Content uploaded by Alistair Mackenzie
Author content
All content in this area was uploaded by Alistair Mackenzie on Sep 28, 2016
Content may be subject to copyright.
A.D.A. Maidment, P.R. Bakic, and S. Gavenonis (Eds.): IWDM 2012, LNCS 7361, pp. 394–401, 2012.
© Springer-Verlag Berlin Heidelberg 2012
Converting One Set of Mammograms to Simulate
a Range of Detector Imaging Characteristics
for Observer Studies
Alistair Mackenzie1, David R. Dance1,
Oliver Diaz2, Annabel Barnard1, and Kenneth C. Young1
1 National Coordinating Centre for the Physics of Mammography, Royal Surrey County
Hospital, Guildford, GU2 7XX, UK and Department of Physics, University of Surrey,
Guildford, GU2 7XH, UK
2 Centre for Vision, Speech and Signal Processing, Faculty of Engineering and Physical
Sciences, University of Surrey, Guildford, GU2 7XH, UK
alistairmackenzie@nhs.net
Abstract. A methodology for adjusting mammographic images taken on a
given imaging system to simulate their appearance if taken on a different
system for use in observer studies is presented. The process involves adjusting
the image sharpness and noise, which takes into account the detector, breast
thickness, and beam quality. The method has been tested by converting images
acquired using an a-Se detector of a CDMAM test object and ‘Rachel’
anthropomorphic breast phantom. They were degraded to appear as if acquired
using a computed radiography (CR) detector. Good agreement was achieved in
the resulting threshold gold thickness for the simulated CR images with
measured real values for CDMAM images. Power spectra comparisons of real
and simulated images of the ‘Rachel’ phantom agree with an average difference
of 4%. This tool in conjunction with observer studies can be used to understand
the effects of the detector characteristics on cancer detection in mammography.
Keywords: simulation, noise power spectra, modulation transfer function.
1 Introduction
Clinical evaluation of image quality is expensive and time-consuming. Clinical trials
to compare the effectiveness of different systems are rarely conducted as they would
require large numbers of patients to achieve both sufficient numbers of detected
cancers and statistical significance. In particular it would be desirable to repeat
exposures on the same breast with the same positioning and compression to minimise
confounding differences in the projection of the breast tissues, but this raises ethical
issues. Alternative methods of evaluation involving some degree of image simulation
have the potential to enable comparisons at reduced cost and time and without
additional radiation exposure. For this purpose it is desirable to be able to acquire
images on a given system and to simulate their appearance on a second system, so the
Adjusting Image Quality of Mammograms 395
performance of the two systems can be compared. This may be possible when the
performance of the second system is inferior to that of the first system. Such a method
would enable the background tissue and compression to be matched in different arms
of a study, either using real cancers or the insertion of simulated cancers.
The aim of this work is therefore to develop and test a methodology for adjusting
mammographic images taken on a given imaging system to simulate their appearance
if taken on a different system. The methodology presented extends previous work [1]
with improved modelling of the noise power spectra, which takes into account the
breast thickness and beam quality.
2 Method
2.1 Summary of Methodology for Changing Image Quality
The conversion methodology blurs the original image to match the blurring of a target
system using measurements of the modulation transfer function. The difference in
noise between the original system and the target system is then calculated and added
in real space to the blurred image, ensuring that the magnitude and correlation of the
noise matches the total noise in the target system. The method also accounts for the
magnitude of the signal in the image. It has been validated using images of a contrast
detail test object for situations where the noise characteristics were measured at the
same beam quality as the image to be converted.
2.2 Linearisation of Images
The analysis below assumes that images have been linearised so that the pixel value is
a measure of the energy absorbed per unit area of the detector. This can be achieved
by a combination of measurements of the incident air kerma at the front face of the
detector and Monte Carlo simulations. The Geant4 Monte Carlo code
(http://geant4.cern.ch/) was used to calculate the absorbed energy per unit area for a
reference beam quality, and to relate this to the incident air kerma at the front face of
the image receptor. X-ray spectra from the work of Boone et al [2] filtered by the X-
ray tube window, filter, compression paddle, object being imaged (test phantom or
breast) and breast support were used for this purpose. The attenuation coefficients
were obtained from Berger et al [3]. The signal transfer properties (STP) thus
calculated for a given detector were assumed to apply to any beam quality for that
detector so that the pixel value was a measure of the energy absorbed per unit area for
all beam qualities.
2.3 Characterisation of Noise Power Spectra (NPS)
The noise was characterised by the NPS (W) which was split into components for
electronic, quantum and structure noise at a reference beam quality. For this purpose,
a series of collimated flat field images were acquired over a wide dose range using a
396 A. Mackenzie et al.
28 kV, Mo/Mo anode/filter combination and 4.5 cm polymethyl methacrylate
(PMMA) at the tube head and the NPS was calculated for each dose. The three noise
components for the reference condition were estimated by fitting a second order
polynomial of NPS against absorbed energy for each spatial frequency.
2.4 Correction of Noise and Signal for Beam Quality
Clinical mammograms are acquired over a range of compressed breast thicknesses
and radiographic factors and so the model needs to be able to take account of a range
of beam qualities. When originally developed the methodology used measurements of
the NPS appropriate to the beam quality used. In the present work we have improved
the modelling of the NPS to take account of beam quality and breast
thickness/composition. For this purpose, it was assumed that the three noise sources
are affected by beam quality as follows:
• Electronic noise is independent of beam quality. No correction is required.
• The quantum noise is comprised of a number of different sources (primary
quantum noise, excess noise, and secondary quantum noise) and is dependent
on the number of photons detected and energy absorbed.
• Structure noise is proportional to the signal from the detector. No correction
for beam quality is required.
The beam quality affects both the proportion of energy absorbed and the quantum
noise. To estimate the effect of beam quality on the quantum noise, a series of flat
field images were taken as described in section 2.3, but using a range of PMMA
thicknesses, tube voltages and anode/filter combinations. These images were then
linearised using the reference STP, so that the linearised pixel value equalled the
absorbed energy per unit area irrespective of the beam quality.
Using the above flat field images, the NPS was calculated for each beam quality,
and the results used to determine the parameters in a model of the NPS applicable to
any beam quality (Eq. 1).
2
),(),()(),(),(
++=
o
s
o
qe E
E
vu
E
E
vuBvuvuW
ωωλω
(1)
In this equation ωe, ωq and ωs (with units of mm2) are ‘noise coefficients’ for
electronic, quantum and structure noise respectively at absorbed energy per unit area
at a reference beam quality Eo, E is the absorbed energy per unit area at the beam
quality under consideration and u and v are spatial frequencies. In accordance with
the assumptions in the bullet points above, the electronic and structure noise
coefficients are independent of beam quality. The quantum noise coefficient, ωq, is
also independent of beam quality; the variation of the quantum noise with beam
quality additional to the factor E in Eq. 1 is accounted for by the beam quality
correction factor (B). The beam quality parameter λ is average photon energy of the
beam incident on the detector and was calculated using the X-ray spectra model
Adjusting Image Quality of Mammograms 397
described in section 2.2. For breast images, λ can be calculated if an assumption about
the breast composition is made [4].
2.5 Validation of the Conversion Using Images of CDMAM Test Object
and Anthropomorphic Phantoms
The validation of the conversion from one detector to another was undertaken using
the following two systems:
ASE: Hologic Selenia X-ray system, amorphous selenium (a-Se) detector. Pixel
pitch 70 µm.
CR: Carestream CR900 reader with EHR-M2 CR plates. Pixel pitch 50 µm.
Sixteen images of the CDMAM contrast detail test object were acquired on both
systems at two beam qualities. The CDMAM test object was imaged on the breast
support on a base of 2 cm PMMA (‘thin’) using 26 kV, Mo/Mo and then with an
additional 4 cm PMMA (‘thick’ – total of 6 cm PMMA) on top of the test object and
imaged using 34 kV, Mo/Rh. The image conversion methodology was applied to the
ASE CDMAM images to convert them to appear at the same image quality and dose
as the CR images. The sets of target and simulated CDMAM images for CR were
automatically read using CDCOM software version 1.5.2 (www.euref.org). Contrast
detail curves were produced for both the simulated and target images, which were
then compared [5].
The Rachel anthropomorphic breast phantom (Gammex RMI, WI, USA) was
designed to mimic a 5 cm compressed breast. It was imaged five times using both
detectors at 31 kV, Mo/Rh. The phantom was shifted slightly between images. The
ASE images were converted to appear with the imaging characteristics of the CR
detector and X-ray system used. The largest rectangular region of interest (ROI) away
from the skin edge was extracted from the same location for the real CR and
simulated CR images. Smaller overlapping sub-ROIs of size 256 × 256 were extracted
from this ROI and the measured power spectra from all of the sub-ROIs were
averaged for each system.
2.6 Clinical Images: Subjective Evaluation and Demonstration of Image
Conversion
Mammography images have been collected for an image database. For the systems
included in this study, images from 234 women (ASE) and 233 (CR) were collected.
The ASE had 31 abnormal cases, while all of the CR images were as normal.
Firstly, a preliminary test of realism of the converted images was undertaken. A set
of normal ASE and CR images were selected. A reasonable match between ASE and
CR images in terms of compressed breast thickness, appearance of the breast and
radiographic factors was made. The ASE images were converted to appear with the
imaging characteristics of the CR system and its associated X-ray system. Both sets of
398 A. Mackenzie et al.
images were processed using Agfa Musica 2. A simulated CR and a real CR image
were shown to a set of 6 observers and they were asked to identify the real CR image.
The observers were shown 10 pairs of images.
Secondly, a sub-set (six images) of the ASE images containing a confirmed cancer
classified as being ‘subtle’ were selected. To demonstrate the conversion process,
these images were changed to appear with the imaging characteristics of the CR
detector using the same X-ray system and grid as the ASE system. Therefore no
correction was made for differences in scatter and grid attenuation. The simulated
images were visually examined for the effects of the conversion.
3 Results and Discussion
3.1 Conversion of CDMAM Test Object Images
The threshold gold thicknesses of simulated CR CDMAM images and the
corresponding real CR images for two beam qualities and PMMA thicknesses show a
close match (Fig. 1). The average differences between the results were 2.5% and 0.3%
for the thin and thick phantoms respectively. This is an encouraging result because the
CDMAM test object provides a good overall measure of the image quality in terms of
noise, sharpness and contrast.
32 mAs (ta rget CR)
32 mAs (AS E)
32 mAs (simu lated CR)
Diameter (mm)
Threshold gold thickness (μm)
0.1 0.13 0.16 0.2 0.250.31 0.4 0.5 0.63 0.8 1.0
0.03
0.05
0.10
0.50
1.00
2.00
CDMAM
+ 2 cm PMM A
125 mAs (ta rget CR)
130 mAs (AS E)
0.1 0.130.16 0.2 0.25 0.31 0.4 0.5 0.63 0.8 1.0
0.03
0.05
0.10
0.50
1.00
2.00
125 mAs (s imula ted CR)
Diameter (mm)
Threshold gold thickness (μm)
CDMA M
+ 6 cm PMMA
Fig. 1. Threshold gold thickness curves for CDMAM images acquired for the thin (2 cm
PMMA) phantom (left), the thick (6 cm PMMA) phantom (right). Results are shown for
simulated and target (real) images obtained with the CR detector and for the original ASE
images from which the CR images were simulated.
3.2 Power Spectra of Conversion of Images of Rachel Anthropomorphic
Phantom
No artefacts were seen in the simulated image of the Rachel phantom. The power
spectra of real and simulated Rachel phantom images obtained with the CR system
are shown in Fig. 2. The results show a good match between the images, with a
Adjusting Image Quality of Mammograms 399
maximum difference of 17% between the simulated and real images, the average
difference was 2% over all spatial frequencies.
Frequency (mm
-1
)
Power spectra (mm
2
)
0.1 110 100
0.001
0.01
0.1
1
10
100
Real CR imag e
Simulated CR image
Fig. 2. Power spectra of ‘Rachel’ phantom of target CR and simulated CR
3.3 Subjective Evaluation of Realism of Converted Images
Using 6 observers and 10 image pairs of a real and simulated CR image, the real
image was correctly identified 32 times out of 60 and the simulated CR image was
incorrectly selected as the real CR image 28 times out of 60. There was a slight
majority of the real image being correctly identified. While these number of results
are very small, it does give an indication that the images produced do look realistic.
3.4 Examples of Conversion of Clinical Images Suitable for an Observer Study
Figs. 3 & 4 show on the left a high quality image of a lesion acquired on a ASE
system. The image on the right shows the image after it has been degraded to have the
image quality of a CR system. There are noticeable differences between the ASE and
simulated CR images in terms of sharpness and noise. The cancers are still visible in
the CR images but the interest of this work is whether the detector used affects the
detection and diagnosis of cancer in breast imaging. An observer study has been
undertaken using this image modification process. The study showed a difference in
the detection of subtle calcifications between ASE detector and a generic CR detector
[6]. The advantage of this method is that the only difference between the two sets of
images is the detector, and that differences due to breasts, compression, anti-scatter
grid, X-ray tube have been removed.
This methodology can be applied to simulate images from different detectors or
even theoretical detectors. A range of observer studies can be undertaken on the effect
of the receptor performance and dose on the detection of different signs of breast
cancer (e.g. masses, calcifications).
400 A. Mackenzie et al.
Fig. 3. Example images of micro-calcification cluster for the original ASE system (left) and
simulated CR (right)
Fig. 4. Example images of mass for the original ASE system (left) and simulated CR (right)
4 Conclusions
We have developed a conversion methodology to change an image to appear as if
acquired on a different imaging system which accounts for the change in detector,
X-ray system and beam quality. The methodology has been successfully applied to
contrast detail measurements. Images with the appearance of a realistic clinical CR
mammogram can be produced without artefacts. The use of this tool in conjunction
with observer studies can be used to understand the effects of detector characteristics
on cancer detection.
Acknowledgements. This work is part of the OPTIMAM project and is supported by
Cancer Research-UK & EPSRC Cancer Imaging Programme in Surrey, in association
with the MRC and Department of Health (England).
The authors are grateful for the help and support of staff from St George’s Hospital,
London, and Jarvis Breast Screening Unit, Guildford. The authors acknowledge Hologic
Inc., MIS Healthcare, and Carestream Healthcare for their help in accessing images and
Agfa Healthcare for the use of their image processing package. We thank our NCCPM
colleagues Lucy Warren and Faith Green who have helped with the collection of
images, and all of our colleagues who took part in the observer study. We thank our
colleagues at Katholieke Universiteit Leuven for helpful discussion of this work.
Adjusting Image Quality of Mammograms 401
References
1. Mackenzie, A., Dance, D.R., Workman, A., Yip, M., Wells, K., Young, K.C.: Development
and validation of a method for converting images to appear with noise and sharpness
characteristics of a different detector and X-ray system. Med. Phys. 39, 2721–2734 (2012)
2. Boone, J.M., Fewell, T.R., Jennings, R.J.: Molybdenum, rhodium, and tungsten anode
spectral models using interpolating polynomials with application to mammography. Med.
Phys. 24, 1863–1874 (1997)
3. Berger, M.J., Hubbell, J.H., Seltzer, S.M., Chang, J., Coursey, J.S., et al.: XCOM: Photon
cross sections database. NIST Standard Reference Database 8, 87-3597 (1998)
4. Dance, D.R., Skinner, C.L., Young, K.C., Beckett, J.R., Kotre, C.J.: Additional factors for
the estimation of mean glandular breast dose using the UK mammography dosimetry
protocol. Phys. Med. Biol. 45, 3225–3240 (2000)
5. Young, K.C., Alsager, A., Oduko, J.M., Bosmans, H., Verbrugge, B., et al.: Evaluation of
software for reading images of the CDMAM test object to assess digital mammography
systems. In: Proc. SPIE, vol. 6913, pp. 69131C-1–69131C-11 (2008)
6. Warren, L.M., Mackenzie, A., Cooke, J., Given-Wilson, R., Wallis, M.G., et al.:
Mammographic calcification cluster detection and threshold gold thickness measurements.
In: Proc. SPIE, vol. 8313, p. 83130J (2012)
