Dr. John Heine from the Moffitt Cancer Research Center has provided three comprehensive EDRN data collections from the protocol Automated Quantitative Measures of Breast Density for our hackathon. These datasets are publicly accessible and stored in the Cancer Biomarker Data Commons ("LabCAS").

In order to make the best use of the time available for the hackathon, all of this data is available for downloading using IBM Aspera.

About the Collections and Links to LabCAS

The following information describes the three collections and contains links into the EDRN biomarker data commons, LabCAS:

Collection 1: Automated System For Breast Cancer Biomarker Analysis

Instrument: General Electric (GE) Senographe 2000D full field digital mammography (FFDM), 2D units

Data Description: 180 case-control pairs; 2D mammograms in both for presentation and for processing image data representations; demographical-clinical; and pathological data from women, 18 years and older.

Grant Number: R01CA114491 breast density and breast cancer risk study (2006-2011).

Collection 2: Automated Quantitative Measures of Breast Density Data

Instrument: Hologic Selenia FFDM 2D units and Dimensions digital breast tomosynthesis (DBT) units operating in the 2D mode

Data Description: 319 case-control pairs; 2D mammographic images in for presentation and for processing image data representations; demographical-clinical; and pathological data, from women 18 years and older.

Grant Number: R01CA166269 breast density and breast cancer study (2013-2017).

Collection 3: Moffitt's Hologic Dimensions 3D Case-Control Mammography Study

Instrument: Hologic Dimensions digital breast tomosynthesis (DBT) mammography units

Description: 348 case-control pairs; mammographic images: (2D) for presentation and for processing image data representations; DBT volumetric (2D slices) and C-View synthetic 2D mammograms; demographical-clinical; and pathological data from women 18 years and older.

Grant Number: U01CA200464 breast density and breast cancer risk study (2016-2022).

Metadata Explorer

We have setup the following GitHub to explore the metadata for all three datasets. That will provide insights to members wanting to formulate problems to be attacked during and after the hackathon.

Smaller Subset

A smaller subset of the data (in the form of a GZIP-compressed TAR archive) is also available. This subset contains selected datasets for ease of downloading, use, and analysis. This subset, as well as the above three collections, are available to download using Aspera.

Results and Conclusions Summary

Select findings from studies using these collections are briefly described, as the investigations spanned roughly 19 years. Concise findings and detailed descriptions of the algorithms and methods can be found from the respective citations.

From Collection 1, a calibration system was established to adjust for processing mammograms to a common normalized effective x-ray attenuation coefficient scale [1]. This system included a serial quality assurance (QA) monitoring system based on the cumulative sum (CUSUM) technique. This QA approach was used to check the serial accuracy of baseline calibration data so it could be updated if calibration accuracy moved beyond a given tolerance [2]. This technique was able to detect signs of impending x-ray tube failure long before the actual failure. Calibration also required an accurate estimate of the compressed breast thickness during the acquisition. Compression paddle force on the breast caused deviations in the compressed breast tissue thickness spatial distribution; a method was developed to compensate for the compression paddle contortion [3]. Using calibrated data, the mean, variation, BI-RADS ordinal breast composition measures, and a PD-type measure produced significant ORs [4-6]. The variation from both for presentation and for processing mammograms produced significant ORs, and that BI-RADS descriptors could be developed without calibration.

From Collection 2, the same calibration system was extended from the Collection 1 study [7], and the QA monitoring approach was advanced. Here we showed how to monitor a given acquisition’s baseline calibration data with a 50/50 percent adipose/glandular phantom and correct a given curve when it drifted out of a present calibration tolerance with the same phantom 50/50 phantom measurement [8]; this system also required a compressed breast thickness correction technique to compensate for the paddle deflection due to the compression force [9]. The calibrated variation measure produced significant ORs [10]. An automated PD measure developed earlier [11, 12], modified to operate on for presentation and for processing images (i.e., non-calibrated data) also produced significant ORs [13] unrelated to the calibration study.

From collections 1 and 2, Fourier ring measurements (texture analyzed in the Fourier domain) produced significant findings across the power spectrum with calibrated, for processing, and for presentation data representations[14]; illustrations were provided showing how these measures translate to image texture and relevant spatial scales. In particular, a low frequency measurement produced significant ORs that was in agreement with earlier work performed with digitized film mammograms [15] and subsequently validated with two large disparate populations with FFDM data [16]. A summarized local spatial correlation measures produced significant ORs related also related to specific spatial scales [17].

From Collection 3, several PD type measures produced significant ORs [18]. PD as a volumetric quantity, PD averaged over the DBT slices (area quantity) and determined from 2D synthetic images. ORs were about the same across these measures. Theoretical derivations predicted the volumetric and average slice measures are the same agreeing with these findings. PD was modeled as a function compressed breast thickness (CBT). Maximum PD location was approximately 0.41×CBT and similar across case-controls. PD determined from the slice where it was a maximum value produced significant ORs very similar to the other PD methods. Both the average pixel values from the DBT volume and from the 2D synthetic images produced significant ORs. Unlike the analysis of 2D FFDM images, variation measured in multiple ways did not produce significant findings.

Method Details

Approach: all collections were developed to make breast density measurements using a matched (1:1) case-control design. Mammograms in the craniocaudal view were used in all studies. Breast density is a generalized term used in our context to imply an arbitrary measurement from a mammogram. For the most part, citations were reserved for the results section unless prior work was required for replicating the algorithms or to put investigations in context. Image definitions and descriptions are provided in the Data collection protocol Section. Collection 1 used two-dimensional (2D) mammograms acquired with full field digital mammography (FFDM). This collection’s main study purpose was to standardize (for processing) mammograms using a calibration phantom imaging approach to account for acquisition technique influences (i.e., target-filter combination, mAs, and compressed breast thickness). This approach was founded on making effective x-ray coefficient measurements from breast tissue equivalent phantoms [19] and developing a serial quality control method using the cumulative sum approach [20]. Images were mapped to a standardized scale (0 – 100) representing total adipose to total glandular tissue respectively, prior to making image measurements. The mapping required developing baseline calibration curves at time-zero (about 160 phantom acquisitions). Differential evolution optimization [21] was used to develop a four-state ordinal measurement system that paralleled the BI-RADS [22] breast composition descriptors using both calibrated and non-calibrated data. The implementation of this DE approach is also described in detail in this work [23]. For risk modeling in all related studies, conditional logistic regression was used to estimate odds ratios (ORs) for given image measurement in a standard deviation increment with confidence intervals (CIs). ORs are the primary risk metric for the matched design with our intent to isolate a given image measurement. For most modeling outputs, ROC curve areas were provided as well with CIs. Both body mass index and ethnicity were controlled for in the modeling with a given image measurement. In all studies discussed here, breast density measurements were made from the non-affected breast for a case and same-side breast for the matched control. Calibrated measures were compared with breast density measured with the user-assisted Cumulus PD method [24-26] (binary labeling of breast density based on thresholding with the dense area normalized by the total breast area). This PD measure is often used as the standard for comparison. Automated measures were based on eroding the breast area first by 25% unless noted. Here we assumed the breast was an approximate half semi-circle and eroded the breast area inward radially by 25%. Erosion was used to approximate the breast area that was in contact with the compression paddle surface during the acquisition. Collection 2 was designed with the same intent as Collection 1 to evaluate if the same calibration approach was applicable to a different type of mammography technology. In both collections, various measures were made both for processing and for presentation images without calibration within the eroded area. Measures included the mean, standard deviation, Fourier based on dividing the power spectrum into concentric rings and summing the power within a given ring [27] (texture analysis in the Fourier domain), local correlation, and automated PD (without erosion). The same image processing techniques and risk modeling used for Collection 1 were applied to Collection 2. When making Fourier measures, the largest rectangle that would fit within the breast area was used as the measurement-region. This rectangle algorithm is described in detail in this work [14]. Collection 3 was designed to make breast density measurements from DBT data and the related 2D synthetic images. Here, an automated PD type measurement was applied to the volume slices producing (1) a volumetric measure, and (2) an average PD measurement taken over the DBT slice measures. The same PD approach was applied to the synthetic 2D images. The same modeling design was used as with the other collections, where a given breast density measurement was modeled with conditional logistic regression; analyzing DBT data did not require breast area erosion.

Data Collection Protocol

All datasets were developed with the same IRB-approved protocol. Cases (unilateral breast disease) were either: (1) women attending the breast clinics at Moffitt Cancer Center (MCC) diagnosed with breast cancer (type-1) or (2) attendees of surrounding area clinics sent to MCC for breast cancer treatment or diagnostic purposes and found to have breast cancer (type-2). Cases have pathology verified unilateral (first time) breast cancer. Controls were attendees of MCC with no history of breast cancer. Controls were individually matched to cases on age (± 2 years), hormone replacement therapy (HRT) usage and current duration, screening history, and mammography unit. The HRT match was based on status of current users or non-users. Nonusers included women that have not taken HRT for at least two years. If a case was a current HRT user, the control was matched on this duration (± 2 years). Controls were matched by screening history using a three-category classification. Group 1 included women with prior screening history by any means; the duration between the last screening and the study image date must be no more than 30 months. Group 2 included women with a screening history that does not fit within in Group 1 or Group 3. Group 3 included women with no screening history. We used mammograms in craniocaudal (CC) views as study images. The unaffected breast was used as the study image for cases (image acquired before treatment) and the matching lateral breast for controls. Women that had breast implants were excluded from this study. Cases were selected retrospectively (type-1) via electronic medical records search or recruited (type-2). Controls were selected retrospectively via electronic medical records search. Multiple suitable controls were matched to a given case and one control was selected randomly for the study.

The collections represent mammograms from different technology designs and manufactures. For conventional 2D images acquired with full field digital mammography (FFDM), there are two sets of related images available at the acquisition time termed for processing and for presentation; for processing mammograms can be considered as raw data and are not used clinically. Manufacturer specific algorithms are applied to these images to produce enhanced for presentation images that are used clinically. Both types of images can be used for experimental measurement investigations. FFDM images from Collection 1 have 100µm pitch and mammograms (FFDM) from collection 2 have 70µm pitch. Digital breast tomosynthesis data includes volumetric images that are 2D slices of the breast about 1mm thick (or about 10 slices per cm of compressed breast thickness) and 2D synthetic images referred to as C-View images for the technology associated with Collection 3. We will refer to these as 2D synthetic images. Pixel spacing in DBT acquired images is about 100µm but varies across women from roughly 80µm to 110µm in these datasets (but is the same for a given woman’s DBT dataset). DBT units can take both 2D FFDM and DBT acquisitions in tandem without patient repositioning; due the way the images from Collection 2 and 3 were acquired, a larger DBT dataset can be constructed by combining elements from both collections (see citation in Collection 3 results). There is an accompanying data dictionary with these collections. Of importance for automated processing, an intricate image file naming convention was developed to tell the user many study and image characteristics such as study number, case-control status, mammographic view, eligible study image, and image type. The file name string convention can be searched automatically to find a given image type relatively easily and to ensure cases and matching controls can be assembled.

Limitations: sampling of cases and controls was not population-based, but rather a mixture of cases ascertained at an NCI-designated comprehensive cancer center inclusive of referrals from the community. There is no evidence from our studies that the cases are not representative, but findings should be replicated in a population-based study. The data fields allow for selection of the population-based cases (discounting the referrals) but will reduce the case-control numbers. Image data from the General Electric Senographe 2000D full field digital mammography units do not include women with large breasts due to the x-ray detector design limitations. Images may contain artifacts such as nipple markers, mole markers, biopsy clips, and scar markers. These artifacts are documented in the data fields. All images were visually inspected and approved for automated processing. Here, a judgement was made to exclude a sample with too many markers or to include the sample because artifacts were deemed negligent.

References

[1] J. J. Heine and J. A. Thomas, "Effective x-ray attenuation coefficient measurements from two full field digital mammography systems for data calibration applications," (in Eng), Biomedical engineering online, vol. 7, no. 1, p. 13, Mar 28 2008. [Online]. Available: http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Citation&list_uids=18373863

[2] J. J. Heine, K. Cao, and C. Beam, "Cumulative sum quality control for calibrated breast density measurements," Med Phys, vol. 36, no. 12, pp. 5380-90, Dec 2009, doi: 10.1118/1.3250842.

[3] J. J. Heine, K. Cao, and J. A. Thomas, "Effective radiation attenuation calibration for breast density: compression thickness influences and correction " BioMedical Engineering OnLine vol. 9, p. 73, 2010.

[4] E. E. Fowler, T. A. Sellers, B. Lu, and J. J. Heine, "Breast Imaging Reporting and Data System (BI-RADS) breast composition descriptors: automated measurement development for full field digital mammography," Med Phys, vol. 40, no. 11, p. 113502, Nov 2013, doi: 10.1118/1.4824319.

[5] J. J. Heine, K. Cao, D. E. Rollison, G. Tiffenberg, and J. A. Thomas, "A Quantitative Description of the Percentage of Breast Density Measurement Using Full-field Digital Mammography," (in eng), Acad Radiol, vol. 18, no. 5, pp. 556-64, May 2011, doi: 10.1016/j.acra.2010.12.015.

[6] J. J. Heine, E. E. E. Fowler, and C. I. Flowers, "Full field digital mammography and breast density: comparison of calibrated and noncalibrated measurements," (in English), Acad Radiol, vol. 18, no. 11, pp. 1430-6, 2011, doi: 10.1186/1475-925X-12-114.

[7] E. E. Fowler, B. Lu, and J. J. Heine, "A comparison of calibration data from full field digital mammography units for breast density measurements," Biomedical engineering online, vol. 12, p. 114, 2013, doi: 10.1186/1475-925X-12-114.

[8] B. Lu, A. M. Smallwood, T. A. Sellers, J. S. Drukteinis, J. J. Heine, and E. E. Fowler, "Calibrated breast density methods for full field digital mammography: a system for serial quality control and inter-system generalization," Med Phys, vol. 42, no. 2, pp. 623-36, Feb 2015, doi: 10.1118/1.4903299.

[9] E. E. E. Fowler, A. M. Smallwood, N. Z. Khan, K. Kilpatrick, T. A. Sellers, and J. Heine, "Technical challenges in generalizing calibration techniques for breast density measurements," Med Phys, vol. 46, no. 2, pp. 679-688, Feb 2019, doi: 10.1002/mp.13325.

[10] E. E. Fowler et al., "Calibrated Breast Density Measurements," Acad Radiol, Dec 10 2018, doi: 10.1016/j.acra.2018.10.009.

[11] J. J. Heine et al., "An automated approach for estimation of breast density," (in eng), Cancer Epidemiol Biomarkers Prev, vol. 17, no. 11, pp. 3090-7, Nov 2008. [Online]. Available: http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Citation&list_uids=18990749

[12] J. J. Heine and R. P. Velthuizen, "A statistical methodology for mammographic density detection," Med Phys, vol. 27, no. 12, pp. 2644-51, Dec 2000. [Online]. Available: http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Citation&list_uids=11190946.

[13] E. E. Fowler, C. M. Vachon, C. G. Scott, T. A. Sellers, and J. J. Heine, "Automated Percentage of Breast Density Measurements for Full-field Digital Mammography Applications," Acad Radiol, vol. 21, no. 8, pp. 958-70, Aug 2014, doi: 10.1016/j.acra.2014.04.006.

[14] E. E. E. Fowler, A. Smallwood, C. Miltich, J. Drukteinis, T. A. Sellers, and J. Heine, "Generalized breast density metrics," (in English), Physics in Medicine and Biology, vol. 64, no. 1, Jan 2019, doi: 10.1088/1361-6560/aaf307.

[15] A. Manduca et al., "Texture features from mammographic images and risk of breast cancer," Cancer Epidemiol Biomarkers Prev, vol. 18, no. 3, pp. 837-45, Mar 2009, doi: 10.1158/1055-9965.EPI-08-0631.

[16] J. Heine et al., "Mammographic Variation Measures, Breast Density, and Breast Cancer Risk," AJR. American journal of roentgenology, vol. 217, no. 2, pp. 326-335, Aug 2021, doi: 10.2214/AJR.20.22794.

[17] E. E. E. Fowler, C. Hathaway, F. Tillman, R. Weinfurtner, T. A. Sellers, and J. Heine, "Spatial correlation and breast cancer risk," Biomedical Physics & Engineering Express, vol. 5, no. 4, p. 045007, 2019/05/22 2019, doi: 10.1088/2057-1976/ab1dad.

[18] J. Heine, E. E. E. Fowler, R. J. Weinfurtner, E. Hume, and S. S. Tworoger, "Breast density analysis of digital breast tomosynthesis," Sci Rep, vol. 13, no. 1, p. 18760, Oct 31 2023, doi: 10.1038/s41598-023-45402-x.

[19] J. J. Heine and M. Behera, "Effective x-ray attenuation measurements with full field digital mammography," Med Phys, vol. 33, no. 11, pp. 4350-66, Nov 2006. [Online]. Available: http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Citation&list_uids=17153414.

[20] D. M. Hawkins and D. H. Olwell, Cumulative Sum Charts and Charting for Quality Improvement (Statistics for Engineering and Physical Science ). New York: Springer-verlag 1997.

[21] K. V. Price, R. M. Storn, and J. A. Lampinen, Differential evolution : a practical approach to global optimization (Natural computing series). Berlin ; New York: Springer, 2005, pp. xix, 538 p.

[22] C. D'Orsi, C. J, and e. al., ACR BI-RADS Atlas: Breast Imaging Reporting and Data System; Mammography, Ultrasound, Magnetic Resonance Imaging, Follow-up and Outcome Monitoring, Data Dictionary. ACR, American College of Radiology, 2013.

[23] E. E. Fowler, A. Berglund, T. A. Sellers, S. Eschrich, and J. Heine, "Empirically-Derived Synthetic Populations to mitigate Small Sample Sizes " Journal of Biomedical Informatics, vol. 105, 2020. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S1532046420300368.

[24] J. W. Byng, N. F. Boyd, E. Fishell, R. A. Jong, and M. J. Yaffe, "Automated analysis of mammographic densities," Phys Med Biol, vol. 41, no. 5, pp. 909-23, May 1996, doi: 10.1088/0031-9155/41/5/007.

[25] J. W. Byng, N. F. Boyd, E. Fishell, R. A. Jong, and M. J. Yaffe, "The quantitative analysis of mammographic densities," Phys Med Biol, vol. 39, no. 10, pp. 1629-38, Oct 1994, doi: 10.1088/0031-9155/39/10/008.

[26] N. F. Boyd et al., "Quantitative classification of mammographic densities and breast cancer risk: results from the Canadian National Breast Screening Study," J Natl Cancer Inst, vol. 87, no. 9, pp. 670-5, May 3 1995, doi: 10.1093/jnci/87.9.670.

[27] J. J. Heine and R. P. Velthuizen, "Spectral analysis of full field digital mammography data," Med Phys, vol. 29, no. 5, pp. 647-61, May 2002. [Online]. Available: http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Citation&list_uids=12033559.