Evaluation of A Photon-Counting Breast Tomosynthesis Imaging System
Evaluation of A Photon-Counting Breast Tomosynthesis Imaging System
ABSTRACT
Digital breast tomosynthesis promises solutions to many of the problems associated with projection mammography,
including elimination of artifactual densities due to the superposition of normal tissues and increasing the conspicuity of
true lesions that would otherwise be masked by superimposed normal tissue. We have investigated tomosynthesis using
a digital camera containing 48 photon counting, orientation sensitive, linear detectors which are precisely aligned with
the focal spot of the x-ray source. The x-ray source and the digital detectors are scanned in a continuous motion across
the object (patient), each linear detector collecting an image at a distinct angle. A preliminary assessment of
tomosynthesis image quality has been performed with both qualitative and quantitative methods. Measured values of
MTF and NPS appear concordant with theoretical values. The MTF in the scanning direction is dominated by scanning
unsharpness and geometric factors, while the NPS is white. The MTF and NPS in the strip direction are somewhat lower
than in the scan direction. The NPS of tomographic images show a slight decrease with increasing spatial frequency,
related to the sampling and interpolation in the reconstruction process. A phase I clinical trial is ongoing; 9 women have
been recruited. Breast positioning is comparable to other imaging systems. The visualization of breast anatomy appears
to be superior to screen-film mammography, at the same average glandular dose. Examination of images reconstructed
with a sub-sampled set of projection images appears to support the hypothesis that image quality is superior when more
projection images are used in the reconstruction.
Keywords: Digital breast tomosynthesis, synthetic tomography, digital mammography, modulation transfer function
(MTF), noise power spectra (NPS), clinical trial.
1. INTRODUCTION
There are two key problems currently associated with projection mammography.1 First, artifactual densities can arise
from the superposition of normal tissues; such densities can result in a loss of specificity. Second, true lesions often lack
conspicuity or are morphologically ill-defined as they are masked by superimposed normal tissue, potentially reducing
both sensitivity and specificity. Tomosynthesis is a promising solution to overcome these problems, providing
tomographic images of the breast while also providing a simple means of localizing lesions in three-dimensions.2-7
A novel tomosynthesis system (XC Mammo-3T, XCounter AB, Danderyd, Sweden) has been developed for
mammography screening and diagnosis. The system differs fundamentally from other breast tomosynthesis systems in
that it simultaneously acquires a large number of images free of electronic noise and scattered x-rays, using a novel
scanning method. Current competing tomosynthesis implementations use a mammographic area detector to repeatedly
acquire images in which the projection angle is slowly altered either continuously or discretely. Such designs naturally
suffer problems which can potentially degrade tomographic image quality. Specifically, the requirement of repeated
exposures from an individual detector requires a trade-off between scan time and the number of acquisition angles. As
demonstrated previously, image artifacts are better suppressed when more projection images are used to produce the
tomographic reconstructions.7 In addition, current area detectors are energy-integrating, and thus suffer from various
sources of electronic noise, as well as ghosting and lag.8 These imperfections in the detector serve to further degrade the
resulting tomographic image quality.
*
[email protected]; Phone + 1 215 746-8763; Fax +1 215 746-8764
Digital breast tomosynthesis is a new and developing field; the results presented here are part of an ongoing work-in-
progress. We have sought to both qualitatively and quantitatively assess system technical and clinical performance, and
to optimize system design. In this paper, we report preliminary measures of imaging performance and a phase I clinical
trial. Continuing the research presented last year, we again examine the role of the number of projection angles on the
image quality of the reconstructed tomosynthesis images.
The system consists of an E-arm that is scanned across the breast, as shown in Figure 1. The E-arm supports the x-ray
tube, pre-patient collimator and camera. The x-ray source is a tungsten-anode mammography tube (Varian, RAD 70T).
The x-ray flux is filtered through a 0.5 mm aluminum filter. The x-ray field is split into 48 fan-shaped beams by a
tungsten pre-patient collimator positioned just above the breast. The camera, consisting of 48 sensor rows enclosed in a
steel gas-tight box, moves under the breast. The camera box is filled with a noble gas at a few atmospheres pressure.
The linear sensors occupy an active area of 24×30 cm2. The pixel size is 60 µm and each sensor is read out every 60 µm
during the scan.
Three dimensional tomographic images are reconstructed using a filtered back-projection method, in slices parallel with
the detector spaced every 1 mm. Both the tomographic and source images are displayed using custom software. The
software presents the images in stack mode. The reader can magnify, pan, and window and level the images.
The individual gaseous avalanche x-ray sensors operate on the principle of photon counting. The x rays interact in the
gas and liberate charges producing an image of the x-ray flux transmitted through the patient. The detector is shown in
schematic in Figure 2. The detector consists of 2 electrodes: one continuous; the other forming pixels aligned with the
divergent x-ray beam. The x rays enter one-by-one into the gas between the electrodes and ionize the gas, liberating
electrons. A voltage applied between the electrodes causes these electrons to move towards the positive anode. These
electrons further ionize the gas, knocking out new electrons and triggering an avalanche of electrons. Hence, the energy
from each x ray is amplified many-fold. The gain of the gaseous avalanche is such that only x rays (or gamma rays) are
sufficient to produce an avalanche, and a simple threshold can be used to discriminate the x rays. As a result, there is no
dark current, and the image noise is limited by the Poisson statistical fluctuation of the incoming x-ray flux.
As discussed previously,5-7 the system is insensitive to scattered x rays; only x rays traveling in a straight line from the
focal spot of the x-ray source can enter between the electrodes and liberate sufficient charge to trigger an avalanche.
The system is typically operated between 30 and 40 kVp, at up to 210 mA. The kVp, mA and integration time (scan
speed) can be altered to achieve a specific breast or detector dose. Forty-eight source images appropriate for
tomosynthesis reconstruction can be acquired over a region of 24×30 cm2 within 10-20 seconds. Optimization as a
function of breast size and thickness is ongoing. The techniques used in the current clinical trial are designed to use an
average glandular dose (AGD) less than or equal to screen-film mammography. It is our belief that the absence of
electronic noise and scatter facilitate this choice. As discussed in Section 3.2, it appears that this choice results in
clinically acceptable image quality.
The noise power spectra (NPS) were measured using the same spectrum (applied voltage and filtration) as for the MTF
measurement, with a dose corresponding to an AGD of 1.5 mGy for a 40 mm breast. Ten planes nearest the bottom edge
of the 40 mm PMMA phantom (vertically) were reconstructed with 1 mm spacing using backprojection without a
frequency-dependent filter. A 1024×2048 pixel region of interest (ROI) close to the chestwall was selected. The 2D
NPS was then calculated from 256×256 pixel ROI overlapped by 128 pixels in each direction, by averaged over all sub
regions and all reconstructed planes. The 1D NPS were then extracted from the 2D spectra. Using the method of Bendat
and Piersol,13 we calculate that the standard error in the NPS estimates is 3.2%
2.3. Clinical Trial Design
A phase I clinical trial involving 30 women is ongoing. The trial is being conducted jointly at Danderyds Sjukhus
AB (Danderyds, Sweden) and XCounter AB. At the time of the conference (Feb. 10, 2006), 9 women had participated
in the imaging trial. Participants of the trials must be at least 40 years of age, and pregnant women are excluded from
the trial. All diagnostic patients at Danderyds Sjukhus are eligible; these specifically include women recalled from
screening, and women referred for clinical reasons. Screen-film mammography is performed at the Danderyds Sjukhus
Mammography Department per clinical practice. Images are acquired on one of two mammography systems: a Siemens
Mammomat 3000 (Siemens, Erlangen, Germany) or a Planmed Sophie (Planmed Oy, Helsinki, Finland). Both systems
have a molybdenum target and filter. Films are typically acquired at 30-31 kVp, and are phototimed.
The tomosynthesis images are acquired at XCounter on the same day by the same radiologic nurse. Only one breast is
imaged in a single view. Due to initial mechanical constraints, the first 6 patients were imaged in a CC orientation. The
most recent three patients were imaged in an MLO orientation. Currently, all patients have been imaged at 35 kVp, 140-
180 mA with a 2 ms integration time, resulting in a scan time of 17 seconds. Note that any one part of the breast is only
in the x-ray beam for 10 seconds, thus minimizing the risk of motion unsharpness. This technique was chosen to result
in an AGD of approximately 1.5 mGy. Tomosynthesis images are being reviewed prospectively by Dr. Per Sunden
(Danderyds Sjukhus AB). A retrospective trial of image quality is planned, but is beyond the scope of this paper.
Measured MTF
0.2
0
0 1 2 3 4 5 6 7 8
-1
Spatial Frequency (mm )
1
Figure 4: MTF of tomographic
images in the scan and strip
directions. The MTF in the strip
0.8 direction is reduced compared to the
scan direction due to simultaneous
Scan Direction triggering of adjacent channels.
0.6
MTF
0.4
Strip Direction
0.2
0
0 1 2 3 4 5 6 7 8
Figure 4 shows the measured MTF in the scan and strip directions. The resolution in the strip direction is lower than that
in the scan direction. This degradation is still under investigation; however, it is likely due to simultaneous triggering of
adjacent channels.
The measured NPS are shown in Figure 5. The NPS have been normalized to unity at their maximum, except the NPS
for the projection images in the scanning direction (solid squares) were normalized by their mean. The NPS in the
scanning direction are shown as solid symbols, while the NPS in the strip direction are shown as open symbols. The
NPS of the projection images (NPSP) are shown as rectangles. The NPS in the scanning direction (solid squares) is
white. This is expected, as each line is acquired independently (sequentially in time), and the system is free of lag; thus,
there should be no correlation between neighboring lines in the image. By comparison, the NPS in the strip direction
(open squares) shows some correlation, which manifests itself as a reduction at high frequencies. It is likely that the
correlation seen in the NPS is due to the same phenomenon responsible for a reduction in the MTF in the strip direction.
0.2
0
0 1 2 3 4 5 6 7 8
-1
Spatial Frequency (mm )
The NPS of the reconstructed tomographic images (NPST) are shown as diamonds. The NPST is reduced in both the
scanning and strip directions. To understand the source of this correlation, we took the ratio of the NPST/NPSP. As
expected, the ratio in the scanning direction (solid triangles) is indistinguishable from NPST in that direction. The ratio
in the strip direction (open triangles) has a similar shape to that in the scanning direction. We believe that both sources
of correlation have the same origins; namely, the correlation arises from interpolation of the projection data in the back-
projection operation. We are currently investigating this further.
It should be noted that both the NPS and the MTF were measured with a 40 mm PMMA phantom in the same position as
the breast. As a result, the presented data should be interpreted as the “system MTF” and “system NPS”.10, 12 There is a
complete absence of a low-frequency drop in either the MTF or the NPS. This is due to the almost complete elimination
of scattered radiation. This differs from area systems, where scatter can account for up to half of the x-ray flux incident
on the detector in the absence of a grid. This benefit will also be present in the DQE (not presented here), as the scatter
elimination is achieved without attenuating the primary radiation.
The images to date have shown very high spatial resolution. In general, we see more calcifications in the tomosynthesis
images than in the screen-film mammograms. Further, the calcifications in the tomosynthesis images are generally
better resolved (sharper and having higher contrast) than in the screen-film images. We find that calcifications rapidly
disappear when out-of-plane. These observations are consistent with our previous findings with phantoms and animals,
and are likely due to the choice of angular range, number of projection images and pixel size.7
Figure 6: The imaging system is shown (left) with a model, demonstrating the positioning used for a MLO tomosynthesis image. A
MLO image is shown in the center. The pectoralis muscle extends to below the perpendicular line from the muscle to the nipple
(arrow). A CC image is shown on the right. This image shows the posterior margin of the glandular tissue. These two images clearly
demonstrate that good clinical breast positioning is possible with the system.
Large low-contrast objects have also been well visualized. The images (see Figures 6-8) depict the breast anatomy well.
The glandular tissue, adipose tissue, Cooper’s ligaments, blood vessels, lymph nodes and other structures of the breast
are well depicted. To date, we have found one cancer (Figure 7), which was quite obvious in the tomosynthesis image,
and only marginally visible in the screen-film image. While anecdotal, we believe that these early images provide
convincing evidence of the superiority of both tomosynthesis and our approach of simultaneously acquiring multiple
images with a scanning photon-counting detector. We believe that the system is capable of clinical quality images, with
adequate tissue penetration and breast positioning. Admittedly, these results are preliminary and lack statistical
significance.
These initial images have been acquired at an AGD that is comparable to the screen-film examination of a 4 cm thick
breast. The technique for the screen-film mammograms ranged from 30 to 31 kVp and from 21 to 166 mAs, resulting in
an entrance skin air kerma (ESAK) of between 2.91 and 21.75 mGy (mean value of 7.31 mGy) and an AGD of between
0.91 and 3.50 mGy (mean value of 1.56 mGy). The digital tomosynthesis images were acquired with a technique
35 kVp, 140 to 180 mAs. This resulted in an ESAK of between 4.33 and 5.61 mGy (mean value of 4.83 mGy) and an
AGD of between 1.11 and 1.67 mGy (mean value of 1.45 mGy). In should be noted that the first 6 patients were imaged
with 40 active linear detectors, while the last 3 patients were imaged with the full set of 48 linear detectors; as such, the
AGD for the first 6 patients (1.42 mGy) is lower than for the last 3 patients (1.50 mGy).
As with any study comparing dose, it is important to realize that the dose in a digital image is somewhat arbitrary, as the
system is linear and has very wide dynamic range. The relevant questions are: (1) is the resultant image x-ray quantum
noise limited to high spatial-frequency; and (2) are the images of clinical quality. We believe that the data presented in
Section 3.1 are proof of the former. We further believe that the results in this section provide anecdotal proof of the
latter, for it is notable that the tomosynthesis images were acquired at a lower dose than the screen-film mammograms,
yet appear to have comparable or superior image quality.
The difference between images (a) and (b) is shown in image (c). The difference images have been offset so that the
average value is in the middle of the grayscale range. Here, the main differences occur on the periphery of the breast,
and along other structures with edges. The difference between images (a) and (d) is shown in image (e). In (e) there are
significant differences seen throughout the breast, including structures within the tumor. These structures again appear
to relate mostly to edge information, but do appear to include some grayscale (i.e., lower spatial frequency) information.
Again, it should be noted that these results are anecdotal. However, after examining many such images, we believe that
it is essential that tomographic images of the breast be acquired with as many unique projection angles as possible. We
further believe that these clinical examples support our prior research.7 A more complete reader study will be necessary
to test whether these differences result in statistically significant differences in clinical performance.
a) Figure 9: Tomographic images of a tumor are
shown. These images were reconstructed with (a) 48
projection angles, (b) 24 projection angles, and (d)
12 projection angles. The images were reconstructed
over the same range of projection angles. The
images were processed so as to have comparable
quantum noise. The difference between (a) and (b)
is shown in (c). The difference images have been
offset so that the average value is in the middle of
the grayscale range. In (c), the main differences
occur on the periphery of the breast, and along other
edges. The difference between (a) and (d) is shown
in (e). In (e) there are significant differences seen
throughout the breast, including structures within the
tumor.
b) c)
d) e)
4. CONCLUSION
A clinical photon-counting tomosynthesis imaging system has been developed and is in clinical trials. A preliminary
assessment of tomosynthesis image quality has been performed with both qualitative and quantitative methods. The
measured values of MTF and NPS appear concordant with theoretical values. The MTF in the scanning direction is
dominated by scanning unsharpness and geometric factors. The NPS in the same direction demonstrates that the
acquired data are uncorrelated between acquisition lines (i.e., separated in time). This would indicate that the system is
free of lag. The MTF and NPS in the strip direction are somewhat lower than in the scan direction; this is believed to be
related to simultaneous triggering of adjacent channels in the detector array. The reconstructed NPS show a slight
decrease with increasing spatial frequency. We believe that this is related to the sampling and interpolation of rays in the
reconstruction process.
A phase I clinical trial is ongoing. To date, 9 of 30 women have been recruited. The clinical image quality has been
assessed qualitatively. Breast positioning appears to be similar to other digital and screen-film imaging systems. The
visualization of breast anatomy (both normal and malignant) appears to be comparable or superior to screen-film
mammography, at an AGD comparable to screen-film mammography. Examination of images reconstructed with a sub-
sampled set of tomographic images appears to support the hypothesis that image quality is superior when more
projection images are used in the reconstruction. These results are anecdotal; a larger reader trial is necessary to
demonstrate whether these results are clinically significant.
ACKNOWLEDGEMENTS
The authors would like to acknowledge the assistance of Dr. Per Sunden of Danderyds Sjukhus AB (Stockholm) for his
assistance in the clinical trial. This work is partially funded by NIH grant PO1 CA85484.
REFERENCES
1. L. Ma, E. Fishell, B. Wright, W. Hanna, S. Allan, and N. F. Boyd, "Case-control study of factors associated
with failure to detect breast cancer by mammography.," Journal of the National Cancer Institute 84(10), 781-
785 (1992).
2. L. T. Niklason, B. T. Christian, L. E. Niklason, D. B. Kopans, D. E. Castleberry, B. H. Opsahl-Ong, C. E.
Landberg, P. J. Slanetz, A. A. Giardino, R. Moore, D. Albagli, M. C. DeJule, P. F. Fitzgerald, D. F. Fobare, B.
W. Giambattista, R. F. Kwasnick, J. Liu, S. J. Lubowski, G. E. Possin, J. F. Richotte, C. Y. Wei, and R. F.
Wirth, "Digital tomosynthesis in breast imaging," Radiology 205(2), 399-406 (1997).
3. J. T. Dobbins, 3rd and D. J. Godfrey, "Digital x-ray tomosynthesis: current state of the art and clinical
potential," Physics in Medicine & Biology 48(19), R65-106 (2003).
4. S. Thunberg and e. al, "Evaluation of a Photon Counting Mammography System," presented at the SPIE
Medical Imaging 2002, 4682, 202-208 (2002).
5. S. Thunberg, A. D. A. Maidment, and e. al., "Dose reduction in Mammography with Photon Counting
Imaging," presented at the SPIE Medical Imaging 2004, 5368, 457-465 (2004).
6. S. Thunberg, A. D. A. Maidment, and e. al., "Tomosynthesis with a Multi-line Photon Counting Camera,"
presented at the 7th International Workshop on Digital Mammography, 459-465 (2004).
7. A. D. A. Maidment, M. Albert, S. Thunberg, L. AdelÖw, O. Blom, J. EgerstrÖm, M. Eklund, T. Francke, U.
Jordung, T. Kristoffersson, K. Lindman, L. Lindqvist, D. Marchal, H. Olla, E. Penton, J. Rantanen, S. Solokov,
C. Ullberg, and N. Weber, "Evaluation of a Photon-Counting Breast Tomosynthesis Imaging System,"
presented at the SPIE Medical Imaging 2005, 5745, 572-582 (2005).
8. J. H. Siewerdsen and D. A. Jaffray, "A ghost story: spatio-temporal response characteristics of an indirect-
detection flat-panel imager," Medical Physics 26(8), 1624-1641 (1999).
9. Y.-H. Kao, A. D. A. Maidment, M. Albert, A.-K. Carton, and H. Bosmans, "Assessment of a software tool for
measurement of the modulation transfer function," presented at the Medical Imaging 2005: Physics of Medical
Imaging, 5745, 1199-1208 (2005).
10. A. K. Carton, D. Vandenbroucke, L. Struye, A. D. Maidment, Y. H. Kao, M. Albert, H. Bosmans, and G.
Marchal, "Validation of MTF measurement for digital mammography quality control," Medical Physics 32(6),
1684-1695 (2005).
11. A. D. A. Maidment and M. Albert, "Conditioning data for calculation of the modulation transfer function,"
Medical Physics 30(2), 248-253 (2003).
12. Z. Shen, G. Mawdsley, A. Bloomquist, J. Mainprize, and M. Yaffe, "Interpreting system MTF and NPS
measured on clinical digital mammography systems," presented at the Digital Mammography IWDM 2002,
123–127 (2002).
13. J. S. Bendat and A. G. Piersol, Random Data: Analysis and Measurement Procedures, 2nd Edition ed. (Wiley-
Interscience, New York, 1986).