the BI-RADS density classification of the validation set was performed by one experienced breast radiologist (R.M.P.). The use of a single reader/radiologist prevented noise due to interobserver disagreement within the sets. Our results give a good indication that the ImageJ model can be used for processed digital images. Besides this finding, the automated method is high-throughput with an image conversion and analysis time of 20 s, making it convenient for measurement of large image sets. Furthermore, no extensive training is needed and only publicdomain software is required. There are important opportunities to use the ImageJ model for both research and clinical purposes. Acknowledgments This work was supported by the Agency for Science, Technology, and Research (A-STAR), Singapore under the 2nd Joint Council Office (JCO) Career Development Grant (13302EG065). We thank Mariette Lokate (M.L.) for providing a cohort of images with Cumulus density estimations for this study.
References 1. Boyd NF, Guo H, Martin LJ, Sun L, Stone J, Fishell E et al (2007) Mammographic density and the risk and detection of breast cancer. N Engl J Med 356(3):227–236 2. McCormack Va, dos Santos, Silva I (2006) Breast density and parenchymal patterns as markers of breast cancer risk: a metaanalysis. Cancer Epidemiol Biomark Prev 15(6):1159–1169 3. Kerlikowske K (2007) The mammogram that cried Wolfe. N Engl J Med 356(3):297–300 4. Nothacker M, Duda V, Hahn M, Warm M, Degenhardt F, Madjar H, Weinbrenner SAU (2009) Early detection of breast cancer: benefits and risks of supplemental breast ultrasound in asymptomatic women with mammographically dense breast tissue. A systematic review. BMC Cancer 20(9):335 5. Kavanagh AM, Byrnes GB, Nickson C, Cawson JN, Giles GG, Hopper JL, Gertig DMED (2008) Using mammographic density to improve breast cancer screening outcomes. Cancer Epidemiol Biomark 10:2818–2824 6. Benndorf M, Baltzer PA, Vag T, Gajda M, Runnebaum IBKW (2010) Breast MRI as an adjunct to mammography: does it really suffer from low specificity? A retrospective analysis stratified by mammographic BI-RADS classes. Acta Radiol 51:715–721 7. Berg WA, Zhang ZLD et al (2012) Detection of breast cancer with addition of annual screening ultrasound or a single screening MRI to mammography in women with elevated breast cancer risk. JAMA 307(13):1394–1404
1043 8. Reston V (1993) American College of Radiology: breast imaging reporting and data system (BIRADS). Am Coll Radiol 9. Yaffe MJ (2008) Mammographic density. Measurement of mammographic density. Breast Cancer Res 10(3):209 10. Boyd NFML (2011) Mammographic density and breast cancer risk: current understanding and future prospects. Breast Cancer Res 13:223 11. Byng JW, Boyd NF, Fishell E, Jong Ra, Yaffe MJ (1994) The quantitative analysis of mammographic densities. Phys Med Biol 39(10):1629–1638 12. Shepherd JA, Herve L, Landau J, Fan B, Kerlikowske KCS (2005) Novel use of single X-ray absorptiometry for measuring breast density. Technol Cancer Res Treat 4(2):173–182 13. Shepherd JA, Kerlikowske K, Ma L, Duewer F, Fan B, Wang J, Malkov S, Vittinghoff ECS (2011) Volume of mammographic density and risk of breast cancer. Cancer Epidemiol Biomark 20(7):1473–1482 14. Kallenberg MGJ (2011) Automatic breast density segmentation: an integration of different approaches. Phys Med Biol 56:2715–2729 15. Highnam R, Brady M, Yaffe MJKN (2010) Robust breast composition measurement—Volpara (TM). Digit Mammogr 6136:342–349 16. Van Engeland S, Snoeren PR, Huisman H, Boetes CKN (2006) Volumetric breast density estimation from full-field digital mammograms. IEEE Trans Med Imaging 25(3):273–282 17. Li J, Szekely L, Eriksson L, Heddson B, Sundbom A, Czene K et al (2012) High-throughput mammographic-density measurement: a tool for risk prediction of breast cancer. Breast Cancer Res 14(4):R114 18. Heine JJ, Carston MJ, Scott CG, Brandt KR, Wu F, Pankratz VS et al (2008) An automated approach for estimation of breast density. Cancer Epidemiol Biomark Prev 17:3090–3097 19. Pawluczyk O, Augustine BJ, Yaffe MJ, Rico D, Yang J, Mawdsley GEBN (2003) A volumetric method for estimation of breast density on digitized screen-film mammograms. Med Phys 30(3):352–364 20. Nickson C, Arzhaeva Y, Aitken Z, Elgindy T, Buckley M, Li M et al (2013) AutoDensity: an automated method to measure mammographic breast density that predicts breast cancer risk and screening outcomes. Breast Cancer Res 15(5):R80 21. ImageJ, U.S. National Institutes of Health, Bethesda, MD, USA. http://imagej.nih.gov/ij/ 22. Sovio U, Li J, Aitken Z, Humphreys K, Czene K, Moss S et al (2014) Comparison of fully and semi-automated area-based methods for measuring mammographic density and predicting breast cancer risk. Br J Cancer 20:1–9 23. Ihaka R, Gentlemen RR (1997) The R Project for Statistical Computing [Internet]. Statistics Department of the University of Auckland. http://www.r-project.org/ 24. Carney PA, Miglioretti DL, Yankaskas BC, Kerlikowske K, Rosenberg R, Rutter CM et al (2003) Individual and combined effects of age, breast density and hormone replacement therapy use on the accuracy of screening mammography. Ann Intern Med 138:168–175
123
