Skip to main content
NCBI home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2015 Oct 7.
Published in final edited form as: Phys Med Biol. 2014 Sep 11;59(19):5883–5902. doi: 10.1088/0031-9155/59/19/5883

Digital breast tomosynthesis: Studies of the effects of acquisition geometry on contrast-to-noise ratio and observer preference of low-contrast objects in breast phantom images

Mitchell M Goodsitt a, Heang-Ping Chan a, Andrea Schmitz b, Scott Zelakiewicz b, Santosh Telang a, Lubomir Hadjiiski a, Kuanwong Watcharotone c, Mark A Helvie a, Chintana Paramagul a, Colleen Neal a, Emmanuel Christodoulou a, Sandra C Larson a, Paul L Carson a
PMCID: PMC4264665  NIHMSID: NIHMS628754  PMID: 25211509

Abstract

The effect of acquisition geometry in digital breast tomosynthesis (DBT) was evaluated with studies of contrast-to-noise ratios (CNRs) and observer preference. Contrast-detail (CD) test objects in 5 cm thick phantoms with breast-like backgrounds were imaged. Twelve different angular acquisitions (average glandular dose for each ~1.1 mGy) were performed ranging from narrow angle 16° with 17 projection views (16d17p) to wide angle 64d17p. Focal slices of SART-reconstructed images of the CD arrays were selected for CNR computations and the reader preference study. For the latter, pairs of images obtained with different acquisition geometries were randomized and scored by 7 trained readers. The total scores for all images and readings for each acquisition geometry were compared as were the CNRs. In general, readers preferred images acquired with wide angle as opposed to narrow angle geometries. The mean percent preferred was highly correlated with tomosynthesis angle (R=0.91). The highest scoring geometries were 60d21p (95%), 64d17p (80%), and 48d17p (72%); the lowest scoring were 16d17p (4%), 24d9p (17%) and 24d13p (33%). The measured CNRs for the various acquisitions showed much overlap but were overall highest for wide-angle acquisitions. Finally, the mean reader scores were well correlated with the mean CNRs (R=0.83).

Keywords: Digital breast tomosynthesis, acquisition parameters, contrast-to-noise ratio, reader study

1. Introduction

Digital breast tomosynthesis (DBT) is a quasi-3D x-ray breast imaging method that reduces the tissue superposition problems associated with 2D mammography. In DBT, a series of slices are generated that improve the ability of radiologists to distinguish true lesions from overlying and underlying tissues. Radiologists can also better discriminate normal tissues that are separated in depth in DBT, but superimpose on 2D mammograms to falsely appear as lesions. Recent studies confirm the improved sensitivity and specificity of DBT over 2D digital mammography. (Skaane et al., 2013; Ciatto et al., 2013; Rafferty et al., 2013; Rose et al., 2013)

DBT involves the acquisition of a number of low-dose projection images of the patient’s breast over a limited angular range. The resulting reconstructed images have very high spatial resolution, similar to that of 2D digital mammography in the plane of the slices (parallel to the detector plane), but poor spatial resolution in the depth direction. As with conventional linear x-ray tomography, the depth resolution is affected by the angular range. In general, narrow angle tomography and DBT have poorer depth resolution than wide angle tomography and DBT. In contrast to conventional tomography in which a single image at the focal plane is integrated on a film; the series of projection images associated with DBT can be combined (reconstructed) into a full set of focal slices throughout the entire extent of the breast in the depth (thickness) dimension. DBT image quality depends not only on the angular range, but also on the number of projection views and the angular increment between projections. Many researchers have performed studies to determine the optimum geometry (angular range and number of projection views) for DBT. (Eberhard et al., 2006; Chawla et al., 2009; Chawla et al., 2008; Das et al., 2009; Gang et al., 2010; Gifford et al., 2008; Hu et al., 2008; Lu et al., 2011; Maidment et al., 2005; Maidment et al., 2006; Mertelmeier et al., 2008; Nishikawa et al., 2007; Reiser and Nishikawa, 2010; Ren et al., 2006; Sechopoulos, 2013; Sechopoulos and Ghetti, 2009; Tucker et al., 2012; Tucker et al., 2013; Van de Sompel et al., 2011; Wu et al., 2004; Ren et al., 2009, Young et al., 2013) Most of these studies have involved modeling of the tomosynthesis systems and some included modeling of the observers. Very few of the investigations have been experimental, in most part due to the unavailability of DBT systems that permit investigation of a wide variety of geometries, and because of the unavailability of realistic breast-simulating physical phantoms. We have a unique prototype General Electric tomosynthesis system at our institution that was recently modified to include an “advanced research mode” for phantom imaging that allows us to vary the acquisition angular range and angular increment. In this paper, we describe studies that were performed to evaluate the effects of acquisition geometry on observer preference of DBT images of contrast-detail (CD) test objects within phantoms with breast-like backgrounds and on the measured contrast-to-noise ratios (CNRs) of those test objects. The preference study allowed us to subjectively evaluate the overall impressions of the observers of their perception of CD test objects in DBT images including factors such as contrast, sharpness, noise, artifacts, background smoothness, etc. without the time-consuming task of examining each factor individually. Preference studies have been used by investigators to determine the optimal kVps for dual energy digital radiography of the chest (Shkumat et al., 2008; Williams et al., 2007), to evaluate radiologists’ visual judgments of the similarity of masses and/or microcalcifications in mammograms (Muramatsu et al., 2007; Tourassi et al., 2013), and to compare digital with screen-film radiography. (Hamers et al., 2001)

2. Methods and Materials

2.1. Prototype tomosynthesis system

A General Electric (GE) GEN 2 prototype DBT system was used to generate the tomosynthesis images in this study. The system (Fig. 1A) has an Rh target, Rh filter x-ray source and a high detective quantum efficiency CsI/a-Si digital detector, which is stationary during exposure. The detector has a 0.1 mm × 0.1 mm pixel pitch and is of the same design as the detector used in the commercial GE Essential digital mammography system (General Electric Healthcare Systems, Milwaukee, WI) except that the overall active area (19.2 cm × 23.4 cm) is smaller. The prototype system was designed and built by scientists at GE Global Research (Niskayuna, NY). It has been n used to image patients in several IRB-approved NIH-funded research projects. (Eberhard et al., 2006) For those projects, the system is operates in a 60° acquisition angle, 21 projection view (3° angular increment) mode (Fig. 1B) with a total scan time of about 7.5 seconds. The projections are acquired in a step-and-shoot manner, whereby the x-ray tube is stationary during the acquisition of each projection view. Recently, we worked with scientists at GE Global Research to develop an advanced mode which permits the acquisition of images with user selectable acquisition geometries. The total tomosynthesis angle can range up to about 80°, the angular increments are integral multiples of 1°, and variable angular increments can be used.

Figure 1.

Figure 1

Open in a new tab

(A) GE GEN2 prototype tomosynthesis system. (B) Standard prototype DBT system acquisition geometry (60°, 21 projection view, 3° angular increment) for imaging patients and phantoms.

2.2 Phantoms

Modular breast-simulating phantoms were imaged in our study. These phantoms consist of multiple 1-cm-thick heterogeneous slabs that mimic the glandular and adipose tissue composition and parenchymal patterns of a human breast. The slabs are made of epoxy resins with x-ray attenuation properties of 50% glandular/50% adipose breast tissue. The slabs were custom manufactured for us by CIRS, Inc. (Norfolk, VA). They are similar to those in the commercial CIRS model 20 mammography 3D phantom, but our slabs have more cloud-like rather than swirl-like patterns. One of the slabs with a cloud-like pattern was machined to include a CD test pattern (Fig 2A). This pattern consists of flat bottomed cylindrical holes of diameters 0.5, 1, 2, 3, 4, and 5 mm and of depths 0.2, 0.4, 0.6, 0.8 and 1 mm. For the present studies which were most concerned with perception of small mass-like objects, we limited our CNR analysis to all holes except the smallest (0.5 mm). A CD array of holes was used because of the technical difficulty and high cost of inserting dense CD objects in the breast phantom slab. The commercial CDMAM phantom has dense gold disks but it has a clear acrylic background that is too uniform even if it is sandwiched between the heterogeneous breast phantom slabs. Although the holes in our phantom are less attenuating than the background as opposed to real masses, which are more attenuating than the background, CD objects like these still provide useful differential contrast for reader preference and perception studies. The CD slab was sandwiched between other heterogeneous slabs (Fig. 2 B and C) to create phantoms with a total thickness of 5 cm. The positions of the slabs including the CD slab itself were varied to create 4 different modular phantoms containing the CD slab. Each of the 4 phantoms was unique in that none had the exact same heterogeneous slabs both immediately above and below the CD slab, and the CD slab was at different heights in 3 of the phantoms (2, 3, and 4 cm from the breast support plate) and upside down in one.

Figure 2.

Figure 2

Open in a new tab

Photographs of: (A) the CD slab, B) one modular phantom configuration (CD phantom in center position), and (C) another modular phantom configuration placed on the breast support plate (CD phantom is second from the top).

2.3. Image acquisition

All phantoms were imaged using x-ray technique factors of Rh target, Rh filter, 29 kVp, with the total mAs (50 mAs) divided equally amongst the projection views. The measured half-value layer (HVL) of the x-ray beam incident on the phantoms was 0.407 mm aluminum. Using Wu et al’s methodology (Sobol and Wu, 1997; Wu et al., 1994) and the fact that the dose weighting factors for each angular projection of a tomosynthesis acquisition are approximately 1.0 (Sechopoulos et al., 2007), the glandular dose in each case was computed to be approximately 1.1 mGy. This is about equal to the average glandular dose in our clinic for digital mammograms of 5-cm-thick, 50% adipose/50% glandular tissue phantoms with a GE Essential digital mammography systems (29 kVp, Rh/Rh, 0.437 mm Al HVL, 46.9 mAs). The twelve combinations of acquisition angles and angular increments that were studied are listed in Table I.

Table I.

Twelve combinations of tomosynthesis angle and angular increments investigated. The angular increments for 40d13pv were 0°, ±1°, ±3°, ±6°, ±10°, ±15°, ±20°. The angular increments for 60d17pv were 0°, ±6°, ±12°, ±15°, ±18°, ±21°, ±24°, ±27°, ±30°. Two of the combinations (24d9p and 32d17p) were repeated and are marked by*.

Tomosynthesis angle (degrees) Number of projection views Projection angular increments (degrees) Descriptor (d=degrees, p=number of projections)
16 17 1 16d17p
24 9 3 24d9p*
24 13 2 24d13p
32 17 2 32d17p*
36 13 3 36d13p
40 11 4 40d11p
40 13 variable 40d13pv
48 13 4 48d13p
48 17 3 48d17p
60 17 variable 60d17pv
60 21 3 60d21p
64 17 4 64d17p
Open in a new tab

In most cases, equal angular increments were used between projections. However, for two acquisitions, variable angles were used. These are the acquisitions with descriptors “40d13pv” and “60d17pv” in Table I. The angular increments for 40d13pv were 0°, ±1°, ±3°, ±6°, ±10°, ±15°, ±20°. This represents an acquisition that is dense in the middle (finer angles) and sparse at the sides. The angular increments for 60d17pv were 0°, ±6°, ±12°, ±15°, ±18°, ±21°, ±24°, ±27°, ±30°. This represents an acquisition that is dense at the sides and sparse in the middle.

The tomosynthesis angles and angular increments were chosen to cover the range used in clinical, prototype, and research systems. For example, 16d17p is similar to the narrow angle, 15 degrees 15 projection view geometry used in the Hologic Selenia Dimensions system (Hologic, Inc. Bedford, MA)(Sechopoulos, 2013), 24d9p is similar to the 25 degrees 9 projection views used in the GE SenoClaire DBT system (General Electric Healthcare, Milwaukee, WI)(Sechopoulos, 2013), 40d13pv is similar to the dense in the middle, sparse at the sides acquisition technique that IMS Giotto has described on their website (Internazionale Medico Scientifica (IMS), Giotto, Bologna, Italy http://www.imsitaly.eu/19/dbt), 32d17p is similar to the 30 degrees 15 projection views used in the Planmed Nuance Excel DBT system (Planmed Oy, Helskinki Finland) (Sechopoulos, 2013), and 60d21p is the wide angle geometry used on the GE Gen 2 research tomosynthesis unit in our laboratory. For our study, to investigate reproducibility, two of the acquisition geometries (24d9p and 32d17p) were repeated a second time for all phantom configurations on a different day. Thus, for the study, there were a total of 14 acquisition geometries performed (the 12 listed in Table I, plus the 2 repeats).

2.4 Image reconstruction and selection

All images were reconstructed using a simultaneous algebraic reconstruction technique (SART) (Zhang et al., 2006) with two iterations. The pixel dimensions in the reconstructed images were 0.1 mm × 0.1 mm, and there was no binning of pixels. The reconstructed slices were 1 mm thick, and in each case, the slice in which the entire CD pattern was best in focus was used for the evaluations and comparisons. The decision to use the single “best overall” slice rather than more slices, such as the ones above and below the single slice was made based on inspection of many images from which it was found that the perception of additional subtle disks was not improved in the adjacent slices. The inspection of the images also indicated that for a given CD pattern position (depth), the best slice was the same for all acquisition geometries.

2.5. Reader study

2.5.1 Image preparation

The reconstructed focal slices containing the CD pattern were cropped to only include rectangular regions of identical size encompassing the CD test pattern. The window level and width settings for the images were determined experimentally, with the goal to maximize the perception of the CD objects for each image with a standardized method. The window level was set at the mean pixel value of the cropped image. The window width was determined using the criteria that the width be narrow enough to enhance the contrast of the CD objects, but not too narrow as to make the image noise distracting. By varying the window width settings and visually judging the overall appearance of the CD images for a subset of images from the different acquisitions, two medical physicists, each with over 30 years of experience in analyzing x-ray images, found that the ratios of the empirically chosen window width to the standard deviation of the pixel values in the cropped images of the phantoms containing the CD insert fell in a narrow range of about 13 to 17, with an average of about 15. To reduce subjectivity, the window width for each acquisition geometry was estimated to be about 15 times the average of the standard deviations over the four phantom configurations, with rounding to the 10’s, as a compromise for all imaging conditions. The window widths derived by this method and the average standard deviations for each acquisition geometry are listed in Table 2. The average standard deviations ranged from 29.9 to 47.8

TABLE 2.

Average standard deviations of the pixel values in the cropped images of the CD test object for the 4 phantoms imaged with the various acquisition geometries and the window widths used in the display of the images for the reader study.

Acquisition Geometry Descriptor (d=degrees, p=number of projections) Average Standard deviation of cropped images Window Width
16d 17p 37.2 550
24d 9p* 29.9/31.3 450/470
24d 13p 35.7 540
32d 17p* 40.8/41.0 620/620
36d 13p 36.8 560
40d 11p 34.0 510
40d 13pv 36.1 550
48d 13p 38.0 580
48d 17p 42.4 640
60d 17pv 44.0 650
60d 21p 47.8 750
64d 17p 45.8 690
Open in a new tab
*

The acquisitions for all phantom configurations were repeated twice for the 24d9p and 32d17p geometries. The standard deviations and window widths are listed for each set individually.

When this method was applied, images with smaller standard deviations (lower noise) were displayed with smaller window widths which increased the contrast; whereas, images with larger standard deviations (greater noise) were displayed with larger window widths, which reduced the contrast. The pixel variations of the background structures were therefore displayed approximately in similar ranges for all images, facilitating comparison of the relative visibility of the CD objects in the paired images in the observer experiment. Examples of images displayed using our method, and methods whereby all images are displayed with the same window widths are shown in Fig. 3.

Figure 3.

Figure 3

Open in a new tab

Focal slices of the CD insert in the same 5-cm-thick modular phantom acquired with 16d17p (left column), 24d9p (middle column) and 60d21p (right column) geometries. Window widths for top row are those corresponding with about 15x the standard deviation (left to right: 550, 450, and 750). Window width for middle row is 450 for all, and window width for bottom row is 750 for all.

Notice that using a fairly narrow window width (450) for all of the images (middle row) results in an overly contrasty appearance of some of the images (e.g., 60d21p); whereas, using a fairly wide window width (750) for all images (bottom row) results in a washed out appearance of some of the images (e.g., 16d17p).

For the reader study, images with window levels and widths set by our method were displayed side by side on a 21.3″, 5 Megapixel LCD monitor (EIZO SMD 21500 D, 2048 × 2560, 800:1 contrast ratio, 750 Cd/m2 maximum luminance). The monitor was calibrated using the DICOM grayscale standard display function. The combination of 14 acquisitions compared 2 at a time resulted in a total of 91 image pairs to evaluate for each phantom configuration. Thus for the 4 phantom configurations, a total of 364 image pairs were rated. The images were randomized so that each reader viewed image pairs in a different order.

2.5.2 Readers and scoring of images

Seven readers participated in this study. Four are MQSA-qualified medical physicists (readers 1–4 in Fig. 6) with 13–28 years of mammography experience and three are MQSA-approved breast imaging radiologists with 5–31 years of mammography experience (readers 5–7 in Fig. 6). The study reading time was unlimited for each reader. The readers were instructed to grade the images on an 11-point (−5 to 5) scale, with −5 indicating the image on the left was much better quality than the image on the right, 0 indicating the images on the left and right were of equal quality, and 5 indicating the image on the right was much better quality than the image on the left. In this context, “quality” referred to the contrast and visibility of the CD holes in the phantom images. Note that the sign of the rating is used only to keep track of whether the left or right image is better. Furthermore, the study was designed such that the chance that an image in a pair was displayed on the left or on the right was basically random. Therefore, only the absolute value of the rating is relevant for the preference analysis and the rating scale is equivalent to 0 to 5, with 0 being equally preferred and 1 to 5 ranging from slightly preferred to strongly preferred. Each reader underwent a training session, examining and rating 30 image pairs to become familiar with the task, the rating scale, and the user interface. The 0–5 scale has an advantage over a 3-point scale of forcing the readers to scrutinize the images more carefully; however, a disadvantage of such a fine scale is the increased variability in the inter-reader scoring. To take advantage of both scales, we used the 0–5 scale for reading and converted the scores to 3 values (1, 0.5, and 0) for the paired comparison analysis, where 1 represents “preferred”, 0.5 represents ”equally preferred” and 0 represents ”not preferred.” The 0–5 scale was converted to the 3-point scale as follows: If one image in a pair was preferred over the other (i.e., an absolute rating of 1 to 5), the preferred image was assigned a score of “1” and the image that was not preferred was assigned a score of “0”. If both images were equally preferred (i.e., an absolute rating of 0 on the 0–5 scale), both images were assigned a score of 0.5. Thus, the maximum score for a given phantom and a given acquisition geometry is 13. This would occur if the focal slice image of the CD phantom generated with a particular geometry received a score of 1 (preferred) for all comparisons with the images generated with the 13 other geometries.

Figure 6.

Figure 6

Open in a new tab

Preference percentage vs. acquisition geometry for all 7 readers.

For the comparisons, a preference percentage was computed that accounted for all acquisition geometries and all modular phantoms. The preference percentage is defined as the number of times the focal slice image of the CD slab is preferred for one acquisition geometry over corresponding images for all other 13 geometries and 4 phantoms divided by the maximum possible score (13 times 4 = 52) multiplied by 100%. The 3-level scoring system is similar to that used in an eye exam to determine a patient’s eyeglass prescription. In that test, the patient is asked which image of an eye chart she or he prefers as an ophthalmologist or optometrist places different lenses in front of the patient’s eyes. However, in our case the images are viewed side by side rather than consecutively.

2.5.3 Contrast–to-Noise Ratio Computation

A plug-in for the public domain NIH ImageJ display program (Schneider et al., 2012) was developed to compute the CNRs of each of the holes in the reconstructed tomosynthesis slices passing through the centers of the CD objects. The ROIs were all positioned manually and they were sized to be well within the holes for the determination of the mean pixel values within the holes, as shown in Fig. 4. The ROI size for the objects varied with the hole size. Four ROIs all of equal size (equal to the size of the ROI in the largest hole (8.12 mm2)) were positioned in the background regions around each hole. The plug-in provides the following functions: (1) storing of templates for the positions of the ROIs so the same ROI set can be used for all images of the same phantom, (2) manual editing of the positions and sizes of individual and grouped ROIs, and (3) storing of the measured mean values and standard deviations for each of the ROIs.

Figure 4.

Figure 4

Open in a new tab

Tomosynthesis slice showing object and background regions of interest for estimation of CNR of each object. The 5, 4, 3, 2 and 1 mm diameter (left to right), and 1, 0.8 and 0.6 mm depth (top to bottom) holes are displayed.

The CNRs were computed using the equation:

CNR=MBackground-MDiskσBackground (1)

where MBackground is the average of the mean pixel values of the 4 background ROIs surrounding the disk, MDisk is the mean pixel value of the ROI inside the target disk, and σBackground is the square root of the total background variance, which, using the law of total variance is equal to the mean of the variances of the 4 background ROIs plus the variance of the means of the 4 background ROIs.

CNRs were compared. For each acquisition geometry, the CNRs of the disks of each size and depth in the focal images of the contrast-detail test object were computed. In addition, the mean CNR for a given acquisition geometry was calculated by averaging the CNRs over all disk sizes and depths on all images acquired with the same geometry. For these calculations and comparisons, the number of disks was kept the same for each geometry. If a particular disk could not be visualized in a particular geometry regardless of window level and width and zoom adjustment, then the CNR for that disk was not computed for any of the geometries. This procedure was used to avoid penalizing geometries that produced images which displayed more subtle disks that would have lower CNRs.

2.5.4 Statistical analysis

The reader study and CNR results were separately evaluated. Reader (e.g., reader 1, reader 2, reader 3, etc.) was treated as a random factor for the reader study, while image replication, CD test object depth, and CD test object diameter were treated as control factors for the CNR study. Mixed model regression analysis was used to assess the geometries. Tukey’s adjustment was used to control experimental error for post hoc tests. A 5% level of significance was used to evaluate statistical significance.

3. Results

3.1. Reader preference study

Focal slices of the CD pattern sandwiched between heterogeneous slabs generated with 16d17p acquisitions for all 4 phantom configurations are shown in Fig. 5, below. Notice that the background varies in these images because the slabs above and below the CD insert are different in each case, which results in the interplane blurring of different patterns from above and below the CD insert extending into the focal plane. The degree of interplane blurring depends on the acquisition geometry.

Figure 5.

Figure 5

Open in a new tab

DBT images of the CD slab sandwiched between heterogeneous slabs in each of the 4 different modular phantom configurations acquired with the narrow angle 16d17p geometry.

The preference percentages for all 7 readers from the reader study are plotted in rank order in Fig. 6. To illustrate some of the results in this plot, consider the values for the 60d21p acquisition geometry. Reader 1 had a score of 51.5 out of a possible 52, which is equal to a preference percentage of 99%, and reader 7 had a score of 42.5, which is equal to a preference percentage of 81.7%. Similarly, for comparisons of the 16d 17p geometry vs. all other geometries, reader 1 had a score of 5 which is equal to a preference percentage of 9.6%; whereas, reader 7 had a score of 0.5, which is equal to a preference percentage of 0.5%. It is noted that the preference percentages for the medical physicist readers (1–4) and radiologist readers (5–7) are very similar. A plot of the mean preference percentage for the data in Fig. 6 vs. the tomosynthesis angles of the acquisition geometries is shown in Fig. 7.

Figure 7.

Figure 7

Open in a new tab

Mean preference percentage over the 7 readers, for a given acquisition geometry plotted against the tomosynthesis angle of that geometry for the CD focal slices. The linear regression equation for this data is y=0.0155x – 0.1097, with a correlation coefficient R of 0.91.

A box plot of the CNR results for the CD phantom images is displayed in Fig. 8. This box plot was generated with vertex42 software and Microsoft Excel (http://www.vertex42.com/ExcelTemplates/box-whisker-plot.html). The reader preference for a given acquisition geometry plotted against the mean CNR of the disks of the CD focal slices for that geometry is shown in Fig. 9.

Figure 8.

Figure 8

Open in a new tab

Box plot of the calculated CNRs of the disks in the CD focal slices for the 4 modular phantom configurations, obtained with various acquisition geometries. The geometries are arranged in the same order as the rankings found in the reader study (Fig. 6). The box part of the plot represents the central 50% of the data, the line in the box is the median, the upper edge of the box is the 75th percentile and the bottom edge is the 25th percentile. The upper whisker extends from the 75th percentile to either 1.5 times the interquartile range (75th percentile – 25th percentile) or the maximum data value, whichever is less, and the lower whisker extends from the 25th percentile to either that value minus 1.5 times the interquartile range or the minimum data value, whichever is less. Maximum and minimum outliers are indicated by asterisks. (Only 1 outlier, a maximum outlier was observed.)

Figure 9.

Figure 9

Open in a new tab

Plot of mean preference percentage for a given acquisition geometry against the mean CNR of the disks in the CD focal slices for that geometry. The linear regression equation for this data is y=0.845x – 0.5435, with a correlation coefficient R of 0.83.

Comparisons of the reader preference percentages and mean CNRs for the acquisition geometries that were repeated and the acquisitions that used the same tomosynthesis angle but different numbers of projection views are listed in Table 3, below.

Table 3.

Mean reader preference percentages and mean CNRs for acquisitions that were repeated and acquisitions that used the same tomosynthesis angles but different numbers of projection views.

A) Repeat exposure study
24d9p(1) 24d9p(2) P-value 32d17p(1) 32d17p(2) p-value
Mean preference percentage 18.3% 15.5% 1.000 59.1% 48.4% 0.178
Mean CNR 1.08 1.04 1.000 1.05 1.04 1.000
B) Acquisitions with same tomosynthesis angles but different numbers of projection views
24d13p 24d9p(1) & 24d9p(2) p-value 40d13p var 40d11p p-value 48d17p 48d13p p-value
Mean preference percentage 32.8% 18.3%
15.5%		 0.009
0.001		 44.5% 44.9% 1.000 71.6% 58.1% 0.024
Mean CNR 0.98 1.08
1.04		 1.000
1.000		 1.24 1.38 0.988 1.37 1.41 1.000
Open in a new tab

4. Discussion

The plots in Figs. 6 and 7 indicate that for perception of CD test objects that are 1 mm to 5 mm in diameter, the readers in our study tended to prefer DBT images generated using acquisition geometries with wider tomosynthesis angles. The average reader preference percentage is highly correlated with the tomosynthesis angle (R= 0.91). The top rated acquisition geometries were 60d21p, 64d17p, 48d17p and 60d17p with variable angular increments, with readers on average preferring images generated with these acquisitions over those generated with other acquisition geometries 94.5%, 80.4%, 71.6% and 69.2% of the time, respectively. The least preferred images of the heterogeneous CD phantom were those generated with narrow angles of 16 and 24 degrees, with average preference percentages of only 3.6% for 16d17p and 15.5% and 18.3% for 24d9p. Average reader preference percentages for repeat scans with the 24d9p and 32d17p acquisition geometries were within 10.7%. These differences in the preference percentages were not statistically significant for the repeated acquisitions.

For the two geometries in which the same tomosynthesis angle was used but with different numbers of equally spaced projections, readers preferred images with larger numbers of projections and this preference was significant (24d13p vs. 24d9p and 48d17p vs. 48d13p in Table 3B.) The use of the same tomosynthesis angle, 40 degrees, but a different distribution of projection views (dense in the middle, sparse at the sides) and slightly different number of projections (13 vs. 11) had minimal effect (40d13pvar vs. 40d11p in Table 3B.)

The CNR values for the CD targets in the images show much overlap in their distributions for the various acquisition geometries (Fig. 8 and Table 3B) but 30 of the 91 pairwise comparisons (33%) are significant (p<0.05). The greatest mean CNRs for the CD test object were obtained for the widest tomosynthesis angle acquisitions (64d17p, 60d17pv, and the 60d21p) with mean CNRs of 1.5 to 1.6. The lowest CNRs were for the narrowest angle 16d17p acquisition for which the mean CNR (0.74) was about half of those for the 3 widest angle and all of the differences are significant (p<0.0001). The mean CNRs for 24d9p were the second lowest, with mean CNRs (1.04 and 1.08) that were about 68% of those for the 3 widest angles, and those differences were also all significant (p<0.006). Despite the large overlaps in the distributions of the CNR values for the images acquired with the different geometries, and some discrepancies in the rankings compared with those based on reader preference (Fig. 8), the mean CNRs are fairly strongly correlated with the mean reader preference scores (R=0.83), as shown in Fig. 9.

Determining the optimum geometry for DBT is not a simple matter. Aside from the improved depth perception with increased tomosynthesis angle, there is a desire to use greater numbers of projections for improved angular sampling. However, at a fixed total x-ray exposure and a fixed tomosynthesis angle for a scan, the total amount of detector electronic noise will increase relative to the total x-ray exposure as the number of projections increases in the scan. This is because each additional projection will introduce additional electronic noise, which will increase the overall noise of the reconstructed volume while the x-ray noise (quantum mottle) of the reconstructed volume stays approximately the same. This problem will be more prominent if the detector noise is high. The impact of increasing the number of projections on the noise of the reconstructed images was well demonstrated in the simulation study by Sechopoulos and Ghetti (Sechopoulos and Ghetti, 2009). In addition, the total scan time affects motion artifacts and generally increases with both the total angle scanned and the number of projection views. Researchers have also found that increased angular range can result in a decrease in planar spatial resolution for both direct and indirect detectors (Acciavatti and Maidment, 2010, 2011; Badano et al., 2011; Badano et al., 2007; Freed et al., 2010; Mainprize et al., 2006; Reiser et al., 2009; Zhao and Zhao, 2008). On the other hand, undersampling artifacts may arise when too few projections are acquired (Maidment et al., 2006; Reiser et al., 2009).

Our reader study results tend to confirm the improved performance for wider tomosynthesis angles that was realized by others. We did not perform a thorough investigation of the effect of the number of projection views on reader preference, but we did find that within the limited range studied, for a given total tomosynthesis angle, use of more projection views results in improved image quality (e.g., 48d17p had a higher preference score over 48d13p, and 24d13p had a higher preference score over 24d9p). This is likely due to the decrease in undersampling artifacts from objects in off-focus planes with the use of more projection views (Hu et al., 2008).

Our results agree with the general trend of higher scores for images acquired with greater tomosynthesis angles found in a simulation study performed by Sechopoulos and Ghetti. (Sechopoulos and Ghetti, 2009), but differ in the effects of the number of projection views. Sechopolous and Ghetti used a quality factor (QF) equal to the normalized CNR for a simulated 3-mm-radius spherical mass made of 65% glandular tissue divided by the artifact spread function (ASF) width for a calcification. Their quality factor did not include human perception modeling or the effect of other artifacts such as the undersampling artifact. Their results generally showed an opposite trend to ours, with higher QFs at lower numbers of projection views. For example, at 24° tomosynthesis angle, the QFs were ranked as 24d5p > 24d9p > 24d13p; at 48° tomosynthesis angle, the QFs were 48d9p > 48d17p > 48d13p > 48d21p > 48d25p > 48d31p > 48d41p; and at 60° tomosynthesis angle, the QFs were 60d13p > 60d9p > 60d17p > 60d25p > 60d21p > 60d31p > 60d41p > 60d61p.

The acquisition geometries that Sechopoulos and Ghetti (Sechopoulos and Ghetti, 2009) found were better than the best one in our study (60d21p) were, in order, 60d13p, 60d9p, 60d17p, 40d9p, 48d9p, and 60d25p. We did not include 60d13p or 60d9p in our study because in a preliminary investigation we observed that angular increments greater than about 3° produced noticeable undersampling artifacts in the reconstructed slices. The Sechopoulos and Ghetti simulation was similar to our experimental study in that they simulated a GE Essential digital mammography flat panel detector that is like our detector except larger in area, a 5-cm-thick 50% glandular/50% adipose breast like our modular phantom, and step-and-shoot acquisition like our system. Some differences include their filtering of 50 different white noise volumes to generate 50 different tissue backgrounds, the use of a single size lesion (3 mm radius sphere made of 65% glandular tissue/35% adipose) and a 0.4 mm × 0.4mm × 0.4mm cube of CaCO3 to simulate a calcifications. They also used different reconstruction (maximum likelihood expectation maximization (MLEM) vs. our SART), average glandular dose (2 mGy vs. our 1.1 mGy) and x-ray spectrum (26 kVp Mo target/Mo filter vs. our 29kVp Rh/Rh). All of these differences may have affected the results.

In an experimental study using a Hologic a-Se flat panel detector and 31 stationary carbon nano-tube x-ray sources, Tucker et al (Tucker et al., 2012) used a quality factor similar to that of Sechopoulos and Ghetti (Sechopoulos and Ghetti, 2009)(Signal-difference-to-noise ratio / width of ASF at half max). They compared DBT images of an ACR mammography accreditation phantom generated with three different combinations of angular span and number of projection views (14d15p, 28d15p and 28d29p) and 3 dose distributions. The best QF was obtained with the 28d15p (2° increments) equal dose distribution (EDD) acquisition (QF=1.211). The lowest QF was obtained with the 14d15p (1° increments) EDD (QF=0.627) acquisition. Their highest rated, 28d15p EDD acquisition is similar to our 32d17p, which produced the fifth most preferred images in our reader study, and their lowest rated, narrowest angle, 14d15p EDD acquisition performed similarly to our narrowest angle 16d17p acquisition which produced the least preferred images in our reader study.

Another study that can be compared to ours is that of Chawla et al (Chawla et al., 2009). Their study combined experiment and simulation. They acquired images of 5 mastectomy samples with a Siemens Mammomat Novation DBT system (Siemens Healthcare, Erlangen, Germany) and embedded 84 synthesized 3 mm 3D lesions at the centers of 84 non-overlapping regions of interest. They used a Laguerre-Gauss channelized Hoteling observer model based measure of lesion detectability. Areas under receiver-operator characteristic (ROC) curves (AUC) were compared for iso-dose acquisitions with 7.5° to 45° angular ranges and 1 to 25 projections. The best performance was obtained for 45° and 15–17 projection views. Like Sechopoulos and Ghetti (Sechopoulos and Ghetti, 2009) and Tucker et al (Tucker et al., 2012), this best performance was at the widest angle in the study. For a fixed dose, Chawla et al found that increasing the number of projections beyond a certain limit decreased performance. They determined that the ideal number of projections increases linearly with the tomosynthesis angle, and the best performance for all tomosynthesis angles occurs when the angular increment is about 2.75°. Our reader preference percentages are in agreement, as the highest percentages are for the widest angle (60°) with an angular increment of 3 °, which is the closest angular increment on our system to 2.75°.

Based on the trend in our results, one might conclude that a tomosynthesis angle larger than 60 degrees might be best. Indeed, the median CNR for 64d17p (1.60) was greater than that for 60d21p (1.35) (Fig. 8). However, in the reader study, 60d21p was preferred over 64d17p (94.5% vs. 80.4%) and the results were significant (p=0.0002). The lower preference of 64d17p for human readers may be attributed to the undersampling reconstruction artifacts from the 4-degree angular increment between the projections; the artifacts could be perceived by human readers but were not accounted for in the CNR calculations. In addition, the image noise tends to increase with increased incidence angle because of the longer pathlength of x-ray attenuation in the breast phantom. Greater incidence angles to the detector also reduce planar spatial resolution for both direct and indirect detectors (Acciavatti and Maidment, 2010, 2011; Badano et al., 2011; Badano et al., 2007; Freed et al., 2010; Mainprize et al., 2006; Reiser et al., 2009; Zhao and Zhao, 2008). All of these factors counter the advantage of reduced effective slice thickness with increased tomosynthesis angle. Thus, there will be a limit to the extent to which image quality improves with increased tomosynthesis angle.

Finally, the results of our study differ from those predicted by the work of Young et al (Young et al, 2013). These researchers “developed a virtual trial framework for task-specific DBT assessment that uses digital phantoms, open-source x-ray transport codes and a projection-space, spatial-domain observer model for quantitative system evaluation.” They performed a study with this framework and found that the detectability (AUC)) of small (3 mm diameter) simulated masses “embedded in randomly-varying anatomical backgrounds” remained constant or increased slightly with tomosynthesis scan angle. There are a number of possible reasons why their predictions are different, including: 1) their AUCs were computed from raw, unfiltered projection images rather than from reconstructed images, 2) their pixel size (500 micron × 500 micron) was much larger than ours (100 micron × 100 micron), 3) their simulated background breast anatomy contained more high frequency Cooper’s ligament-like structures than ours, 4) they neglected various characteristics of a real detector including spatial blur, and 5) their virtual trial model did not take into account inter-plane and in-plane blurring, which depend strongly on the tomosynthesis scan angle and angular increment and can affect readers’ perception of targets on the reconstructed images.

5. Summary and Conclusions

Our reader preference study and CNR measurements of DBT images of CD test objects in heterogeneous backgrounds sandwiched between heterogeneous slabs indicate images acquired with wide tomosynthesis angles (e.g., 60 degrees) have higher CNR and are preferred by readers over those acquired with narrow angles (e.g., 16 and 24 degrees). The reader preference scores are highly correlated with the tomosynthesis angle (R=0.91) and well correlated with the mean CNR measurements (R=0.83). Although the trends for the reader study and CNR measurements are similar, the rankings of the various acquisition geometries for the reader preference study and the CNR measurements show differences. One reason for these differences is that human readers are known to pre-whiten; therefore, the frequency characteristics of the signal and noise components in the image will affect human performance; whereas, CNR measurements do not include the frequency information. In addition, CNR does not account for background correlation, which depends on tomosynthesis acquisition parameters, in particular the angular range. The general conclusion that wider angles are preferred is in agreement with the work of other researchers.

The results of our study are for objects 1 to 5 mm in diameter and may be applicable to mass perception. The optimal acquisition geometry should be a compromise that accounts for the effects of tomosynthesis angle, angular increment and number of projections on perception of calcifications, architectural distortions, and soft tissue lesions and accounts for increased motion blur with increased scan time. It is interesting to note that we recently completed a phantom study of calcification detection using the same GE prototype system and found that calcification detection is best for narrow angle tomosynthesis (Chan, 2014).

Finally, the following limitations must be acknowledged with respect to extrapolation of the results of the present study to those for some commercial and prototype DBT systems. Specifically, although the acquisition angles and angular increments of some of the commercial and prototype DBT systems were approximated in this study, other aspects of these systems were not, including: the continuous x-ray tube motion of the Hologic, Planmed, and Siemens systems, the rotation of the detector in the Hologic and Planmed systems, the use of different x-ray tube targets and filters (e.g., W targets with Al filter for Hologic, with Rh or Ag filters for IMS Giotto and Planmed, and with an Rh filter for Siemens), the use of different detectors (e.g., a-Se for Hologic, IMS Giotto, Planmed and Siemens), the use of pixel binning (Hologic), the use of different reconstruction algorithms (e.g., filtered back projection for Hologic and Siemens, iterative for Planmed, adaptive statistical iterative reconstruction (ASiR) for DBT for the GE SenoClaire, and iterative with total variation regularization for IMS Giotto), and the use of a physical moving scatter rejection grid for the GE SenoClaire system.

Acknowledgments

This work is supported by USPHS grants RO1 CA151443 and RO1 CA091713. The digital breast tomosynthesis system was developed by scientists at GE Global Research (Niskayuna, NY), with input and some revisions from the University of Michigan investigators, through the Biomedical Research Partnership (USPHS grant CA091713, PI: Paul Carson, Ph.D.) collaboration. The biostatistician Kuanwong Watcharotone, PhD was supported in part by the National Center for Advancing Translational Sciences of the National Institutes of Health under Award Number UL1TR000433. The content of this paper does not necessarily reflect the position of the funding agencies and no official endorsement of any equipment and product of any companies mentioned should be inferred. MMG and HPC had control of the data and analysis submitted for publication. AS and SZ are employees of GE Global Research, but their efforts in this project were supported by subcontracts from the University of Michigan NIH grants CA151443 and RO1 CA091713.

References

  1. Acciavatti RJ, Maidment ADA. Calculation of OTF, NPS, and DQE for Oblique X-Ray Incidence on Turbid Granular Phosphors. In: Marti J, et al., editors. Digital Mammography. 2010. pp. 436–43. [DOI] [PMC free article] [PubMed] [Google Scholar]
  2. Acciavatti RJ, Maidment ADA. Optimization of phosphor-based detector design for oblique x-ray incidence in digital breast tomosynthesis. Medical physics. 2011;38:6188–202. doi: 10.1118/1.3639999. [DOI] [PMC free article] [PubMed] [Google Scholar]
  3. Badano A, Freed M, Fang Y. Oblique incidence effects in direct x-ray detectors: A first-order approximation using a physics-based analytical model. Medical physics. 2011;38:2095–8. doi: 10.1118/1.3567497. [DOI] [PubMed] [Google Scholar]
  4. Badano A, Kyprianou IS, Jennings RJ, Sempau J. Anisotropic imaging performance in breast tomosynthesis. Medical physics. 2007;34:4076–91. doi: 10.1118/1.2779943. [DOI] [PubMed] [Google Scholar]
  5. Chan H-P, Goodsitt MM, Helvie MA, Zelakiewicz S, Schmitz A, Paramagul C, Roubidoux MA, Nees AV, Noroozian M, Neal C, Carson PL, Lu Y, Hadjiiski L, Wei J. Digital breast tomosynthesis: Observer performance study of the detection of clustered microcalcifications in breast phantom images acquired with an advanced experimental DBT system Accepted for publication in. Radiology pending revision (November 2013) 2014 doi: 10.1148/radiol.14132722. [DOI] [PMC free article] [PubMed] [Google Scholar]
  6. Chawla AS, Lo JY, Baker JA, Samei E. Optimized image acquisition for breast tomosynthesis in projection and reconstruction space. Medical physics. 2009;36:4859–69. doi: 10.1118/1.3231814. [DOI] [PMC free article] [PubMed] [Google Scholar]
  7. Chawla AS, Samei E, Saunders RS, Lo JY, Baker JA. A mathematical model platform for optimizing a multiprojection breast imaging system. Medical physics. 2008;35:1337–45. doi: 10.1118/1.2885367. [DOI] [PMC free article] [PubMed] [Google Scholar]
  8. Ciatto S, Houssami N, Bernardi D, Caumo F, Pellegrini M, Brunelli S, Tuttobene P, Bricolo P, Fanto C, Valentini M, Montemezzi S, Macaskill P. Integration of 3D digital mammography with tomosynthesis for population breast-cancer screening (STORM): a prospective comparison study. The lancet oncology. 2013;14:583–9. doi: 10.1016/S1470-2045(13)70134-7. [DOI] [PubMed] [Google Scholar]
  9. Das M, Gifford HC, O’Connor JM, Glick SJ. Evaluation of a variable dose acquisition technique for microcalcification and mass detection in digital breast tomosynthesis. Medical physics. 2009;36:1976–84. doi: 10.1118/1.3116902. [DOI] [PMC free article] [PubMed] [Google Scholar]
  10. Eberhard JW, Staudinger P, Smolenski J, Ding J, Schmitz A, McCoy J, Rumsey M, Al-Khalidy A, Ross W, Landberg CE, Claus BEH, Carson P, Goodsitt M, Chan H-P, Roubidoux M, Thomas JA, Osland J. High speed, large angle mammography tomosynthesis system - art. no. 61420C. In: Flynn MJ, Hsieh J, editors. Medical Imaging 2006: Physics of Medical Imaging, Pts 1–3. 2006. pp. C1420–C. [Google Scholar]
  11. Freed M, Park S, Badano A. A fast, angle-dependent, analytical model of CsI detector response for optimization of 3D x-ray breast imaging systems. Medical physics. 2010;37:2593–605. doi: 10.1118/1.3397462. [DOI] [PMC free article] [PubMed] [Google Scholar]
  12. Gang GJ, Tward DJ, Lee J, Siewerdsen JH. Anatomical background and generalized detectability in tomosynthesis and cone-beam CT. Medical physics. 2010;37:1948–65. doi: 10.1118/1.3352586. [DOI] [PMC free article] [PubMed] [Google Scholar]
  13. Gifford HC, Didier CS, Das M, Glick SJ. Optimizing breast-tomosynthesis acquisition parameters with scanning model observers - art. no. 69170S. In: Sahiner B, Manning DJ, editors. Medical Imaging 2008: Image Perception, Observer Performance, and Technology Assessment. 2008. pp. S9170–S. [Google Scholar]
  14. Hamers S, Freyschmidt J, Neitzel U. Digital radiography with a large-scale electronic flat-panel detector vs screen-film radiography: observer preference in clinical skeletal diagnostics. European Radiology. 2001;11:1753–9. doi: 10.1007/s003300100830. [DOI] [PubMed] [Google Scholar]
  15. Hu YH, Zhao B, Zhao W. Image artifacts in digital breast tomosynthesis: investigation of the effects of system geometry and reconstruction parameters using a linear system approach. Medical physics. 2008;35:5242–52. doi: 10.1118/1.2996110. [DOI] [PMC free article] [PubMed] [Google Scholar]
  16. Lu Y, Chan HP, Wei J, Goodsitt M, Carson PL, Hadjiiski L, Schmitz A, Eberhard JW, Claus BE. Image quality of microcalcifications in digital breast tomosynthesis: effects of projection-view distributions. Medical physics. 2011;38:5703–12. doi: 10.1118/1.3637492. [DOI] [PMC free article] [PubMed] [Google Scholar]
  17. Maidment A, Albert M, Thunberg S, Adelow L, Blom O, Egerstrom J, Eklund M, Francke T, Jordung U, Kristoffersson T, Lindman K, Lindqvist L, Marchal D, Olla H, Penton E, Rantanen J, Solokov S, Ullberg C, Weber N. Evaluation of a photon-counting breast tomosynthesis imaging system. In: Flynn MJ, editor. Medical Imaging 2005: Physics of Medical Imaging, Pts 1 and 2. 2005. pp. 572–82. [Google Scholar]
  18. Maidment ADA, Ullberg C, Lindman K, Adelow L, Egerstrom J, Eklund M, Francke T, Jordung U, Kristoffersson T, Lindqvist L, Marchal D, Olla H, Penton E, Rantanen J, Solokov S, Weber N, Westerberg H. Evaluation of a photon-counting breast tomosynthesis imaging system - art. no. 61420B. In: Flynn MJ, Hsieh J, editors. Medical Imaging 2006: Physics of Medical Imaging, Pts 1–3. 2006. pp. B1420–B. [Google Scholar]
  19. Mainprize JG, Bloomquist AK, Kempston MR, Yaffe MJ. Resolution at oblique incidence angles of a flat panel imager for breast tomosynthesis. Medical physics. 2006;33:3159–64. doi: 10.1118/1.2241994. [DOI] [PubMed] [Google Scholar]
  20. Mertelmeier T, Ludwig J, Zhao B, Zhao W. Optimization of tomosynthesis acquisition parameters: Angular range and number of projections. In: Krupinski EA, editor. Digital Mammography, Proceedings. 2008. pp. 220–7. [Google Scholar]
  21. Muramatsu C, Li Q, Schmidt RA, Shiraishi J, Suzuki K, Newstead GM, Doi K. Determination of subjective similarity for pairs of masses and pairs of clustered microcalcifications on mammograms: Comparison of similarity ranking scores and absolute similarity ratings. Medical physics. 2007;34:2890–5. doi: 10.1118/1.2745937. [DOI] [PubMed] [Google Scholar]
  22. Nishikawa RM, Reiser I, Seifi P. A new approach to digital breast tomosynthesis for breast cancer screening. In: Hsieh J, Flynn MJ, editors. Medical Imaging 2007: Physics of Medical Imaging, Pts 1–3. 2007. pp. U1464–U71. [Google Scholar]
  23. Rafferty EA, Park JM, Philpotts LE, Poplack SP, Sumkin JH, Halpern EF, Niklason LT. Assessing radiologist performance using combined digital mammography and breast tomosynthesis compared with digital mammography alone: results of a multicenter, multireader trial. Radiology. 2013;266:104–13. doi: 10.1148/radiol.12120674. [DOI] [PMC free article] [PubMed] [Google Scholar]
  24. Reiser I, Nishikawa RM. Task-based assessment of breast tomosynthesis: effect of acquisition parameters and quantum noise. Medical physics. 2010;37:1591–600. doi: 10.1118/1.3357288. [DOI] [PMC free article] [PubMed] [Google Scholar]
  25. Reiser IS, Nishikawa RM, Lau BA. Effect of non-isotropic detector blur on microcalcification detectability in tomosynthesis. Proc. SPIE 7258, Medical Imaging; 2009; 2009. p. 72585Z-Z-10. [Google Scholar]
  26. Ren B, Ruth C, Zhang Y, Smith A, Williams C, Polischuk B, Jing Z. The CNR method in scan angle optimization of tomosynthesis and its limitations. In: Hsieh ESJ, editor. Medical Imaging 2009: Physics of Medical Imaging. The international Society for Optical Engineering; 2009. p. 72585W-W-10. [Google Scholar]
  27. Ren B, Wu T, Smith A, Ruth C, Niklason L, Jing Z, Stein J. The dependence of tomosynthesis imaging performance on the number of scan projections. In: Astley SM, et al., editors. Digital Mamography, Proceedings. 2006. pp. 517–24. [Google Scholar]
  28. Rose SL, Tidwell AL, Bujnoch LJ, Kushwaha AC, Nordmann AS, Sexton R., Jr Implementation of breast tomosynthesis in a routine screening practice: an observational study. AJR American journal of roentgenology. 2013;200:1401–8. doi: 10.2214/AJR.12.9672. [DOI] [PubMed] [Google Scholar]
  29. Schneider CA, Rasband WS, Eliceiri KW. NIH Image to ImageJ: 25 years of image analysis. Nature Methods. 2012;9:671–5. doi: 10.1038/nmeth.2089. [DOI] [PMC free article] [PubMed] [Google Scholar]
  30. Sechopoulos I. A review of breast tomosynthesis. Part I. The image acquisition process. Medical physics. 2013;(40):014301. doi: 10.1118/1.4770279. [DOI] [PMC free article] [PubMed] [Google Scholar]
  31. Sechopoulos I, Ghetti C. Optimization of the acquisition geometry in digital tomosynthesis of the breast. Medical physics. 2009;36:1199–207. doi: 10.1118/1.3090889. [DOI] [PubMed] [Google Scholar]
  32. Sechopoulos I, Suryanarayanan S, Vedantham S, D’Orsi C, Karellas A. Computation of the glandular radiation dose in digital tomosynthesis of the breast. Medical physics. 2007;34:221–32. doi: 10.1118/1.2400836. [DOI] [PMC free article] [PubMed] [Google Scholar]
  33. Shkumat NA, Siewerdsen JH, Richard S, Paul NS, Yorkston J, Van Metter R. Dual-energy imaging of the chest: Optimization of image acquisition techniques for the ‘bone-only’ image. Medical physics. 2008;35:629–32. doi: 10.1118/1.2828186. [DOI] [PubMed] [Google Scholar]
  34. Skaane P, Bandos AI, Gullien R, Eben EB, Ekseth U, Haakenaasen U, Izadi M, Jebsen IN, Jahr G, Krager M, Niklason LT, Hofvind S, Gur D. Comparison of digital mammography alone and digital mammography plus tomosynthesis in a population-based screening program. Radiology. 2013;267:47–56. doi: 10.1148/radiol.12121373. [DOI] [PubMed] [Google Scholar]
  35. Sobol WT, Wu XZ. Parametrization of mammography normalized average glandular dose tables. Medical physics. 1997;24:547–54. doi: 10.1118/1.597937. [DOI] [PubMed] [Google Scholar]
  36. Tourassi G, Yoon H-J, Xu S, Morin-Ducote G, Hudson K. Comparative Analysis of Data Collection Methods for Individualized Modeling of Radiologists’ Visual Similarity Judgments in Mammograms. Academic Radiology. 2013;20:1371–80. doi: 10.1016/j.acra.2013.08.002. [DOI] [PubMed] [Google Scholar]
  37. Tucker A, Qian X, Gidcumb E, Spronk D, Sprenger F, Kuo J, Ng S, Lu J, Zhou O. Optimizing configuration parameters of a stationary digital breast tomosynthesis system based on carbon nanotube X-ray sources. In: Pelc NJ, et al., editors. Medical Imaging 2012: Physics of Medical Imaging; Proc. SPIE 8313, Medical Imaging 2012: Physics of Medical Imaging; 2012. p. 831307. [Google Scholar]
  38. Tucker AW, Lu J, Zhou O. Dependency of image quality on system configuration parameters in a stationary digital breast tomosynthesis system. Medical physics. 2013;40:031917. doi: 10.1118/1.4792296. [DOI] [PMC free article] [PubMed] [Google Scholar]
  39. Van de Sompel D, Brady SM, Boone J. Task-based performance analysis of FBP, SART and ML for digital breast tomosynthesis using signal CNR and Channelised Hotelling Observers. Medical image analysis. 2011;15:53–70. doi: 10.1016/j.media.2010.07.004. [DOI] [PubMed] [Google Scholar]
  40. Williams DB, Siewerdsen JH, Tward DJ, Paul NS, Dhanantwari AC, Shkumat NA, Richard S, Yorkston J, Van Metter R. Optimal kVp selection for dual-energy imaging of the chest: Evaluation by task-specific observer preference tests. Medical physics. 2007;34:3916–25. doi: 10.1118/1.2776239. [DOI] [PubMed] [Google Scholar]
  41. Wu T, Moore RH, Rafferty EA, Kopans DB. A comparison of reconstruction algorithms for breast tomosynthesis. Medical physics. 2004;31:2636–47. doi: 10.1118/1.1786692. [DOI] [PubMed] [Google Scholar]
  42. Wu XZ, Gingold EL, Barnes GT, Tucker DM. NORMALIZED AVERAGE GLANDULAR DOSE IN MOLYBDENUM TARGET RHODIUM FILTER AND RHODIUM TARGET RHODIUM FILTER MAMMOGRAPHY. Radiology. 1994;193:83–9. doi: 10.1148/radiology.193.1.8090926. [DOI] [PubMed] [Google Scholar]
  43. Young S, Bakic PR, Myers KJ, Jennings RJ, Park S. A virtual trial framework for quantifying the detectability of masses in breast tomosynthesis projection data. Medical physics. 2013;40:051914-1-15. doi: 10.1118/1.4800501. [DOI] [PMC free article] [PubMed] [Google Scholar]
  44. Zhang Y, Chan H-P, Sahiner B, Wei J, Goodsitt MM, Hadjiiski LM, Ge J, Zhou C. A comparative study of limited-angle cone-beam reconstruction methods for breast tomosynthesis. Medical physics. 2006;33:3781–95. doi: 10.1118/1.223754. [DOI] [PMC free article] [PubMed] [Google Scholar]
  45. Zhao B, Zhao W. Imaging performance of an amorphous selenium digital mammography detector in a breast tomosynthesis system. Medical physics. 2008;35:1978–87. doi: 10.1118/1.2903425. [DOI] [PMC free article] [PubMed] [Google Scholar]

RESOURCES