Characterization of the imaging settings in screening mammography using a tracking and reporting system: A multi-center and multi-vendor analysis

Abstract

A tracking and reporting system was developed to monitor radiation dose in X-ray breast imaging. We used our tracking system to characterize and compare the mammographic practices of five breast imaging centers located in the United States and Brazil. Clinical data were acquired using eight mammography systems comprising three modalities: computed radiography (CR), full-field digital mammography (FFDM), and digital breast tomosynthesis (DBT). Our database consists of metadata extracted from 334,234 images. We analyzed distributions and correlations of compressed breast thickness (CBT), compression force, target-filter combinations, X-ray tube voltage, and average glandular dose (AGD). AGD reference curves were calculated based on AGD distributions as a function of CBT. These curves represent an AGD reference for a particular population and system. Differences in AGD and imaging settings were attributed to a combination of factors, such as improvements in technology, imaging protocol, and patient demographics. The tracking system allows the comparison of various imaging settings used in screening mammography, as well as the tracking of patient-and population-specific breast data collected from different populations.

1. Introduction

Every medical image contains a plethora of useful information for conducting quality assurance, which encoded in the image data and reported in the digital information and communications in medicine (DICOM) header [1]. The DICOM header has been used in many applications to monitor the patient radiation dose in order to better manage radiation risks and to improve patient care [2–7]. In a previous study, we presented a tracking system [7] that can be used for radiation dose monitoring in X-ray breast imaging. Our tracking system allows us to characterize acquisitions in real-time [8–11,7].

The current study presents the implementation of our tracking and reporting system used to compare and contrast the breast imaging practices in two countries, the United States and Brazil. Our tracking system was used in five clinical institutions for screening and diagnostic mammography. At the time of this report, the database contained the metadata of 334,234 clinical images acquired using three modalities: computed radiography (CR), full-field digital mammography (FFDM), and digital breast tomosynthesis (DBT). This study demonstrates the use of the tracking system to monitor and to report the clinical imaging parameters used in screening mammography, such as breast compression, radiation dose, and X-ray technique.

2. Materials and methods

2.1. Data collection

The clinical studies were provided by five breast imaging centers located at the University of Pennsylvania Perelman Center for Advanced Medicine (PCAM), Penn Presbyterian Medical Center (Presby), Pennsylvania Hospital (PAH), Hospital of Sao Paulo (HSP), and Sao Paulo Institute of Radiology (INRad). All imaging centers offer screening and diagnostic mammography, together with other breast imaging services. The institutional review boards waived the require-ment to obtain written consent for this retrospective Health Insurance Portability and Accountability Act (HIPAA)-compliant study.

In this study, we collected clinical images from the screening populations at the breast imaging centers named above (Table 1). For all breast imaging centers, the screening mammography protocol requires two images per breast: cranio-caudal (CC) and medio-lateral oblique (MLO). Additional CC and MLO exposures may be needed to correct acquisition problems such as poor positioning and image artifacts. In addition, multiple exposures may be required for large breasts that do not fit on the detector and for women with breast implants. Additional breast views (i.e., not CC or MLO) were excluded from the data analyses. Automatic exposure control (AEC) was the acquisition method used for the majority of the images. Manual mode was used for particular cases, such as dense breasts and breasts with implants. In this work, clinical images acquired using manual mode were excluded from the data analyses.

Table 1.

Summary of mammography acquisition systems and data collected in this study.

X-ray Unit (ID)	Vendor -	Site -	Modality -	Units n	Patients n	Patient Age avg (SD)	Images n
(A) Senographe Essential	GE Healthcare	PCAM	FFDM	3	3,644	58 (11)	18,452
(B) Senographe DS	GE Healthcare	PCAM	FFDM	1	271	58 (11)	1,321
(C) LORAD Selenia	Hologic Inc.	Presby	FFDM	2	3,214	57 (11)	14,052
(D) Selenia Dimensions	Hologic Inc.	PCAM	FFDM/DBT	4	54,244	56 (11)	235,225
(E) MAMMOMAT Novation	Siemens	PAH	FFDM	3	13,028	58 (12)	54,733
(F) MAMMOMAT Inspiration	Siemens	PAH	FFDM/DBT	2	1,258	56 (11)	5,006
(G) Senographe DS	GE Healthcare	INRad	FFDM	2	598	59 (12)	2,245
(H) Performa1	GE Healthcare	HSP	CR	1	249	57 (11)	988
(I) LORAD MIV	Hologic Inc.	HSP	CR	1	554	56 (10)	2,212
Total				19	77,060		334,234

¹GE Instrumentarium Performa Mammography.

2.2. Database creation

We used a tracking and reporting system to store the DICOM attributes (i.e., imaging settings, patient information, operator, etc.) in a structured query language (SQL) database [7]. This tracking and reporting system consists of two major components: a custom DICOM Service Class Provider (SCP) and a client-side application that is used in clinical practice to generate physics reports for radiation monitoring in breast imaging.

The custom DICOM SCP identifies DICOM message service elements (e.g., C-FIND, C-MOVE, and C-STORE) from the picture archiving and communication system (PACS) [1], using an established association via transmission control protocol/internet protocol (TCP/IP) [7]. To reduce disk space usage, the DICOM SCP discards the image content from the DICOM files and stores only the DICOM header information (i.e., metadata) into the SQL database. It is important to mention that the SQL database resides at the source institution and the metadata are stored in compliance with local regulations.

All of the metadata were stored and unified into a single SQL database for data analysis. As we described in the previous study [7], we used the DICOM hierarchical model [12] and DICOM UID as primary keys to avoid duplicate entries in the database.

For images acquired using CR and FFDM modalities, we only stored the metadata of the processed mammograms (“For Presentation”). However, the CR vendors do not provide the DICOM header in its entirety. For example, the acquisition information of compression force, half value layer, and average glandular dose need to be calculated or registered manually into the DICOM header.

For images acquired using the DBT modality, only the metadata of reconstructed images (“For Presentation”) were stored into our SQL database. The patient data from the projections were not saved in the database to avoid duplication.

2.3. Data analysis

The access to the database is independent of the PACS associations; database queries can be performed in real-time. In this study, we retrieved the metadata information to characterize the mammography imaging settings such as compression force, X-ray spectrum (target-filter and tube voltage selections), and average glandular dose. We also identified clinical images that are under and over the expected population dose levels using polynomial curves.

2.3.1. Compression force

The breast compression should be both appropriate and tolerable, so that the system selects the correct acquisition parameters [13–15]. Optimized breast compression reduces radiation burden to the patients, minimizes superimposition of breast tissues, and limits patient motion [13,14]. Depending upon the mammography system, the compression force information can either be automatically recorded in the DICOM header after the image acquisition or entered manually by the technician.

In this study, systems A-G provided the compression force information in the DICOM header. We retrieved this information from the SQL database to analyze the force applied to the patients’ breasts. The compression force was analyzed as a function of compressed breast thickness (CBT). The technicians did not enter the compression force information for CR images acquired using systems H and I.

2.3.2. X-ray Spectrum

The target-filter and tube voltage selection influences dose and contrast in mammography [16,17]. Due to the diversity of target-filter combinations, we only compared target-filter distributions of images acquired using systems manufactured by the same vendor. We compared the following three pairs of vendors: A and H (GE Healthcare), C and I (Hologic Inc.), and E and F (Siemens). We investigated the target-filter distributions as a function of CBT. We also investigated the AEC selection of target-filter and tube voltage as a function of CBT.

2.3.3. Average glandular dose

It is important to achieve optimal radiation dose levels during mammography screening while achieving the best clinical outcomes. Average glandular dose (AGD) is one of the measures used to evaluate patient dose and radiation risk of breast imaging [18], including the theoretical risk of iatrogenic breast cancer [19,20]. In this study, we monitored and estimated AGD on both an individual and a population basis.

Due to the diverse patient populations and use of multiple mammography systems in this study, we only compared AGD distributions of images acquired using systems manufactured by the same vendor. Similarly, we characterized the AGD as a function of CBT and target-filter combination using the pairs of systems: A and H, C and I, and E and F.

We used second-degree polynomial regressions to develop fitting curves of AGD as a function of CBT. These fitting curves represent the estimated AGD on a population basis. Our tracking system was used to select clinical mammograms that are below and above the AGD fitting curves (AGD reference curves).

2.3.4. Image quality

The AEC should determine not only the optimal imaging settings, but also maintain appropriate image quality for screening mammography. Usually, the quality assessment of images acquired in clinical practice requires the inspection of experienced readers [21]. In our study, an experienced radiologist evaluated the quality of selected clinical images.

The radiologist compared the quality of images acquired using systems E and F. We selected six studies above, six studies below, and three studies at the AGD reference curves of system E. For images acquired above and below the AGD reference curve, the relative difference between AGD reference and the AGD reported in the DICOM header was greater than 25%. For images acquired at the AGD reference curve, the relative difference between AGD reference and the AGD reported in the DICOM header was less than 10%. For these patients, we also selected the corresponding images that were acquired with system F. The radiologist evaluated the quality of these 15 pairs of images side-by-side and reported the quality in terms of noise, contrast, image processing, and positioning.

Note that systems E and F were installed in the same breast imaging center (PAH), but at different time-periods. It should also be noted that these systems are equipped with different detectors; the pixel element sizes are 0.07 mm and 0.085 mm for systems E and F, respectively.

2.4. Assessment of AGD

2.4.1. AGD in CR systems

The CR systems used in this work are Kodak DirectView Classic CR (Mod. 975, Eastman Kodak Company, Rochester, NY) used together with KODAK CR cassettes (Mod. EHR-M2, Eastman Kodak Company, Rochester, NY). We used a Radcal ionization chamber (mod. 9010, Radical Corporation, Monrovia, CA) and aluminum foil of size 100 × 100 × 0.1 mm with certified purity greater than 99.999% (Nuclear Associates, Cleveland, OH) to measure the half-value layer (HVL) and the entrance surface air kerma (ESAK) for entry into the database. The ESAK measurements were calculated using up to seven polymethyl methacrylate (PMMA) slabs with a size of 200 × 250 × 10 mm and one PMMA slab with a size of 200 × 250 × 5 mm.

The method used for calculating the HVL and ESAK was described by Alves [22]. This work is based on guidelines for quality assurance in digital mammography [23,24]. The HVL and ESAK measurements were used to calculate the AGD in CR, following Eq. 1 described by Dance et al. [25–27]:

$$D_g\text{=}Kgcs$$ (1)

where D_g and K represent AGD and ESAK, respectively; g converts K to AGD for a breast having a 50% fibroglandular and 50% fat composition with a particular thickness, using an appropriate value of the HVL; c corrects D_g for any difference in breast composition; and s corrects for differences the X-ray spectrum selection. In this work, c was interpolated as a function of HVL, CBT, and patient age for all the images (Tables 7 and 8 in [26]). We used s = 1 for all images, because the X-ray spectrum used to acquire the CR images was Mo/Mo.

2.4.2. AGD in FFDM and DBT systems

Vendors use different methods to assess and estimate AGD, which is displayed as “organ dose” in the DICOM header and measured in dGy. For instance, Hologic Inc. uses the method of Boone [28–30], GE Healthcare uses the method of Wu [31], and Siemens uses the method of Dance [25–27] to estimate organ dose (NB: the dose method also varies by region, thus the reader may find that the dosimetry method used at their site may differ from that stated here). To avoid any disparity in dose calculations across vendors, we only compared the AGD calculated in CR with the AGD reported in the DICOM header of FFDM systems produced and sold by the same vendor. The accuracy of AGD reported in the DICOM header was evaluated in a previous study [8].

We also compared the AGD in FFDM with the AGD reported in the DICOM header of DBT systems. We used the AGD of the Tomo Combo mode, which represents FFDM and DBT acquisitions below the same breast compression. We characterized the AGD as function of CBT for images acquired using Tomo Combo mode of systems D and F. We also calculated separately the summary AGD of 2D and 3D components of the Tomo Combo. The AGD analysis of components 2D and 3D was also segregated by breast view (CC and MLO).

2.5. Statistical analysis

The radiation dose-specific information estimated by the vendor and recorded in the DICOM header varies by vendor, system model, and imaging modality. We used our tracking and reporting system to select, filter, aggregate, and correlate the DICOM attributes using the metadata stored in the SQL database. We analyzed various imaging and patient parameters (e.g., compression force, breast thickness, patient age, target-filter combination, kV, etc.) that affect the radiation dose provided by the vendor.

In this study, we use tables, histograms, boxplots, and scatter plots to present statistical analyses of the clinical data collected. In the boxplots and tables, we report the first (Q1) and the third (Q3) quartiles to present the range of the data distribution. The quartiles are, respectively, the 25 percent and 75 percent quantiles.

We verified correlations between the categorical variables using Spearman rank correlation coefficient (ρ). We also used regression models to estimate a linear relationship between predictor-response variables. The Kolmogorov-Smirnov (K-S) test was used to verify the normality of distributions using continuous data. Non-parametric Kruskal-Wallis (K-W) test by rank and Mann-Whitney-Wilcoxon (MWW) test were used to estimate differences in means and to analyze significant differences in the pairwise comparison of continuous variables.

3. Results

3.1. Compressed breast thickness

The amount of breast compression was analyzed based on imaging systems (Fig. 1a) and segregated by categorical CBT (Fig. 1b). These results can be used to support quality control (QC) tests in DM [23,24]: The Mammography Quality Standards Act (MQSA, United States) and the National Program for Quality in Mammography (PNQM, Brazil) regulate the requirements for the maximum compression force [32]. All of the systems (A through I) had passed the compression force QC test. It should be noted that the compression force information was not provided for the clinical images in systems H and I [22].

Fig. 1.

(a) Boxplot of compression force for images acquired using the systems A through G. (b) Boxplot of compression force based on categorical CBT for images acquired using the systems A through G. Note that there are y-scale differences and outliers are not presented in (b).

The MQSA and PNQM state that the maximum compression force for the initial power drive shall be between 111 N (25 lb) and 200 N (45 lb) [32]. Subsequent manual compression can be applied in excess of 200 N. Images acquired using compression force higher than 200 N were considered outliers in this study (Fig. 1a). Only 253 (0.80%), 126 (0.78%), 2,753 (0.08%), and 34 (≈0%) images exceeded 200 N for systems A, C, D and E, respectively. Systems B, F, and G did not acquire images using compression force greater than 200 N. The majority of these images (77%) were acquired on thick breasts (CBT > 65 mm), which usually require more compression for optimal radiation exposures in the mammography exam.

The compression force values did not follow normal distributions (D > 0.99 and p < 0.001); thus, we compared the differences in means of the compression force using the K-W test by rank, and the MWW test was used to compare statistical differences in compression force as a function of systems. There were significant differences in the pairwise comparison between the compression force distributions (p < 0.001), except between systems A and B (p = 0.21), as shown in Fig. 1a. However, we observed significant differences in the pairwise comparison between systems A and B using categorical CBT (Fig. 1b). For the thinnest and thickest breasts (i.e., CBT of sizes between 21–30 and 81–90 mm), the differences in compression force were significant in the pairwise comparison (p < 0.001).

Compression force is correlated to CBT in all cases, except for system G (p = 0.83 and ρ=0.00). The ρ values obtained were 0.26, 0.09, 0.16, 0.29, 0.38, and 0.27 for systems A through F (p < 0.05), respectively. Differences in compression force are due to differences in the imaging protocol that is implemented at each breast imaging center, e.g., use of different compression paddles, technologists’ education, training, and experience, as well as the extent of physician oversight and physicians’ preferences.

3.2. Target-filter

We analyzed the distribution of target-filter combinations as a function of CBT (Fig. 2).The CBT distributions categorized by target-filter combinations did not follow normal distributions (D = 1 and p < 0.001). The CR (Fig. 2a and Fig. 2c) images were acquired using Mo/Mo, since this is the only target-filter combination available. The second generation of FFDM systems manufactured by GE Healthcare has Rh/Rh and Mo/Rh as options for target-filter combinations (Fig. 2b). In total, 85.4%, 13.3%, and 1.3% of the images were acquired using the target-filter combinations Rh/Rh, Mo/Rh, and Mo/Mo on system A, respectively. The systems manufactured by Hologic replaced Mo/Mo with W/Rh and W/Ag as the target-filters combinations (Fig. 2d). In total, 80.4% and 19.6% of the images were acquired using the target-filter combinations W/Rh and W/Ag on system C, respectively. Note that the AEC is typically set to switch the target-filter combination from W/Rh to W/Ag at a CBT 70 mm, while system A uses a different combination of imaging parameters to switch the target-filter combination (Fig. 2b). Systems E (Fig. 2e) and F (Fig. 2f) acquire the majority of images using W/Rh as target-filter combination. Only 0.5% and 0.2% of the images were acquired using the target-filter combinations Mo/Mo and Mo/Rh.

Fig. 2.

CBT distribution as a function of target-filter combinations for the pairs of systems: (a) H and (b) A (GE Healthcare), (c) I and (d) C (Hologic Inc.), and (e) E and (f) F (Siemens). Note that there are y-scale differences.

3.3. Tube voltage

We present the kV and target-filter usage using the same pairs of systems: A and H, C and I, and E and F (Fig. 3). The differences in tube voltage between systems are due to differences in detector response and efficiency, and changes to the optimization strategies, resulting in different AEC programming. A red asterisk was used to identify data clusters that comprise at least 10% of the total number of images acquired for a particular target-filter combination.

Fig. 3.

CBT boxplot as a function of tube voltage and target-filter combinations for the pairs of systems: (a) H and (b) A (GE Healthcare), (c) I and (d) C (Hologic Inc.), and (e) E and (f) F (Siemens).

We expect that thicker breasts are imaged using higher tube voltage. The systems used as exemplars have a strong correlation between target-filter, kV and CBT (ρ > 0.69). kV < 26, kV = 33 or kV = 35 were used to acquire only 10% of the images on system H (Fig. 3a). kV < 26 or kV > 31 were used to acquire only 0.04% of the images on system A (Fig. 3b). Note that thicker breasts tend to be imaged using Rh/Rh and higher kV on this system. kV < 26 or kV > 32 were used to acquire only 5.5% of the images on system I (Fig. 3c). As shown in Fig. 2d, the system I switched the target-filter combinations from Mo/Mo to W/Rh and W/Ag in system C. Fig. 3d shows that no images were acquired using 33 kV. We could not predict a clear relationship between CBT, kV, and target-filter combinations on system E (Fig. 3e). kV < 27 or kV > 32 were used to acquire only 0.03% of the images on system F (Fig. 3f). System F also acquired only 10 images with W/Rh and kV > 32 combinations.

3.4. Average glandular dose

The summary of the AGD and CBT statistics is shown in Table 2. We compared the differences in mean CBT and mean AGD using images acquired in two distinct populations, Brazil and US. The mean CBT and mean AGD of images acquired using system B is 25.4% lower and 46.1% lower than the mean CBT and mean AGD of images acquired using system G (p < 0.001), respectively. The difference in the mean AGD is reduced to 41.5% when we consider only the AGD at 50 mm CBT (p < 0.001). Differences in demographics, equipment calibration, and imaging techniques can affect substantially the AGD estimate.

Table 2.

Summary of mean CBT, mean AGD, and mean AGD at 50 mm CBT, for all systems used in this study (A through I).

X-ray Unit (ID)	Vendor	Modality	CBT, mm (avg ± sd)	AGD, mGy (avg [Q1, Q3])	AGD at 50 mm, mGy (avg ± sd)
(A) Senographe Essential	GE Healthcare	FFDM	53.90 ± 14.57	1.44 [1.20, 1.59]	1.38 ± 0.29
(B) Senographe DS	GE Healthcare	FFDM	42.85 ± 12.77	1.32 [1.07, 1.42]	1.31 ± 0.24
(C) LORAD Selenia	Hologic Inc.	FFDM	57.83 ± 13.56	1.46 [1.16, 1.68]	1.29 ± 0.34
(D) Selenia Dimensions	Hologic Inc.	FFDM	59.65 ± 15.26	1.73 [1.20, 2.12]	1.36 ± 0.40
(E) MAMMOMAT Novation	Siemens	FFDM	43.67 ± 14.13	1.90 [1.50, 2.20]	2.01 ± 0.54
(F) MAMMOMAT Inspiration	Siemens	FFDM	54.50 ± 14.54	0.95 [0.69, 1.20]	1.00 ± 0.30
(G) Senographe DS	GE Healthcare	FFDM	57.41 ± 13.92	2.45 [2.21, 2.74]	2.24 ± 0.28
(H) GE Performa	GE Healthcare	CR	51.94 ± 12.03	2.72 [1.80, 3.24]	2.64 ± 0.54
(I) LORAD MIV	Hologic Inc.	CR	58.46 ± 11.31	3.08 [2.35, 3.55]	2.81 ± 0.64

Since there were significant differences in CBT for the analysed population, we also compared the mean AGD at 50 mm CBT. The differences in mean AGD of systems A and B are not statistically significant when the patients are matched at 50 mm CBT (p = 0.12). The mean AGD of system C is 5.2% lower than the mean AGD of system D (p < 0.001). The mean AGD of images acquired using the FFDM system E is 49% higher than the mean AGD of system F (p < 0.001). Differences in equipment calibration, imaging techniques, and system technology can substantially affect the AGD estimate.

The AGD distribution as a function of CBT is shown in Fig. 4. It should be emphasized that CR systems acquire images using higher radiation dose levels to achieve appropriate image quality for the mammography exam (Figs. 4a and 4c). By comparison, FFDM systems usually achieve appropriate image quality using lower levels of radiation dose (Fig. 4b and 4d).

Fig. 4.

Scatter plot of AGD as a function of CBT and target-filter combinations for the pairs of systems: (a) H and (b) A (GE Healthcare), (c) I and (d) C (Hologic Inc.), and (e) E and (f) F (Siemens). Note that there are y-scale differences.

We expect a positive correlation between AGD and CBT. The systems used as exemplars have a positive correlation between AGD and CBT (ρ > 0.50 and p < 0.001). The regression curves represent the AGD reference for a particular population and system. The AGD reference was calculated as a function of CBT. These curves were used as references to select images that are below, above, and on the curves.

3.5. Image quality

Ultimately, AGD alone does not determine which system has better (or worse) image quality. We used the AGD distributions and reference curves for systems E and F (Figs. 4e and 4f) to select and analyze the image quality of 15 study pairs, in which the same patients were imaged on both systems at different times. An experienced radiologist compared the quality of images acquired on the two systems. Overall, the radiologist reported a personal preference for images acquired using system F, even though system F utilizes much lower dose than system E. Largely, this can be attributed to differences in X-ray spectra, lower noise detector electronics, and image processing advances. Yet, as we illustrate below, neither image quality nor dose alone tells the whole story.

Fig. 5 illustrates images of two patients imaged on systems E and F. In Figs. 5a and 5b, a smaller (30–36 mm thick) breast with an admixture of adipose and glandular tissue is shown, while in Figs. 5c and 5d a thicker (51–60 mm thick) denser breast is shown. Figs. 5a and 5b were chosen because the AGD and thickness reported in the DICOM header of each image was at or near the AGD reference curve for that device, while Figs. 5c and 5d were chosen because the AGD reported was near the upper limit of the AGD distribution for all patients with that breast thickness. In both of these examples, we conclude that the dose to the patient is appropriate to the device in question, even though the dose in the two cases is reduced 30% and 48% respectively when using system F. We draw this conclusion based upon the fact that both devices independently selected similar doses for those patients, when expressed in reference to the doses for that population of women. This conclusion could not be made without dose monitoring, as the values of the doses differ substantially between systems; similarly, doses also vary between institutions and patient populations, as we have shown above.

Fig. 5.

Mammograms acquired using systems (a, c) E and (b, d) F. (a) CBT = 36 mm and AGD = 1.90 mGy, (b) CBT = 30 mm and AGD = 1.33 mGy, (c) CBT = 51 mm and AGD = 4.50 mGy, and (d) CBT = 60 mm and AGD = 2.45 mGy.

Image-wise, the radiologist preferred the images acquired using system F (Figs. 5b and 5d). That said, the radiologist reported that the noise and contrast of the images acquired using system E are also clinically acceptable (Figs. 5a and 5c). The selected and highlighted regions illustrate the differences in contrast and noise seen (Figs. 5a and 5b). The radiologist also reported that the tissue differentiation between fat and glandular was improved with system F.

By comparison, Fig. 6 shows an example in which the dose for system F is again less than system E (18% reduction), yet relative to the AGD reference curve, system F uses substantially more radiation than system E. The image quality in system F is again superior to E, but the image from system F is over-exposed, and thus it is more difficult to determine what benefit was obtained from the use of different imaging technologies and what benefit was obtained from the radiation dose differences. Note that while the region highlighted in red and the calcifications indicated with yellow arrows are more poorly depicted using system E (Fig. 6a), the conspicuity of the calcifications highlighted in blue is superior with system E. It is also important to note that the radiologist reported that the breast compression was better in the images from system F, which further confounds analysis.

Fig. 6.

Mammograms acquired using systems (a) E and (b) F. (a) CBT = 50 mm and AGD = 3.30 mGy, and (b) CBT = 50 mm and AGD = 2.72 mGy.

Finally, in Fig. 7 we see two images of the same patient, both acquired on system E. In this instance, poor breast positioning on the breast resulted in Fig. 7a being acquired with an artificially low dose (28% below the reference curve), while Fig. 7b was correctly positioned and the breast correctly received a dose 6% above the reference curve. Thus even within a single system, there are many factors which determine the dose a women receives-all the more reason to routinely monitor radiation dose and image quality.

Fig. 7.

Mammograms acquired using system E. Images reported with (a) poor and (b) proper breast positioning. (a) CBT = 65 mm and AGD = 2.00 mGy, and (b) CBT = 67 mm and AGD = 2.70 mGy.

3.6. Tomo Combo in DBT

In 2010–11, the U.S. Food and Administration (FDA) approved the use of DBT as an adjunct to FFDM (i.e., Tomo Combo exams). In this study, we analyzed the AGD distributions for images acquired using Tomo Combo with Selenia Dimensions (D) and MAMMOMAT Inspiration (F) systems. In total, 134,779 and 11,332 images were analyzed for systems D and F, respectively.

The AGD of the 2D and 3D components (termed AGD 2D and AGD 3D) were analyzed separately as a function of CBT (Fig. 8a). Fig. 8a shows a strong positive correlation between AGD 2D and CBT (ρ=0.75 and r²=0.48), as well as AGD 3D and CBT (ρ=0.91 and r²=0.85). Similarly, Fig. 8b shows a positive correlation between AGD 2D and CBT (ρ=0.54 and r²=0.27), as well as AGD 3D and CBT (ρ=0.55 and r²=0.64). The Spearman’s rank coefficients indicate that system F adjusts the imaging settings based on additional imaging parameters besides CBT, such as breast density.

Fig. 8.

Polynomial fitting curves of AGD 2D and AGD 3D as a function of CBT using the pairs of (a) systems D (Hologic Inc.) and (b) F (Siemens).

The proportional coefficients of the regression equation shown in Fig. 8b indicate a potentially strong linear correlation between AGD 2D and AGD 3D. We calculated the linear correlation between AGD 2D and AGD 3D to have a y-intercept of 0.00 and slope of 2.0 (r² = 1.00). By comparison, system D did not show a linear correlation between AGD 2D and AGD 3D (r² = 0.56).

We also analyzed differences in mean AGD 2D and AGD 3D using the Tomo Combo datasets. Table 3 shows that system D acquires CC images with mean AGD 3D 11% higher than the AGD 2D, and MLO images with mean AGD 3D 6.8% higher than the AGD 2D. System F acquires CC and MLO images with mean AGD 3D 50% higher than the mean AGD 2D (i.e., mean AGD 3D is twice the mean AGD 2D).

Table 3.

Summary of mean CBT, mean AGD 2D, and mean AGD 3D using the Tomo Combo mode.

DBT System (ID)	No. Images	Mode	Breast View	CBT, mm (avg ± sd)	AGD, mGy (avg ± sd)
(D) Selenia Dimensions	134,779	2D	CC MLO	56.4 ± 12.9 62.2 ± 15.9	1.71 ± 0.67 2.04 ± 0.79
		3D	CC	56.4 ± 12.9	1.93 ± 0.56
			MLO	62.2 ± 15.9	2.19 ± 0.71
		2D + 3D	CC	56.4 ± 12.9	3.75 ± 0.87
			MLO	62.2 ± 15.9	4.11 ± 1.06
(F) MAMMOMAT Inspiration	11,332	2D	CC	53.4 ± 12.9	1.04 ± 0.36
			MLO	55.5 ± 15.1	1.12 ± 0.40
		3D	CC	53.4 ± 12.9	2.08 ± 0.72
			MLO	55.5 ± 15.1	2.24 ± 0.79
		2D + 3D	CC	53.4 ± 12.9	2.16 ± 0.80
			MLO	55.5 ± 15.1	4.32 ± 0.89

4. Discussion

The characterization of the imaging settings used in clinical practice is challenging because it depends on the efficient and accurate performance of the AEC, and the exact formulation of the AEC programming. The selection of optimal imaging settings necessitates further scrutiny by medical physicists [33–35]. Most of the AEC methods use a scout image of the breast at a very low dose, which is used to determine the settings for the clinical image acquisition [36]. We have been using our tracking system at our breast imaging centers to characterize the imaging settings of the AEC modes and to compare the imaging settings of various systems [8–11,7].

The MSQA and PNQM do not state exact guidelines for breast compression. However, the MSQA and PNQM suggest that compression should be applied gently to ensure that the breast is positioned correctly and that the force should not exceed 200 N. The majority of images were acquired using compression force below 200 N (Fig. 1a) for all of the mammography systems examined. We noted a positive correlation between compression force and CBT (Fig. 1b). The majority of images that exceeded the limit of compression force were for patients who have thicker breasts (CBT > 65 mm) for all of the systems. Although the current study does not explore correlations between compression force and breast density or volume, Mercer et. al have shown that there is a general tendency to apply higher compression for larger volumes and higher BI-RADS grades [37,38]. These studies also have shown that practitioners vary in the amount of compression force applied to breast tissue during mammography exam. In future work, we will investigate the correlations between compression force (and other imaging settings), practitioners, and the patient data that are not present in the DICOM header, such as volume and breast density.

The target-filter selection affects both the radiation dose used for acquiring clinical mammograms and the image quality for clinical diagnosis. In FFDM and DBT, the vendors replaced the traditional Mo targets (found in CR and screen-film mammography) with Rh or W (Fig. 2). In our study, the images acquired with target Rh and W result in a substantial dose reduction in the pair of mammography systems analyzed (Figs. 4a-4b and 4c-4d). The traditional Mo/Mo spectra are no longer considered optimal for mammography [16], particularly for thick and dense breasts. Rh and W have higher atomic numbers, higher K-absorption edges, and higher X-ray energies [16]. Thus, we can acquire images with higher quality and lower radiation dose when using Rh and W targets. Previous studies have shown that the use of the target-filter combinations W/Rh and W/Ag with a digital amorphous selenium detector achieve equivalent contrast-to-noise ratio values with reduced AGD, when compared to Mo/Mo and Mo/Rh [39,40].

Together with target-filter, the tube voltage determines the photon energy of x-rays. Fig. 3 shows that the tube voltage increases with CBT, except in the extreme limits of kV (i.e., high and low kVs) or CBT. The lack of images in these extremes affected our analyses. Fig. 3d showed that system C did not acquire images using 33 kV in our patient population. Similarly, we showed in our previous study that the DBT system D, which is fabricated by the same vendor (Hologic Inc., Bedford, MA), does not frequently acquire images using 37 kV, 39 kV or 41 kV [7]. Note that some vendors, such as Hologic, use tabulated values [28–30] to select appropriate imaging parameters for the AEC [41]. The optimal tube voltage can rise steeply as a result, and breast density largely does not affect the AEC operation.

The tracking system and database developed for this study allowed us to compare differences between systems. The differences in AGD were primarily due to the imaging protocol and technology implemented in each center. For example, for thick and/or radio-graphically dense breasts, alternative target-filter selections can achieve contrast comparable to or better than that obtained with the traditional molybdenum target while using lower radiation dose levels [16]. In this study, the target-filter combinations Mo/Mo and Mo/Rh (where available) are rarely used in systems E and F, respectively (Figs. 2e and 2f).

Patient demographics (e.g., breast density and thickness) also affect the AGD estimation. For example, AGD substantially increases with increasing CBT (Figs. 4 and 8). To reduce the AGD variability that is the result of differences in CBT, we calculated the mean AGD calculated at 50 mm CBT (Table 2). The differences in AGD at 50 mm CBT may be due to equipment calibration, clinical protocol, and radiologists’ preferences for imaging. For instance, systems A and B are distinct models but fabricated by the same vendor and used at the same breast imaging center. The system calibration and imaging protocol used to acquire images in both systems are very similar. In addition, both systems have similar hardware and software technology, which resulted in images acquired using similar levels of radiation dose.

The tracking system also allowed us to contrast differences between populations. For instance, systems B and G are equipped with the same vendor specification (Senographe DS, GE Healthcare), but used in different breast imaging centers and used in different populations. The differences in demographics and imaging techniques substantially affected the AGD estimate (Table 2). We investigated in system B that 69.9%, 7.9%, 14.4% of the images were acquired using the AEC modes standard, dose, and contrast, respectively. On the other hand, we noticed in system G that 6.8%, 0.3%, 91.6% of the images were acquired using the AEC modes standard, dose, and contrast, respectively. We analyzed the use of the imaging parameters using the combined AEC modes because of the differences in the imaging protocols between centers. The lack of images acquired using specific AEC modes would affect the statistics in this study.

While still heavily used worldwide, screening mammography with CR has evolved to FFDM. The sensitivity of breast cancer detection in CR is significantly lower than screen-film mammography [42], and CR images require higher radiation dose [43–45]. CR images have a lower noise (i.e., lower standard deviation in small regions) than FFDM but it has a loss of sharpness that hampers the signal visibility (conspicuity); thus, CR requires more radiation dose to achieve the same image quality and signal-to-noise ratio as FFDM for small objects [36]. In this study, the two CR systems used the highest levels of radiation dose (Fig. 4 and Table 2). However, for the CR systems, the patient AGD was calculated using measurements acquired from dosimetry instruments [22], unlike the AGD reported by the vendor in the DICOM header.

FFDM is evolving towards to DBT because DBT provides reconstructions of the breast volume, which reduces the impact of superimposed breast tissues. There is a growing body of evidence that DBT has increased sensitivity and specificity compared to FFDM [46–48]. The Tomo Combo mode has shown reductions in recall rates and increases in cancer detection rates compared to FFDM alone [49]. However, the total radiation dose delivered to the patient using Tomo Combo mode can be increased by as much as threefold (e.g., the AGD 3D of system F is twice the AGD 2D (Fig. 8 and Table 3). In particular, for system F, which showed the highest correlation between AGD 2D and AGD 3D (slope = 2.0, r²=1.00 and ρ=1.00). System F bases the decision for the 3D exposure most probably on the same input parameters as in 2D, but it applies an extra dose scaling factor that results in double the 2D dose.

A better estimation of the imaging settings can be obtained if patient-specific breast thickness and density are considered. For example, Siemens uses individual breast characteristics determined from a scout image (e.g. CBT and breast density) to calculate the imaging settings. The characterization of these systems in terms of X-ray spectrum is challenging because the AEC acquisition technique is adaptive (Figs. 3e and 3f).

Since the patient-specific breast density is not provided by the vendors, we used the AGD information reported in the DICOM header to develop reference AGD curves that represent the AGD reference for a particular population and system. In this study, an experienced radiologist evaluated the image quality aspects of 15 pairs of patient’s images acquired using systems fabricated by the same vendor. The radiologist reported many benefits with the transition to the newer system F (Figs. 5 and 6). Ideally, to avoid subjectivity and bias in the evaluation, the visual inspection should be performed by more radiologists. The image quality assessment was performed by one radiologist as this was not the primary goal of the project. It is important to highlight that not only the dose but also the system technology and image processing methods can improve image quality. In this study, we could not separate changes in dose and image processing. The differences in image quality were primarily due to the imaging processing. Additionally, the images were acquired using systems equipped with different technologies and equipment calibration, which makes a fair comparison between systems challenging. Differences in CBT are also caused by disparities in the calibration of the height indicator on each system. Note that although we used images collected at the same breast imaging center, the CBT distribution also varied over time (Figs. 2e and 2f).

We are currently investigating more deeply the effect of radiation dose in the image quality of clinical mammograms. The image quality is evaluated in clinical practice using an anisotropic quality index, which assesses the signal-difference-to-noise ratio and dose efficiency in DM and DBT [50–52]. In the future, we will include image quality features and metrics for risk of breast cancer in our tracking system. It is important to stress that our system can identify outliers, such as under and over exposures, which can lead us to look for problems in equipment calibration and/or poor breast positioning (Fig. 7).

We also evaluated 10 mammography exams considered outliers in the data analysis. Fig. 3f showed that the range of CBT dropped significantly for the target/filter combination W/Rh and kV > 32. We noticed that 9 out of 10 images were acquired on patients with breast implants. The breast imaging center has a specific imaging protocol used for breast implants. The imaging protocol information can be extracted from the DICOM header using the tag “protocol name”. However, for these 9 images, the imaging protocol was not filled appropriately. It is important to mention that this represents an exception in this breast imaging center, and again demonstrates the value of on-going, real-time monitoring of image quality and dose.

This study presents some limitations. The AGD calculated using measurements acquired by dosimetry devices (Eq. 1) for the CR systems was not calculated for the FFDM and DBT systems. We have shown in a retrospective study that our tracking system can be used to estimate the patient AGD by including the breast density information and measurements acquired from calibrated devices [8]. We also verified the accuracy of the AGD reported by the FFDM and DBT vendors. For future work, we will use these measurements acquired retrospectively to monitor and estimate the patient AGD prospectively.

The accuracy of the CBT reported in the DICOM header information varies between mammography systems [53]. Correction factors are needed to standardize CBT for each paddle selection (flexible versus non-flexible) and compression force applied. Previous studies have shown that CBT varies over the area of the breast in contact with the compression paddle [54,55]. Differences in compression force applied using flexible and non-flexible paddles are not explored in this study. The equipment calibration and use of different paddles to acquire the clinical images could have affected the poor correlation between compression force and CBT.

The accuracy of the AGD reported in the DICOM header also varies between mammography systems. Although we opted for Dance’s methods [25–27] to estimate AGD in CR, vendors have the choice to report and estimate AGD using alternative methods [31,28–30]. In addition, details of the methods used to estimate AGD are proprietary and not completely clear [41]. The use of alternative methods to monitor and calculate AGD uniformly across vendors is encouraged.

The characterization of X-ray breast imaging systems includes optimization work in terms of the selection of the most suitable imaging parameters, which maximizes the image quality within the limits imposed by breast dosimetry. In DBT, we must also consider optimization of the X-ray tube angular range, number of projections, and reconstruction algorithms. Maldera et al. provide a detailed characterization of four commercial DBT systems, which differ in detector technology (direct/indirect), scan angle, number of projections, tube motion and reconstruction algorithms [56].

5. Conclusion

This study provides a detailed analysis of the imaging settings used in mammography practice of various mammography systems. The systems are used in breast imaging centers located in United States and Brazil. We analyzed the imaging settings (e.g., compression force, target-filter and kV combinations, and radiation dose levels) used in AEC mode. We also analyzed the image quality of patient’s mammograms acquired in different length of time and mammography systems. Differences in the imaging settings were attributed to a combination of factors, such as improvements in technology, imaging protocol and patient demographics.

Our tracking system can provide the estimated AGD based on a specific population and system. The tracking system can match patient-and population-specific breast data using the DICOM information collected from different mammography systems. For future work, we will include image quality metrics in our tracking system to evaluate optimal radiation doses and to minimize the risks of breast imaging.