Professional Documents
Culture Documents
SDCT Noise Properties
SDCT Noise Properties
Jeffrey F. Williamson
Medical College of Virginia, Virginia Commonwealth University, Richmond, Virginia 23298
Received 25 January 2006; revised 16 June 2006; accepted for publication 22 June 2006;
published 24 August 2006
The accurate determination of x-ray signal properties is important to several computed tomography
CT research and development areas, notably for statistical reconstruction algorithms and dosereduction simulation. The most commonly used model of CT signal formation, assuming monoenergetic x-ray sources with quantum counting detectors obeying simple Poisson statistics, does not
reflect the actual physics of CT acquisition. This paper describes a more accurate model, taking into
account the energy-integrating detection process, nonuniform flux profiles, and data-conditioning
processes. Methods are developed to experimentally measure and theoretically calculate statistical
distributions, as well as techniques to analyze CT signal properties. Results indicate the limitations
of current models and suggest improvements for the description of CT signal properties. 2006
American Association of Physicists in Medicine. DOI: 10.1118/1.2230762
Key words: x ray, computed tomography, probability distribution function, flux uniformity, signal
detection, energy-integrating detection
I. INTRODUCTION
The introduction of computed tomography CT in the early
1970s revolutionized diagnostic imaging. In the past decade
CT has experienced rapid advancement in technical performance in terms of expanded volume coverage and faster scan
times, enabling new clinical applications such as population
screening and real-time cardiac imaging. In addition to improving hardware design, challenges for continued progress
in CT include better reconstruction algorithms to extract
more information from scan measurements, and developing
scan protocols to reduce the risk of associated radiation exposure. Potential solutions for these two tasks rely on accurate physical models of CT signal properties.1
Iterative reconstruction algorithms, based on statistical
optimization principles, hold promise to minimize noise and
artifact impact, allowing more quantitative diagnostic
information.2 Linear reconstruction algorithms, such as filtered back projection, equally weight the contributions of all
measurements to the reconstructed image, implicitly assuming that the signal-to-noise ratio SNR is constant for all
measurements. Typically, even statistically based reconstruction algorithms, which attempt to improve results by inversely weighting the contribution of measurements to an
image by the estimated noise in the measurement, have assumed a homogenous beam flux incident on the object in
actual implementations. In order to minimize patient radiation dose, however, modern scanner designs have evolved to
produce a beam flux that is nonuniform in space, due to
bowtie equalization filters, and time, due to adaptive tubecurrent modulation. The development of statistical recon3290
0094-2405/2006/339/3290/14/$23.00
3290
3291
3291
ability of observing a certain number of quanta within a particular energy range during a certain measurement period
given by a Poisson distribution that is characterized by a
mean number of quanta per measurement, i.e., I0EE. As
this primary x-ray beam passes through physical material, it
is attenuated by random absorption and scattering events in
an energy-dependent fashion, and the mean number of surviving quanta, IsE, is described by Beers law,17
2. Detectors
The attenuated x-ray beam described above that exits an
object is measured by detectors, which convert the x-ray flux
into an electrical signal and eventually to a digital value. The
interaction of each x-ray quantum with a detector is also a
random process, with an energy-dependent absorption E
= 1 expdetEtdet, where detEtdet is the attenuation
in the detector. For modern solid state detectors such as
1.2 mm of Gd2O2S this absorption efficiency is relatively
high21 almost unity for energies 70 keV, 70% at
150 keV and the interaction process is governed by binomial statistics. When a binomial process acts on a Poisson
process, the result is a Poisson process with an effective
flux lowered to an amount equal to the absorbed signal,
EIsE.8
Once an x-ray quantum is absorbed in the detector, the
conversion process is characterized by a mean energydependent gain, gE, so that the generalized mean signal S
is defined as
3292
FIG. 1. Input spectra generated for a CT scanner 7 tube angle, 120 kVp,
104 cm distance, 12 mm Al Cu filtration for air a and 30 cm water attenuation b. The HVL of the incident spectrum is 8.8 mm Al, with a flux
level of 1.09 106 photons per mAs mm2. The mean photon energy increases from 66.1 to 78.4 due to beam hardening. The energy resolution is
1 keV; characteristic peaks are actually 1 2 keV wide.
S = I
o
gEEEexp EtdE,
3292
3. Scanner model
Besides the x-ray source and detection process, several
additional factors significantly influence each individual
x-ray CT measurement and must be accounted for.
Nonuniform flux across fan beam. This is generated
by at least two effects. First, bowtie filters are introduced into the beam to minimize patient dose. Because most human anatomy is nominally circular in
cross section, with more attenuation in central areas
compared to the periphery, a uniform flux that gives
acceptable signal-to-noise ratio in the center of the
patient would give higher SNR through the edge of
the patient but at the expense of unnecessary dose
there. A tapered filter is used to attenuate flux toward
the fan-beam edges while maintaining adequate
overall SNR. The magnitude of this exposure reduction can be quite large, up to a factor of 10 or more.
A second source of nonuniformity is due to the heel
effect: Modern x-ray tubes generate a fan-beam
plane that is normal to the scan axis, while the plane
of the target anode surface is at an oblique angle to
the scan axis. Rays at an increasing angle in the fan
beam have a longer exit path in the anode and experience more attenuation in the target; hence there is a
falloff in intensity at a higher fan-beam angle, estimated to be on the order of 4%. Consider a 60-keV
electron striking a tungsten anode with a tube angle
of 7 and a fan-beam angle of +/26. Using the
method of Fritz and Livingston,26 the mean electronstopping distance will be 6 m, resulting in an exit
distance in the anode ranging between 49 and 54 m
across the fan beam. The self-attenuation of tungsten
will lead to a decrease in intensity of 4% at the edge
of the fan beam relative to the central ray. Therefore, the heel effect is much smaller in magnitude
3293
Ameas = log
aS0 + b
,
aSm + b
3a
3293
2
Acorr = Ameas + cAmeas
,
3b
N ID
e
ID
.
N!
PS, =
N=0
N
ID
S/G NeID
,
N!
3294
3294
FIG. 2. Plot of first ten pdfs for exactly N quanta formed by convolution of
original spectra from 35.4 cm PMMA cylinder and overall pdf scaled by a
factor of 10 for visibility for mean quanta number= 5. Note that the fine
structure of the spectrum is blurred after several convolutions.
B. Experimental techniques
1 =
kVp
gEEdE,
g2EEdE,
g3EEdE,
g4E 322EdE.
2 =
kVp
3 =
kVp
4 =
kVp
Experiments were performed to gather data to characterize signals in clinical CT scanner systems. Several generations of CT scanners all manufactured by Siemens Medical
Solutions, Forchheim Germany were tested, including
single-row helical scanners Somatom Plus 4, four-row
scanners Somatom Volume Zoom, and sixteen-row scanners Somatom Sensation 16. Sinogram data files were exported from the system via optical disks or network transfer
for analysis.
The sinogram data consisted of attenuation measurements, i.e., the logarithm of the ratio of unattenuated signal
to measured signal, as well as a recording of the tube current
for each measurement. The attenuation data has been corrected for beam hardening, detector nonuniformity, and tubecurrent fluctuation Eq. 3. To execute these measurements
and corrections at a rate of millions per second requires
implementation of a signal-processing chain with low-level
hardware and firmware; as a result, the original data measurements or coefficients used by individual detectors are not
accessible for inspection. Numerous inquiries were made to
several CT manufactures regarding availability of uncorrected measurements or individual correction coefficients.
Such information was not available for any of the scanners
presented in this paper. Of the eight scanner designs accessible in the authors institutions, only two of them provided
access to uncorrected raw data in a service-maintenance
mode. While this means that the statistics of the dataconditioning chain cannot be examined directly, the requirements for human visual sensitivity to detecting artifacts place
constraints on the resultant errors caused by such signal
processing.32 That is, artifacts introduced by mismatched detectors or tube current fluctuations, such as linear streaks or
ring artifacts in images, will be detectable at a contrast level
of about 0.15% on the order of one Hounsfield Unit, corresponding to an NEQ of 500 000. The absence of such artifacts in image slices implies that the conditioning is accurate to at least this visualization level, and hence the impact
3295
3295
The exact form of beam filtration for each scanner was not
known, but could include inherent tube filtration, bowtie filters, and gantry shroud encasements. However, physical
specifications for each scanner model are reported by the
manufacturer or governmental agencies e.g., ImPACT, UK,
http://www.impactscan.org/, and QC measurements are performed locally on each device, including the beam quality
HVL and air kerma. In this study, the specified HVL and
air kerma for a device were used as a goal for adjusting the
filtration and the output of Xcomp5, and the resulting spectrum is used in the calculations as the incident spectrum
Fig. 1. As will be shown in the analysis, due to the central
limit theorem, the absolute accuracy of the spectrum is not
essential to the conclusions presented here.
C. Analysis methods
3296
3296
the attenuation is the NEQ since the mean of the unattenuated signal is nominally 1.0. Because analysis of the excess
noise due to spectrum shift bowtie-filter beam hardening
indicates only small changes in variance dependence, the
NEQ can be used as an estimate of the flux level.
For sinograms, there are typically between 1050 and 2320
gantry steps per revolution, so reasonable accuracy was obtained by using data from one rotation. Once the variance
profile for the detectors has been calculated, the resulting
profile was fit to an eighth-order polynomial and this fitprofile function was used to estimate the relative flux level
for analyzing detection statistics or in noise synthesis.
2. Cylinder phantom profiles
To measure the distribution of sinogram fluctuations about
a range of mean values, measured sinogram profiles for
simple cylindrical phantoms were fit to an expected meanvalue profile, and the statistical distribution of detector readings was approximated by the histogram of the fit residuals.33
These histograms were used to generate pdfs and their moments for different signal levels.
When scanned, the PMMA cylinders were centered and
aligned precisely but not perfectly. Ideally for a monoenergetic beam, their profile would be the fan-beam projection of
a circle, described by
PD = 2R2 L2sin2D D0,
10
11
FIG. 5. Plot of measured cylinder attenuation and parametrized fit of cylinder profile. Data from 16-row scanner, with 25 mAs exposure.
3297
3297
B. Fan-beam profile
A. Attenuation linearity
3298
3298
FIG. 9. Patient attenuation sinogram profiles for an adult head scan, abdominal scan, and pelvic scan in patient with bilateral hip implant. Note that
the latter scan demonstrates the range of maximum attenuations observed
clinically.
C. Flux levels
can be combined to improve the SNR of data. Clinical protocols are usually adjusted to limit the amount of noise
present at the expense of spatial resolution.
D. Pdf forms
TABLE I. Summary statistics for theoretically calculated pdfs. The columns are: diameter of cylinder phantom, effective energy of exit spectrum, excess
factors for cumulants variance, skewness, and kurtosis relative to Poisson distribution for a monoenergetic source, the total flux, and absolute value of NEQ,
skew, and kurtosis for a 25-mAs scan.
Diameter
cm
Mean
energy
keV
Excess
variance
Excess
skew
Excess
kurtosis
Qo
NEQ
Skew
Kurtosis
0
5
10
15
20
25
30
35.4
66.6
69.0
71.2
73.2
75.2
77.0
78.7
80.4
1.082
1.076
1.071
1.067
1.063
1.060
1.057
1.053
1.257
1.238
1.222
1.207
1.195
1.183
1.172
1.161
1.558
1.510
1.471
1.437
1.407
1.380
1.354
1.329
46982.8
15974.1
5534.6
1945.1
695.3
249.0
89.3
32.9
43442.2
14847.2
5166.7
1822.8
653.9
234.9
84.5
31.2
6.66E-10
5.61E-09
4.58E-08
3.63E-07
2.79E-06
2.14E-05
1.64E-04
1.19E-03
1.90E-14
4.61E-13
1.07E-11
2.37E-10
5.03E-09
1.06E-07
2.25E-06
4.36E-05
3299
3299
FIG. 12. Plot of measured pdf for single-row scanner for mean transmission
4.45 104, corresponding to region in Fig. 11 with excess system noise.
The calculated primary beam NEQ is 32 but the measured NEQ is 2.5, due
to high system noise.
FIG. 13. Plots of measured signals for six central detectors over one gantry
revolution, at low-flux level 50 mAs in four row scanner in 35.4-cm-diam
cylinder phantom. The measurements cluster at quantization levels corresponding to a step size of 5 105. Note that the overall range of measurements is large 103 relative to the quantization steps, so that the contribution of quantization to total variance is not significant.
3300
3300
not be resolved. Even for a well-designed ADC, the resolution limit is likely to be several hundred keV and will prevent the detection of individual quanta.
IV. DISCUSSION
The methods and measurements presented above allow a
comparison between actual CT scanner data, the conventional monoenergetic-counter model, and a more accurate
energy-integrating model. The first noticeable observation is
that real CT data distributions consist of continuous rather
than discrete values, as is correctly predicted by the energyintegrating model. To first order, the variance of actual signal
measurements is proportional to the mean of the signal for
all cases. Theoretical models predict that the absolute variance of an ideal energy integrator would be 2%10% higher
than that of an ideal counting detector over the range of
clinical signals, being a function of object attenuation and
beam spectrum, which is consistent with measurements to
within our experimental accuracy. In many applications the
counter model can be simply rescaled to the data as an effective NEQ, so the actual residual error due to such spectral
variation could be limited to only a few percent. Current
clinical CT protocols normally produce measurements with
fairly high minimum NEQ, in excess of 30. Under these
conditions, measurements are normally distributed and
choice of a statistical model is not critical. As the flux to the
detector decreases, however, the forms of the distribution
functions become quite different and need to be accounted
for.
A second observation of CT signal statistics is the dominant role that nonuniform beam flux has on signal levels. To
minimize radiation risk to the patient, CT scanners are designed to reduce the flux levels in certain areas of the fanbeam measurements by substantial amounts, using bowtie
filters or tube current modulation. With proper calibration,
this will leave the means of attenuation measurements unaffected but can strongly influence noise properties.
A third influence on CT signal statistics comes from data
conditioning, such as calibrating individual detector measurements or compensating for polyenergetic beamhardening nonlinearities. This must be done very accurately
to avoid visual artifacts in the image, but can result in significant modification of statistical properties of noise. However, the effect on small-signal contrast-to-noise ratio is not
significant, as both signal and noise are propagated through
the processing.
The overall trends in image noise are summarized in Fig.
14, which presents the relative variance ratio of variance to
that of an ideal monoenergetic counter as a function of reported attenuation. At very low or very high attenuations,
effects such as quantization or data conditioning will significantly effect the variance of signals, depending on scanner
design and operating parameters. In mid-attenuation regions,
the Compound Poisson statistics of energy-integrating detection contribute to an increase in variance of less than 10%
relative to ideal-counter detection, which is on the order of
preprocessing contributions. The dominant determinate of
Medical Physics, Vol. 33, No. 9, September 2006
FIG. 14. Plot of ratio of noise variance of total system model including
energy integrating excess noise, quantization noise in ADC and log data,
system noise, beam-hardening corrections, variation in electronic gain and
offset, and bowtie filter relative to conventional counter model monoenergetic, uniform flux, no additional noise sources as a function of corrected
attenuation. The dashed line represents ideal variance reference unity,
lower curve is for central ray minimum bowtie attenuation, and upper
curve is for bowtie transmittance of 0.25. Parameters used include 170 000
quanta incident per measurement, ADC quantization of 105 and log quantization of 4.4 104, electronic noise of 5 quanta, gain and offset accuracy
of 0.14%, and quadratic beam-hardening correction. At high-flux levels low
attenuation in central ray, excess noise is dominated by quantization and
compensation error; at low flux, system noise and beam-hardening corrections contribute to increases; at mid-range flux levels, polyenergetic excess
noise increases the relative variance. In all cases, the flux levels produced by
the bowtie filter dominate the relative variance ratio.
3301
Cf, = expeif 1,
kVp
Q0EeiGEf 1dE .
Cf, = exp
A2
kVp
Q0EeiGEf 1 dE
exp
kVp
Q0EdE
= exp
Q0EeiGEf dE
A3
kVp
= exp QT + QGf,
3301
Equations 3a and 3b represent the conditioning operations performed on measurements to generate data appropriate for image reconstruction. Errors or variability in the coefficients of these operations could possibly influence the
variance in the recorded signals. To analyze this effect, assuming a linear system model,38 the measurement attenuation is restated by noting that in the numerator b aS0, and
the gain and offset coefficients can be combined into one
term,
Am = log
eiSf eQT+QGfdf
B1
= eQT
=e
QT
aS0 + b
aS0
S0
log
= log
.
aSm + b
aSm + b
Sm + b/a
k=0
= eQT
k=0
eiSf
k=0
QGfk
df
k!
eiSf QGfkdf
A2 S0
m
k!
Q0S/Gk
.
k!
A4
2
Am 2 2 S0 S20T2
=
S0 = 2 = 2 = T2 ,
S0
S0
S0
B2
3302
Am
b/a
2
2
b/a
=
2
b/a
.
Sm + b/a2
B3
In order to avoid the generation of visual artifacts in reconstructed images, the electronic factors must be properly
matched to within a fractional signal error 0.0015Sm
+ b / a, limiting the variance to
A2 b/a =
m
2
b/a
Sm + b/a2
=
0.00152 = 2 106 .
Sm + b/a2 Sm + b/a2
B4
As a point of reference, the contribution to
quantum noise is
A2 QN
m
A2
m
from
S0
Am 2 2
=
Sm =
Sm
Sm + b/a2
2
E2
E + b/a
1
b/a
1+
E
2,
B5
where is the mean number of quanta and E is the average
energy of the quanta. Therefore, the quantum noise variance
is inversely proportional to the number of quanta, except
when the system noise becomes comparable to the signal
level. The contributions to measurement variance due to data
conditioning will thus be small, except for very large signals
where 1 Ameas + A2 b/a or very small signals
m
E b / a.
Application of a beam-hardening correction Eq. 3a
will also affect the variance properties of the processed signal, computed as
A2 corr =
=
Acorr 2 2
Ameas
Ameas
2
Ameas + cAmeas
Ameas
= 1 + 2cAmeas2A2
meas
2
Ameas
B6
A2 corr
A2 meas
expAmeas
expAmeas
=
2
expAcorr expAmeas + cAmeas
2
,
= exp cAmeas
B7
i.e., the variance of the corrected signal is less than the exMedical Physics, Vol. 33, No. 9, September 2006
3302
pected value, due to the fact that beam hardening lessens the
attenuation of the beam signal. In total, the relative attenua2
tion signal variance will differ by the amount corr
2
= 1 + 2cAmeas2expcAmeas
, shown in Fig. 14.
a
3303
3303