M e d i c a l P hy s i c s a n d I n f o r m a t i c s • O r i g i n a l R e s e a r c h

Solomon et al.
Noise and Image Quality Indicators in CT Using Tube Current
Modulation

Medical Physics and Informatics

Original Research
Downloaded from www.ajronline.org by Duke University Medical Center on 03/03/13 from IP address 152.16.51.172. Copyright ARRS. For personal use only; all rights reserved

Relating Noise to Image Quality

Indicators in CT Examinations
With Tube Current Modulation
Justin B. Solomon1 OBJECTIVE. Modern CT systems use surrogates of noise—noise index (NI) and quality
Xiang Li reference effective tube current-time product (Q)—to infer the quality of images acquired us-
Ehsan Samei ing tube current modulation. This study aimed to determine the relationship between actual
noise and these surrogates for two CT scanners from two different manufacturers.
Solomon JB, Li X, Samei E MATERIALS AND METHODS. Two phantoms (adult and 1-year-old child) were im-
aged on two CT scanners (64 and 128 MDCT) using a clinical range of NI (6–22) and Q (30–300
mA). Each scan was performed twice, and noise was measured in the mediastinum, lung, and ab-
domen using an image subtraction technique. The effect on noise from changing other imaging
parameters, such as beam collimation, pitch, peak kilovoltage, slice thickness, FOV, reconstruc-
tion kernel or algorithm, and patient age category (adult or pediatric), was investigated.
RESULTS. On the 64-MDCT scanner, noise increased linearly along with NI, with the
slope affected by changing the anatomy of interest, peak kilovoltage, reconstruction algo-
rithm, and convolution kernel. The noise–NI relationship was independent of phantom size,
slice thickness, pitch, FOV, and beam width. On the 128-MDCT scanner, noise decreased
nonlinearly along with increasing Q, slice thickness, and peak tube voltage. The noise-Q rela-
tionship also depended on anatomy of interest, phantom size, age selection, and reconstruction
algorithm but was independent of pitch, FOV, and detector configuration.
CONCLUSION. We established how noise changes with changing image quality indi-
cators across a clinically relevant range of imaging parameters. This work can aid in optimiz-
ing protocols by targeting specific noise levels for different types of CT examinations.

Keywords: CT, noise, noise index, reference tube
current, tube current modulation
DOI:10.2214/AJR.12.8580
Received January 17, 2012; accepted after revision
June 4, 2012.
1
All authors: Carl E. Ravin Advanced Imaging Laborato-
ries and Clinical Imaging Physics Group, Department of
Radiology, Medical Physics Graduate Program, Duke
University Medical Center, Hock Plaza, 2424 Erwin Rd,
Ste 302, Durham, NC 27705. Address correspondence to
J. B. Solomon (justin.solomon@duke.edu).
AJR 2013; 200:592–600
0361–803X/13/2003–592
© American Roentgen Ray Society
A
dvances in x-ray CT technology
and realization of its unprecedent-
ed clinical utility have resulted in
an increase in the number of CT
examinations performed in the United States.
This has resulted in a large increase in radiation
dose received by patients as a result of CT ex-
aminations [1, 2]. Automatic tube current mod-
ulation (ATCM) techniques were developed,
in part, in an effort to reduce dose and main-
tain image quality for patients of different siz-
es and shapes [3–6]. In CT examinations using
ATCM, a level of image quality is targeted by
setting an image quality indicator. The ATCM
systems then use patient size and attenuation
properties to modulate the x-ray tube current
(i.e., photon fluence) to create images of the
desired quality [7]. Generally, modulating the
tube current according to patient size and shape
is sought to control the photon fluence at the
detectors. Because image noise is inversely
proportional to the square root of fluence, these
systems are attempting to control quantum
noise according to the selected image quality
indicator. Therefore, an image quality indica-
tor can be thought of as a surrogate for expect-
ed noise. The ATCM system eliminates the
need for a technologist to manually adjust the
tube current for different patients while at the
same time automatically adjusting the tube
current to accommodate differing attenuation
properties within a patient habitus.
When designing and optimizing CT proto-
cols, one is required to balance the need for
acceptable image quality with the mandate to
keep radiation exposure “as low as reason-
ably achievable.” This is especially important
for clinical tasks, such as pediatric lung nod-
ule detection, that are particularly sensitive to
image quality. Li et al. [8] have shown that it
is possible to establish a relationship between
measureable image quality and radiologist per-
formance in pediatric lung nodule detection.
ATCM makes it possible to design protocols
that will produce images of similar quality for
patients of different shapes and sizes [9], but

592 AJR:200, March 2013

 Noise and Image Quality Indicators in CT Using Tube Current Modulation

what measureable noise levels can be expect-
ed for a given quality reference indicator? In
other words, if we know what noise level is
acceptable to achieve a certain level of di-
Downloaded from www.ajronline.org by Duke University Medical Center on 03/03/13 from IP address 152.16.51.172. Copyright ARRS. For personal use only; all rights reserved
according to the Q and individual patient size ferent patient categories was further explored on
and attenuation properties [9]. The Siemens the Somatom Definition Flash.
Healthcare ATCM algorithm does not alter the Because of scan time restraints and the num-
tube current according to any reconstruction ber of parameters investigated, data were not col-

agnostic performance, how can we choose a
quality reference indicator to produce imag-
es with the desired noise level? Complicating
this answer is the fact that different manufac-
turers apply ATCM differently and use dif-
ferent quality reference indicators.
The purpose of this study was to quantita-
tively assess the relationship between actual
quantum noise and image quality indicators
across imaging parameters in the context of
two specific commercial CT platforms man-
ufactured by GE Healthcare and Siemens
Healthcare. Both platforms use simultaneous
longitudinal (z-axis) and angular (x–y) mod-
ulation, which has been shown to decrease
dose compared with systems that use only an-
gular modulation or no modulation [10]. On
the GE Healthcare scanner, one selects a noise
index (NI) and specifies minimum and maxi-
mum tube current levels. NI is expected to re-
flect the noise in the central region of a uni-
form phantom. The GE Healthcare ATCM
algorithm will attempt to produce images with
noise consistent with the selected NI, indepen-
dent of peak kilovoltage, patient size, and pro-
spectively chosen reconstructed slice thick-
ness [7]; therefore, one would expect actual
image noise to correlate positively with NI
and be independent of peak kilovoltage, pa-
tient size, and prospective slice thickness. On
a Siemens Healthcare scanner, one selects an
image quality reference effective tube current-
time product (Q). The Q represents the mean
effective tube current used on a reference pa-
tient. The system then adapts the tube current
parameters, nor does it attempt to achieve con-
stant noise levels across patients of different
sizes. Given such a design and known physics
of CT, one would expect actual image noise
to be inversely related to the Q and be depen-
dent on reconstruction parameters and patient
size. The purpose of the study was to investi-
gate how the actual noise in a patient image
varies as a function of NI and Q across a range
of imaging parameters.
Materials and Methods
A 64-MDCT scanner (Discovery CT 750 HD,
GE Healthcare) and a 128-MDCT scanner (So-
matom Definition Flash, Siemens Healthcare)
were chosen for this study. We investigated the
ATCM algorithm SmartmA (GE Healthcare) on
the Discovery CT 750 HD scanner and the ATCM
algorithm CARE Dose 4D (Siemens Healthcare)
on the Somatom Definition Flash scanner. Two
anthropomorphic phantoms, representing an adult
male (73 kg) and a 1-year-old child (10 kg) (ATOM
models 701 and 704, Computerized Imaging Ref-
erence Systems), were imaged on the two systems.
A clinical range of imaging parameters was
chosen to investigate the relationship between im-
age quality indicators and measured noise. The
parameters of interest were beam collimation, he-
lical pitch factor, peak tube voltage, slice thick-
ness, reconstructed FOV, reconstruction algo-
rithm, and the convolution kernel (reconstruction
filter) (Table 1). Because Siemens Healthcare’s
CARE Dose 4D modulation uses different atten-
uation curves for different patient age categories
(e.g., adult or pediatric), the effect of selecting dif-
lected for every combination of the parameters
listed. Also, system limitations on tube output
made some acquisitions unachievable. For exam-
ple, at 80 kVp, the Discovery CT 750 HD cannot
achieve a tube current sufficient to image an adult
patient at an NI of 6. In general, data were collect-
ed by first choosing a reference protocol accord-
ing to clinical settings. The phantom was imaged
using this reference protocol at various values of
the system’s image quality indicator. The refer-
ence protocol was then changed one parameter
at a time, each time imaging with several image
quality indicator values. This made it possible to
systematically explore how each parameter af-
fects the relationship between noise and the im-
age quality indicator in question. Reference pa-
rameters are indicated with footnotes in Table 1.
Note that, on the Discovery CT 750 HD, the mini-
mum and maximum tube current were set to their
lowest and highest available values to ensure that
the required tube current (as determined by the
SmartmA algorithm) was not clipped during the
scan. The phantoms were scanned from the top of
the lungs through the abdomen to include a range
of anatomic and attenuation variations. Each scan
was repeated yielding duplicate series. Duplicates
were then subtracted from each other to remove
anatomy (i.e., signal) (Fig. 1).
For each series, three square regions of interest
(ROIs) were defined: one in the mediastinum, one in
the lung, and one in the abdomen. Noise magnitude
(pixel SD) in the subtracted images was measured
over multiple slices and averaged. These noise mea-
surements were divided by the square root of 2 to ac-
count for the noise added as a result of the subtraction

TABLE 1: Imaging Parameters Used to Acquire Data

Scanner Model and Manufacturer, Parameter Value
Parameter Discovery CT 750 HD, GE Healthcare Somatom Definition Flash, Siemens Healthcare
Phantom Adult malea and 1-year-old child Adult malea and 1-year-old child
Image quality indicator Noise index: 6, 9, 11.57, 16, 19, and 22 Quality reference effective tube current-time product (mAs): 30,
60, 75, 120, 150, 200, and 300
Beam collimation (mm) 40 (64 × 0.625)a and 20 (32 × 0.625) 38.4 (128 × 0.6)a,b and 38.4 (32 × 1.2)
Pitch factor 0.969, 0.984, and 1.375a 0.8a and 1.4
Tube voltage (kVp) 80, 100, 120,a and 140 80, 100, 120,a and 140
Slice thickness (mm) 0.625, 1.25, 2.5, and 5.0a 0.6, 1.0, 3.0, and 5.0a
Reconstructed FOV (mm) 150 and 340a 150, 340,a and 400
Reconstruction algorithm FBPa and 30% adaptive statistical iterative reconstruction FBPa and iterative reconstruction in image space
Convolution kernel Standard,a bone, soft, and detail B31fa and I31f
Note—FBP = filtered back projection.
aReference protocol against which other protocols were compared.
bThe actual number of detector rows is 64, but the use of a flying focal spot allows double sampling, resulting in 128 data channels.

AJR:200, March 2013 593


Downloaded from www.ajronline.org by Duke University Medical Center on 03/03/13 from IP address 152.16.51.172. Copyright ARRS. For personal use only; all rights reserved
Fig. 1—Example image of adult phantom (left) with subtracted image (right) showing location of lung and
mediastinum regions of interest (squares).
process (i.e., error propagation). Also, the SD and co-
efficient of variation of noise across slices were re-
corded. The number of slices measured in each series
depended on the slice thickness, because a constant
axial length of approximately 6 cm was considered
for each volume ROI. The in-slice size of the ROI de-
pended on the phantom and the FOV of the particular
series. ROI size varied between 1 and 4 cm2, which
corresponded to about 360–1000 pixels per ROI.
The measured noise data were analyzed to de-
termine the effect of scan and reconstruction pa-
rameters on the noise–image quality indicator re-
lationship. For each system, noise was examined
as a function of its image quality indicator using
a least-squares fit technique (Matlab 2010, Math-
works). Also, other parameters that were found to
have a significant effect on the noise–image qual-
22
20
18
16
Noise (HU)
14
12
10
8 400 to 1600 pixels) and location (variable
6 within the anatomic region) were 3% and
4
4 9 14 19
Noise Index
Fig. 2—Plot of noise in mediastinum against noise
index for adult phantom under reference conditions
(120 kVp, pitch of 1.375, 40-mm beam collimation,
slice thickness of 5 mm, and reconstruction with
34-cm FOV using filtered back projection and
standard filter). Dashed line is added as reference to
show where noise index is equal to noise. Error bars
represent 1 SD.
594
Solomon et al.
ity indicator relationship were incorporated into
the models. To determine whether a parameter
had a significant effect on the noise–quality indi-
cator relationship, for parameters from which only
a few comparable noise values were available, ei-
ther the percentage difference or coefficient of
variation was used. If the percentage difference or
coefficient of variation was greater than the aver-
age percentage error of the measurements, the pa-
rameter was deemed to have a significant effect.
For parameters with many comparable noise mea-
surements, a paired Student t test (JMP Pro ver-
sion 9, SAS Institute) was used. If the p value was
less than 0.05, the parameter was deemed to have
a significant impact on the relationship.
Results
Because each vendor implements ATCM
in a different way, the results for each scan-
ner will be reported separately. However,
there were consistent findings about the un-
certainty associated with each noise measure-
ment that should be mentioned initially. First,
changing the ROI size or location within the
anatomic structure of interest was found to
have little effect on the noise measurement.
Overall, the variations due to ROI size (from
8%, respectively. Second, the SD associat-
ed with each noise measurement increased as
the average noise increased. The average SD
was 0.6 HU, and the maximum was 2.6 HU.
The average coefficient of variation was 9%,
and the maximum was 19%.
On the Discovery CT 750 HD, for every
set of scan parameters, noise increased lin-
early with NI. An example is illustrated in
Figure 2 for the reference protocol. At a giv-
en NI, the noise-NI relationship was indepen-
dent of phantom size (< 7% difference), slice
thickness (< 5% difference), pitch (< 4% dif-
ference), FOV (< 7% difference), and beam
width (< 4% difference) (Fig. 3). However,
anatomy of interest, peak kilovoltage, con-
volution kernel, and the reconstruction algo-
rithm all yielded a significant effect on the
relationship according the criteria described
in Materials and Methods (Figs. 3 and 4). On
average, for a given image series, lung noise
was 31% lower and abdomen noise was 25%
higher compared with mediastinum noise
measured in the same image series. This is
illustrated by the plots in Figure 4.
Given the design of the SmartmA algo-
rithm [7] and the results already presented,
the noise-NI relationship is best described by
Equation 1:
σ = σref · A = 0.936 · NI · A (1),
where σ is the noise under conditions of in-
terest, σref is the noise under reference condi-
tions, A is an adjustment factor to account for
deviations from reference conditions, and NI
is the NI. The adjustment factor was found
by taking ratios of noise under conditions of
interest to noise under reference conditions
and describes the dependency of the noise-
NI relationship on peak kilovoltage, anato-
my of interest, convolution kernel, and re-
construction algorithm. Adjustment factors
were calculated for each affecting parameter
so that an overall adjustment factor could be
a multiplication of the parameter-based ad-
justment factors, as shown in Equation 2:
A = A Anatomy · AkVp · AKernel · A Algorithm (2).
These adjustment factors are tabulated in Ta-
ble 2.
The model described by Equations 1 and
2 maintains a close fit to measured data. Us-
ing this model, the average residual magnitude
was found to be 0.46 HU (5% error) with a
maximum at 2.5 HU (18% error). The root-
mean square of residuals was 0.7 HU. In gen-
eral, residual magnitude was larger for noisier
images, whereas the less noisy images resulted
in a higher percentage error. Please note that
the adjustment factors listed in Table 2 pertain
only to the range of values studied for a given
parameter. The factors may not be directly ex-
trapolated to other parameter values.
On the Somatom Definition Flash, for every
set of scan parameters, noise decreased with in-
creasing image Q. An example is given in Fig-
ure 5 for the reference protocol. The noise-Q
relationship had little dependence on helical
AJR:200, March 2013
 Noise and Image Quality Indicators in CT Using Tube Current Modulation

A B C
25 Coefficient of 14 Coefficient of 12 Difference = 3%
variation = 3% variation = 7%
12
Downloaded from www.ajronline.org by Duke University Medical Center on 03/03/13 from IP address 152.16.51.172. Copyright ARRS. For personal use only; all rights reserved

20 10
10
Coefficient of 8
Noise (HU)

Noise (HU)

Noise (HU)
Coefficient of
15 variation = 3% 8
variation = 8% Difference = 2%
6
10 6
4
4
5 2
2

0 0 0
11.57 22.00 6.0 11.57 6.0 11.57
Noise Index Noise Index Noise Index
Slice thickness = 0.625 mm Slice thickness = 1.25 mm kVp = 80 kVp = 100 Pitch = 0.984 Pitch = 1.375
Slice thickness = 5.00 mm kVp = 120 kVp = 140

D E F
12 Difference = 6% 14 12 Difference = 6%
Difference = 4%

10 12 10
10
8 Difference = 1% 8
Noise (HU)

Noise (HU)

Noise (HU)
8 Difference = 3%
Difference = 4%
6 6
6
4 4
4
2 2 2

0 0 0
6.0 11.57 6.0 11.57 6.0 11.57
Noise Index Noise Index Noise Index
FOV = 34 cm FOV = 40 cm Beam width = 40 mm Beam width = 20 mm Adult One-Year-Old

G H
35 Coefficient of 14 Difference = 21%
variation = 64%
30 10
25
Coefficient of 8
Noise (HU)

Noise (HU)

20 variation = 64% Difference = 19%

6 Fig. 3—Noise in mediastinum compared with noise
15 index.
4 A–H, Bar graphs show effect of slice thickness
10
(a), peak kilovoltage (b), pitch (c), FOV (d), beam
5 2 collimation (e), phantom size (f), convolution
kernel (g), and reconstruction algorithm (H) on
0 0 relationship between noise and noise index. Each
6.0 11.57 6.0 11.57
plot represents alteration with respect to reference
Noise Index Noise Index
technique by one parameter. Reference parameters
Standard Bone Soft Detail Filtered back projection are underlined in legend of each plot. Error bars
Adaptive statistical iterativere reconstruction 30% represent 1 SD.

pitch factor (< 6% difference), FOV (< 5% dif-
ference), or detector configuration (< 8% dif-
ference) (Fig. 6). However, the relationship was
dependent on slice thickness, peak kilovoltage,
phantom size, patient age selection, anatomy
of interest, and reconstruction algorithm (Figs.
6 and 7). The significant impact on the noise-
Q relationship due to slice thickness, peak ki-
lovoltage, phantom size, age selection, anato-
my of interest, or reconstruction algorithm was
found to be significant according to the criteria
described in Materials and Methods. On aver-
age, for a given image series, lung noise was
27% lower and abdomen noise was 24% higher
compared with mediastinum noise measured in
the same image series. This is illustrated by the
plots in Figure 7.
Given the design of the CARE Dose4D al-
gorithm [11], and the results already present-
ed, the noise-Q relationship is best described
by Equation 3:
eα
σ = A • σref = A •
(Q • t)β • kVpγ
where σ, σref, and A are the same as in Equation
1; e is Euler’s number, α, β, and γ are fitting
parameters; Q is the quality reference effective
tube current; t is the slice thickness in millime-
ters; and kVp is the peak kilovoltage of the x-
ray tube. The fitting parameters α, β, and γ
were found using a linear least-squares fit to
the natural log of noise. The adjustment fac-
tor, A, was found by taking ratios of noise un-
der conditions of interest to noise under refer-
(3), ence conditions and describes the dependency
of the noise-Q relationship on the anatomy of

AJR:200, March 2013 595

 Solomon et al.

40 y = 0.694x 30 y = 1.254x for the same phantom and scan settings, ex-
R2 = 0.99143 R2 = 0.99602 pected noise may be different depending on
35
25 the selection of an adult or pediatric patient.

Abdomen Noise (HU)

30 The model described by Equation 3 fits
Downloaded from www.ajronline.org by Duke University Medical Center on 03/03/13 from IP address 152.16.51.172. Copyright ARRS. For personal use only; all rights reserved

Lung Noise (HU)

20 closely to measured data. The average residu-

25
al magnitude was 0.7 HU (9% error), and the
20 15
maximum was 5.7 HU (53% error). As with
15 the Discovery CT 750 HD, residual magni-
10
10 tude was larger for noisier images, whereas
5
5 the less noisy images resulted in a higher per-
centage error. Again, it should be noted that
0 0
0 10 20 30 40 50 60 0 5 10 15 20 25 the adjustment factors listed in Table 3 per-
A Mediastinum Noise (HU) B Mediastinum Noise (HU) tain only to the range of values studied for a
given parameter. The factors may not be di-
Fig. 4—Mediastinum noise compared with anatomic noise for all acquisitions on Discovery CT 750 HD scanner rectly extrapolated beyond those ranges.
(GE Healthcare).
A and B, Plots show lung (A) and abdomen (b) noise. Each data point represents different image series, but for
given data point, noise values being compared were measured from same image series. Plots provide way to Discussion
translate mediastinum noise to lung and abdomen noise. We have shown how noise changes as a
function of NI and Q across many imaging pa-
interest, phantom size, patient age selection, 3 is different from that of Equation 1 because rameters for two common brands of CT scan-
and reconstruction algorithm. Adjustment of the difference in the design of each manu- ners. To our knowledge, similar evaluations
factors were calculated for each affecting pa- facturer’s ATCM algorithm. have not been reported before. This work has
rameter so that an overall adjustment factor An explanation of the age adjustment fac- the potential to aid in protocol optimization.
could be a multiplication of the parameter- tor is warranted. On the Somatom Definition For example, observer studies can quantita-
based adjustment factors, as shown in Equa- Flash, when entering patient information, the tively show how radiologist performance for
tion 4: technologist selects the relative age of the pa- specific tasks is related to physical metrics of
tient (e.g., adult or pediatric). That selection image quality such as noise [8]. Equations 1–4
A = A Anatomy · APhantom · A Age · A Algorithm (4).
changes the attenuation curves used by the relate noise to image quality indicators and
These adjustment factors found are tabulated system to modulate tube current and there- other image acquisition settings. Therefore,
in Table 3 along with the values of the fitting fore affects noise. This effect is reflected by using Equations 1–4 in conjunction with the
parameters α, β, and γ. The form of Equation the age adjustment factor. This means that, results of observer studies helps one predict
radiologist performance for a given set of im-
TABLE 2: Adjustment Factors Used to Estimate Noise on Discovery CT 750 age acquisition settings. Translating image ac-
HD Scanner (GE Healthcare) for a Given Protocol Governed by quisition settings to radiologist performance
Equations 1 and 3 is necessary to optimize protocols. Further-
Protocol Adjustment Factor more, this work can help in achieving con-
sistent image quality for institutions that use
Tube voltage
80 kVp 1.08
16
100 kVp 1.02
14
120 kVp 1.00
12
140 kVp 0.94
10
Noise (HU)

Anatomic area of interest

Mediastinum 1.00 8

Lung 0.69 6

Abdomen 1.25 4

Convolution kernel 2
Standard 1.00 0
0 50 100 150 200 250 300 350
Bone 2.62
Q (mAs)
Soft 0.81
Fig. 5—Plot of noise in mediastinum against image
Detail 1.05 quality reference effective tube current-time product
Reconstruction algorithm (Q) for adult phantom under reference conditions
(120 kVp, pitch of 0.8, 32.8-mm beam collimation
Filtered back projection 1.00 with flying focal spot, 5-mm slice thickness, and
reconstruction with 34-cm FOV using filtered back
30% adaptive statistical iterative reconstruction 0.84
projection and B31f filter). Error bars represent 1 SD.

596 AJR:200, March 2013

 Noise and Image Quality Indicators in CT Using Tube Current Modulation

A B C
60 35 Coefficient 18
Coefficient of Coefficient Difference = 5%
of variation = 40% of
variation = 55%
30 variation = 49% 16
Downloaded from www.ajronline.org by Duke University Medical Center on 03/03/13 from IP address 152.16.51.172. Copyright ARRS. For personal use only; all rights reserved

50
Coefficient 14
of 25 Difference = 1%
40 variation = 55% 12
Noise (HU)

Noise (HU)

Noise (HU)
Coefficient
of 20 10
30 variation = 55% Coefficient
15 of
variation = 53%
8
20 6
10
4
10 5
2
0 0 0
75 150 300 75 150 300 75 150
Q (mAs) Q (mAs) Q (mAs)
Slice thickness = 0.6 mm Slice thickness = 1.0 mm kVp = 80 kVp = 100 Pitch = 0.8 Pitch = 1.4

Slice thickness = 3.0 mm Slice thickness = 5.0 mm kVp = 120 kVp = 140

D E F
18 7 Difference = 7% 16 Difference = 59%
Difference < 1%
16 6 14
14 Difference < 1% 12 Difference = 57%
Difference = 2%
5
12
Noise (HU)*
Noise (HU)

Noise (HU)
10
10 4 Difference = 2% Difference = 58%
Difference = 4% 8
8 3
6
6
2
4 4
2 1 2
0 0 0
75 150 300 30 60 120 75 150 300
Q (mAs) Q (mAs) Q (mAs)
FOV = 34 cm FOV = 40 cm 128 × 0.6 32 × 1.2 Adult One-Year-Old

G H
8 18
Difference = 37% Difference = 35%
7 16

6 14
12 Difference = 36%
Noise (HU)*

Difference = 33%
Noise (HU)

5
10 Fig. 6—Noise in mediastinum against image quality
Difference = 36%
4 Difference = 30% reference effective tube current-time product (Q).
8 A–H, Bar graphs show effect of slice thickness (a),
3
6 peak kilovoltage (b), pitch (c), FOV (d), detector
2 4 configuration (e), phantom size (f), patient age
1 selection (g), and reconstruction algorithm (H)
2
on relationship between noise and Q. Each plot
0 0 represents alteration with respect to reference
75 150 300 75 150 300
technique by one parameter. Reference parameters
Q (mAs) Q (mAs)
are underlined in key of each plot. Asterisks indicate
Adult selected Pediatric selected Filtered back projection that these noise values were measured on 1-year-old
Iterative reconstruction in image space phantom. Error bars represent 1 SD.

scanners from multiple manufacturers. Us-
ing Equations 1–4, protocols could be written
to achieve similar noise levels from different
scanners for patients who are similar in size to
the phantoms used in this study.
Results on the Discovery CT 750 HD are
mostly consistent with the design of SmartmA
modulation [7]—that is, the system was able
to modulate the tube current to achieve im-
ages of similar noise for different slice thick-
nesses, pitches, and phantom sizes. Changing
peak kilovoltage, however, did slightly affect
the noise-NI relationship. The peak kilovolt-
age dependence of the noise-NI relationship
is not consistent with the design of SmartmA
and will be discussed further later in this arti-
cle. Measurable noise in different ROIs can be
predicted from the selection of NI. One caveat
is that the SmartmA algorithm will modulate
the tube current according to the prospective-
ly chosen reconstructed slice thickness. This
implies that changing the slice thickness pro-
spectively could have a large effect on patient
exposure. A retrospective reconstruction for a
different slice thickness will result in a change
in noise magnitude as predicted by known
physics principles (i.e., the noise is roughly in-
versely proportional to the square root of slice
thickness) [12].
As mentioned already, on the Discovery
CT 750 HD, changing the peak kilovoltage
did affect the noise-NI relationship, but as
Table 2 suggests, that effect was small com-
pared with other factors. One must keep in
mind that noise magnitude is only one of sev-
eral metrics of image quality. This means that
two images with identical noise magnitude
acquired at different peak kilovoltages do not
necessarily have the same image quality. For

AJR:200, March 2013 597


40 y = 0.7322x
R2 = 0.99296
35
30
Downloaded from www.ajronline.org by Duke University Medical Center on 03/03/13 from IP address 152.16.51.172. Copyright ARRS. For personal use only; all rights reserved
Lung Noise (HU)
25
20
15
10
5 10
0 0
0 10 20 30 40
A Mediastinum Noise (HU)
Fig. 7—Mediastinum noise compared with anatomic noise for all acquisitions on Somatom Definition Flash
scanner (Siemens Healthcare).
A and B, Plots show lung (A) and abdomen (B) noise. Each data point represents different image series, but for
given data point, noise values being compared were measured from same image series. Plots provide way to
translate mediastinum noise to lung and abdomen noise.
example, lowering the peak kilovoltage can
increase contrast [13] but can also introduce
streak artifact.
The results show that decreasing the beam
collimation width has little effect on the
noise-NI relationship. However, changing the
beam width has other implications on image
quality. Total scan time must be longer for a
narrower beam to achieve the same scan cov-
erage. This has the potential to increase mo-
tion artifacts. Also, a narrower beam is gener-
ally associated with lower dose efficiency as
a result of increased overbeaming and over-
scanning [14–16]. The use of a wider beam
will result in shorter scan times but will also
introduce more x-ray scatter. Increased scat-
ter generally degrades image quality [17].
It is important to note that neither the
SmartmA nor CARE Dose4D algorithm al-
tered the tube current according to the pro-
spectively chosen convolution kernel. There-
fore, the choice of convolution kernel had a
large impact on noise (Fig. 3G). On the Dis-
covery CT 750 HD, the bone kernel had the
largest impact. The convolution kernel not
only affects the noise magnitude but also af-
fects noise texture as described by the noise
power spectrum (NPS). The NPS encapsulates
both the variance and the spatial frequency
content of noise. Boedeker and McNitt-Gray
[18] have shown that images with the same
noise magnitude but different NPSs appear
to have a different noise texture. Noise power
concentrated at low frequencies is representa-
tive of images with coarse graininess, whereas
that at high frequencies implies a finer graini-
ness. The texture of the CT image may affect
detectability of certain pathologic abnormali-
ties. Thus, maintaining noise magnitude levels
598
Solomon et al.
70 y = 1.2362x
R2 = 0.99508
60
Abdomen Noise (HU)
50
40
30
20
50 0 10 20 30 40
B Mediastinum Noise (HU)
across various convolution kernels does not
guarantee constant image quality. This is also
true in comparing images across vendors be-
cause each vendor uses different convolution
kernels. Current work is being done to quanti-
tatively compare noise texture across scanners
from different manufacturers; however, such a
comparison is beyond the scope of this study.
Iterative reconstruction algorithms, such
as adaptive statistical iterative reconstruction
and iterative reconstruction in image space,
have been shown to reduce noise for the same
dose compared with filtered back projection
[19–22]. The results of this study (Figs. 3H
and 6H) are consistent with those of other
studies. This result was expected given that
neither SmartmA nor CARE Dose4D alters the
tube current according to use of iterative re-
construction. It should be noted that the adap-
tive statistical iterative reconstruction images
analyzed were 30% blended images (i.e., 30%
adaptive statistical iterative reconstruction and
70% filtered back projection). Further noise re-
duction would be expected for a higher percent-
age, but a recent observer study by Miéville et
al. [19] suggests that, if adaptive statistical it-
erative reconstruction is used above 50%, sub-
jective image quality can actually decrease.
The results from the Somatom Definition
Flash are consistent with the design of CARE
Dose4D modulation [11], and Equation 2 is
consistent with known physics of CT—name-
ly, noise should decrease with increasing tube
current and peak kilovoltage. In contrast to
the SmartmA algorithm, CARE Dose4D does
not alter the tube current modulation profile
according to the prospective reconstructed
slice thickness. As such, changing slice thick-
ness affects expected noise and, therefore, the
noise-Q relationship. For this reason, Equation
1 (representing SmartmA) has no dependence
on slice thickness, whereas Equation 3 (rep-
resenting CARE Dose4D) does. The effect of
changing slice thickness and peak kilovoltage
is accounted for in Equation 2.
Our study shows that, for the same beam
width, double sampling (made possible by
use of a z-axis flying focal spot) had little ef-
fect on the noise-Q relationship (Fig. 6E).
Double sampling is analogous to halving the
50 width of each detector row. This means that
each projection measurement is noisier be-
cause there is half as much total x-ray fluence
per projection, but, because there are twice as
many projections in a fixed length along the
z-axis, the noisiness of each projection can be
averaged out with slices that are thicker than
the detectors’ widths.
Our study also shows that, on the Somatom
Definition Flash, images of the 1-year-old
phantom had less noise compared with images
of the adult phantom acquired with identical
scan settings (Fig. 6F). Thus, the noise-Q re-
lationship is patient size dependent. Also, our
results show that, if one scans the same phan-
tom but changes the selected patient age from
adult to pediatric, the modulation algorithm
will produce images with less noise (Fig. 6G).
Thus, the noise-Q relationship is also depen-
dent on which relative patient age (adult or
pediatric) is selected by the technologist. The
size and patient age dependencies of the noise-
Q relationship were expected, given the design
of CARE Dose 4D, and highlight an impor-
tant distinction between SmartmA and CARE
Dose 4D. As described already, SmartmA at-
tempts to produce images with the same noise
levels irrespective of patient size or shape,
whereas CARE Dose 4D attempts to maintain
constant diagnostic image quality across pa-
tient sizes. Here, diagnostic image quality is
based on clinical assessment. As noted by oth-
ers, this system operates on the principle that
different-sized patients require different levels
of noise to maintain adequate image quality
[23]. The result of such an approach is that, for
identical Q, the system will produce images
with less noise for smaller patients and more
noise for larger patients. This is the reason that
a phantom size adjustment factor was needed
in Equation 2. Such an approach to ATCM
could be justified because images of pediat-
ric patients are more prone to breathing arti-
fact and often lack soft-tissue contrast. Thus,
less noise can be tolerated in pediatric images.
The difference in noise between the medi-
astinum, lung, and abdomen that was observed
AJR:200, March 2013
 Noise and Image Quality Indicators in CT Using Tube Current Modulation
TABLE 3: Fitting Constants and Adjustment Factors Needed to Calculate
Expected Noise for a Given Protocol Using Equations 2 and 4
Protocol
Downloaded from www.ajronline.org by Duke University Medical Center on 03/03/13 from IP address 152.16.51.172. Copyright ARRS. For personal use only; all rights reserved
Anatomic area of interest
Mediastinum
Lung
Abdomen
Phantom
Adult
1-year-old child
Patient age
Adult
Pediatric
Reconstruction algorithm
Filtered back projection
Iterative reconstruction in image space
Fitting constant
α 12.52
β 0.50
γ 1.44
on both systems is not by the design of either
SmartmA or CARE Dose4D but is due to the
difference in average attenuation path length
through these tissues. In the back projection
reconstruction process (ignoring effects of fil-
tering), the variance of a pixel in the recon-
structed image is proportional to the sum of
variances of all the projections through that
pixel [4]. The variance of a single projection is
proportional to the attenuation path length of
that projection. Thus, on average, projections
through pixels in the lungs have shorter atten-
uation path lengths and those in the abdomen
have longer attenuation paths when compared
with the attenuation path lengths of projec-
tions through pixels in the mediastinum. This
leads to images with nonuniform noise. When
using Equation 1 or Equation 2 to calculate ex-
pected noise, the anatomy of interest would be
an important consideration.
This study was limited in two ways with
respect to available resources. First, only
two anthropomorphic phantoms of drasti-
cally different size were available. Further
data are needed to characterize the response
of the CARE Dose 4D algorithm to patients
of many different sizes. Second, only two
CT vendors were considered. Because im-
plementation of ATCM varies greatly across
vendors, the results reported in this study can
be used to predict noise only if SmartmA or
CARE Dose 4D is used. Another limitation
AJR:200, March 2013
Adjustment Factor
1.00
0.73
1.24
1.00
0.42
1.00
0.69
1.00
0.60
is that not every possible factor that could af-
fect noise was considered. For example, only
one convolution kernel was considered for
the Somatom Definition Flash. Thus, there
is no adjustment factor to account for chang-
ing convolution kernels on Table 3. Future
study will include adjustment factors for the
many different convolution kernels offered
by Siemens Healthcare. Also, modulation
strength is a CARE Dose 4D parameter that
determines the lookup table used to relate ef-
fective tube current to attenuation. This pa-
rameter can be selected as weak, average, or
strong and controls the degree to which the
tube current is adjusted for patient size [23].
The effect of changing modulation strength
on image quality has been previously con-
sidered by other researchers but was not ex-
plored in this study [24].
In conclusion, for the SmartmA tube cur-
rent modulation algorithm on the Discovery
CT 750 HD, expected noise is a linear func-
tion of NI. The slope of this function can be
adjusted to account for differing anatomy,
peak kilovoltage, convolution kernels, and
reconstruction algorithms. This function is
not affected by phantom size, slice thickness,
collimated beam width, pitch, or FOV. For
the CARE Dose 4D tube current modulation
algorithm on the Somatom Definition Flash,
expected noise can be modeled as a power
function of Q, slice thickness, and peak kilo-
voltage. This function can be further adjust-
ed to account for differing anatomy, phantom
size, patient age selection, and reconstruc-
tion algorithm. This function is independent
of pitch, FOV, and detector configuration set-
tings. This work has potential to aid in op-
timizing protocols that employ ATCM and
to aid in achieving consistent image quality
across different scanners.
References
1. Brenner DJ, Hall EJ. Computed tomography: an
increasing source of radiation exposure. N Engl J
Med 2007; 357:2277–2284
2. Schauer DA, Linton OW. NCRP Report No. 160,
ionizing radiation exposure of the population of
the United States, medical exposure: are we doing
less with more, and is there a role for health phys-
icists? Health Phys 2009; 97:1–5
3. Lee TY, Chhem RK. Impact of new technologies
on dose reduction in CT. Eur J Radiol 2010;
76:28–35
4. Gies M, Kalender WA, Wolf H, Suess C. Dose re-
duction in CT by anatomically adapted tube cur-
rent modulation. Part I. Simulation studies. Med
Phys 1999; 26:2235–2247
5. Kalender WA, Wolf H, Suess C. Dose reduction in
CT by anatomically adapted tube current modula-
tion. Part II. Phantom measurements. Med Phys
1999; 26:2248–2253
6. Greess H, Nomayr A, Wolf H, et al. Dose reduc-
tion in CT examination of children by an attenua-
tion-based on-line modulation of tube current
(CARE Dose). Eur Radiol 2002; 12:1571–1576
7. Kalra MK, Maher MM, Toth TL, et al. Techniques
and applications of automatic tube current modula-
tion for CT. Radiology 2004; 233:649–657
8. Li X, Samei E, Barnhart HX, et al. Lung nodule
detection in pediatric chest CT: quantitative rela-
tionship between image quality and radiologist
performance. Med Phys 2011; 38:2609–2618
9. Kalra MK, Naz N, Rizzo SM, Blake MA. Com-
puted tomography radiation dose optimization:
scanning protocols and clinical applications of
automatic exposure control. Curr Probl Diagn
Radiol 2005; 34:171–181
10. Mulkens TH, Bellinck P, Baeyaert M, et al. Use of
an automatic exposure control mechanism for
dose optimization in multi-detector row CT ex-
aminations: clinical evaluation. Radiology 2005;
237:213–223
11. Flohr T. CARE Dose4D: white paper. Siemens
Healthcare website. www.medical.siemens.com/
siemens/en_US/gg_ct_FBAs/files/Case_Studies/
A9115-111236_Care_Dose_4D.pdf. Published 2011.
Accessed November 2, 2012
12. Bushberg JT, Seibert JA, Leidholdt EM, Boone
JM. The essential physics of medical imaging,
599

2nd ed. Philadelphia, PA: Lippincott Williams &
Wilkins, 2001:367
13. Bogot NR, Fingerle A, Shaham D, Nissenbaum I,
Sosna J. Image quality of low-energy pulmonary
Downloaded from www.ajronline.org by Duke University Medical Center on 03/03/13 from IP address 152.16.51.172. Copyright ARRS. For personal use only; all rights reserved
CT angiography: comparison with standard CT.
AJR 2011; 197:384; [web]W273–W278
14. Flohr TG, Schaller S, Stierstorfer K, Bruder H,
Ohnesorge BM, Schoepf UJ. Multi-detector row
CT systems and image-reconstruction techniques.
Radiology 2005; 235:756–773
15. Nicholson R, Fetherston S. Primary radiation out-
side the imaged volume of a multislice helical CT
scan. Br J Radiol 2002; 75:518–522
16. Niu YT, Olszewski ME, Zhang YX, Liu YF, Xian
JF, Wang ZC. Experimental study and optimiza-
tion of scan parameters that influence radiation
600
Solomon et al.
dose in temporal bone high-resolution multidetec-
tor row CT. AJNR 2011; 32:1783–1788
17. Glover GH. Compton scatter effects in CT recon-
structions. Med Phys 1982; 9:860–867
18. Boedeker KL, McNitt-Gray MF. Application of
the noise power spectrum in modern diagnostic
MDCT. Part II. Noise power spectra and signal to
noise. Phys Med Biol 2007; 52:4047–4061
19. Miéville FA, Gudinchet F, Rizzo E, et al. Paediat-
ric cardiac CT examinations: impact of the itera-
tive reconstruction method ASIR on image quali-
ty—preliminary findings. Pediatr Radiol 2011;
41:1154–1164
20. Silva AC, Lawder HJ, Hara A, Kujak J, Pavlicek
W. Innovations in CT dose reduction strategy: ap-
plication of the adaptive statistical iterative recon-
struction algorithm. AJR 2010; 194:191–199
21. Leipsic J, Labounty TM, Heilbron B, et al. Adap-
tive statistical iterative reconstruction: assessment
of image noise and image quality in coronary CT
angiography. AJR 2010; 195:649–654
22. Ghetti C, Ortenzia O, Serreli G. CT iterative re-
construction in image space: a phantom study.
Phys Med 2012; 28:161–165
23. Lee CH, Goo JM, Ye HJ, et al. Radiation dose
modulation techniques in the multidetector CT
era: from basics to practice. RadioGraphics 2008;
28:1451–1459
24. Söderberg M, Gunnarsson M. The effect of differ-
ent adaptation strengths on image quality and ra-
diation dose using Siemens Care Dose 4D. Radiat
Prot Dosimetry 2010; 139:173–179
AJR:200, March 2013