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

Christianson et al.
Noise Measurement in Clinical CT

Medical Physics and Informatics

Original Research
Downloaded from www.ajronline.org by 200.28.58.4 on 05/09/22 from IP address 200.28.58.4. Copyright ARRS. For personal use only; all rights reserved

Automated Technique to Measure

Noise in Clinical CT Examinations
Olav Christianson1 OBJECTIVE. The purpose of this study was to develop and validate an automated meth-
James Winslow 1 od to measure noise in clinical CT examinations.
Donald P. Frush2 MATERIALS AND METHODS. An automated algorithm was developed to measure
Ehsan Samei1 noise in CT images. To assess its validity, the global noise level was compared with image
noise measured using an image subtraction technique in an anthropomorphic phantom. The
Christianson O, Winslow J, Frush DP, Samei E global noise level was further compared with image noise values from clinical patient CT im-
ages obtained by an observer study. Finally, the clinical utility of the global noise level was
shown by assessing variability of image noise across scanner models for abdominopelvic CT
examinations performed in 2358 patients.
RESULTS. The global noise level agreed well with the phantom-based and clinical image–
based noise measurements, with an average difference of 3.4% and 4.7% from each of these
measures, respectively. No significant difference was detected between the global noise level
and the validation dataset in either case. It further indicated differences across scanners, with
the median global noise level varying significantly between different scanner models (15–35%).
CONCLUSION. The global noise level provides an accurate, robust, and automated
method to measure CT noise in clinical examinations for quality assurance programs. The
significant difference in noise across scanner models indicates the unexploited potential to
efficiently assess and subsequently improve protocol consistency. Combined with other auto-
mated characterization of imaging performance (e.g., dose monitoring), the global noise lev-
el may offer a promising platform for the standardization and optimization of CT protocols.

Keywords: CT, dose monitoring, image quality, protocol
optimization, quality assurance
DOI:10.2214/AJR.14.13613
Received August 6, 2014; accepted after revision
December 31, 2014.
1
Ravin Advanced Imaging Laboratories, Duke University,
2424 Erwin Rd, Ste 302, Durham, NC 27705. Address
correspondence to O. Christianson
(olav.christianson@gmail.com).
2
Department of Radiology, Duke University Medical
Center, Durham, NC.
WEB
This is a web exclusive article.
AJR 2015; 205:W93–W99
0361–803X/15/2051–W93
© American Roentgen Ray Society
M
ore than 70 million CT examina-
tions are performed annually in
the United States, and CT now
accounts for approximately 25%
of the annual radiation exposure to the U.S.
population [1, 2]. Accordingly, recent studies
have suggested that radiation exposure from
CT examinations may be associated with an
increased risk of developing cancer [3, 4].
Therefore, there is a growing emphasis placed
on using the minimum radiation dose neces-
sary. However, using a relatively low radiation
dose may have a detrimental impact on diag-
nostic accuracy. Therefore, research and clini-
cal efforts are essential to optimize CT proto-
cols to deliver the minimum radiation dose
while maintaining diagnostic image quality.
In a clinical environment, it is a daunt-
ing task to ensure that the multitude of ever-
changing CT protocols is consistent across
a fleet of CT scanners to maintain a proper
balance of radiation dose and image quali-
ty. To this end, tracking radiation dose and
comparing it across institutions and scanner
models is becoming routine clinical practice.
For example, The Joint Commission will re-
quire recording the radiation dose for every
CT study to maintain accreditation effec-
tive July 1, 2015 [5]. Although radiation dose
tracking is a valuable tool for identifying ra-
diation overdoses and evaluating operation-
al consistency, this monitoring provides no
direct information on achieving the desired
image quality [6]. Importantly, because im-
age quality is affected by several reconstruc-
tion parameters (e.g., kernel, slice thickness,
iterative reconstruction, and display FOV)
that are not factors in the delivered radiation
dose, radiation dose tracking alone is insuffi-
cient to determine whether CT protocols are
consistent or optimized [7].
An ideal approach to optimize and moni-
tor CT protocols is to directly evaluate the
level of image quality in clinical images.
However, current methods to evaluate image
quality in patient CT images (e.g., assessing

AJR:205, July 2015 W93

 Christianson et al.

600

500

400

Frequency
300
Downloaded from www.ajronline.org by 200.28.58.4 on 05/09/22 from IP address 200.28.58.4. Copyright ARRS. For personal use only; all rights reserved

200

100

0
0 20 40 60 80 100
SD (HU)

A B C
Fig. 1—Noise detection process.
A, Original CT image of patient.
B, Noise map shows soft-tissue regions.
C, Histogram shows noise map for soft-tissue regions. Mode of histogram corresponds to global noise level.

noise, contrast, or the ratio of the two) are
highly labor intensive. For example, methods
for noise assessment rely on manual selection
of ROIs in uniform areas to measure image
noise [8], making them impractical for large-
scale quality assurance programs. A more re-
cently reported automated platform is limit-
ed by its reliance on phantom measurements
and the complexity associated with image
quality in nonuniform patient images [9, 10].
Automated approaches to assess image qual-
ity in chest radiography have been notewor-
thy [11]. However, such methods have not
been implemented for CT.
The objective of this study was to take the
first step toward automated assessment of im-
age quality in clinical CT images. More spe-
cifically, the aims of this work were to develop
an automated computationally efficient meth-
od for measuring the global noise level from
patient CT images, to validate this method
in a complex abdominal-thoracic surrogate
phantom as well as in clinical patient images,
and to show how this new metric can be used
as a tool to evaluate the consistency of image
noise across multiple CT scanner models.
Materials and Methods
In accordance with our institutional policy, this
quality improvement activity did not constitute
human subject research because it pertained to
a clinical quality control initiative. All data were
anonymized, and the project was in compliance
with HIPAA guidelines.
The Noise Detection Algorithm
We developed an automated method to calcu-
late noise in patient CT images. Because CT noise
varies with location in the image, noise was char-
acterized in terms of the most frequent noise level
in areas of homogeneous tissue. The metric was
termed the “global noise level.”
The global noise level was calculated in three
steps. First, the image was segmented into dif-
ferent tissue types (Fig. 1). The tissue types were
roughly separated into four categories: aerated tis-
sue (HU< −800), fat (−300 to < 0 HU), soft-tissue
(0 < to 100 HU), and bone (> 300 HU). For the
purposes of this study, we limited our analysis to
the soft-tissue components of the image.
Second, a noise map was generated in which the
value at each location corresponded to the SD in an

A B
Fig. 2—Meleagris gallopavo phantom.
A, Representative CT image of phantom shows homogeneous areas as well as regions containing bone, air cavities, and fluid recesses. Boxes indicate ROIs used for
noise measurement in subtracted image.
B, CT image resulting from subtraction of repeated series. Boxes indicate ROIs used for noise measurement in subtracted image.

W94 AJR:205, July 2015


TABLE 1: Image Acquisition Parameters for Cadaver Phantom
Parameter
Dose levels (% of default clinical setting)
Slice thicknesses (mm)
Reconstruction kernels
Downloaded from www.ajronline.org by 200.28.58.4 on 05/09/22 from IP address 200.28.58.4. Copyright ARRS. For personal use only; all rights reserved
aUses SAFIRE (Sinogram Affirmed Iterative Reconstruction) 3 (Siemens Healthcare).
area (i.e., kernel) surrounding that pixel. A design
consideration was the kernel size. Kernels that are too
small may not fully sample the low-frequency com-
ponents of the noise, whereas those that are too large
may be overly contaminated with anatomic transi-
tions. Multiple kernel sizes from 2 to 30 mm were
investigated across abdominal and thoracic CT pa-
tient images to determine which would give the most
robust and accurate results that were relatively in-
sensitive to the size. A convolution method was then
used to calculate the SD for each kernel. The aver-
age computing time for this processing on a standard
commercially available laptop computer (ThinkPad
T420, Lenovo) was 0.05 seconds per image.
Third, a histogram was generated for the SDs
corresponding to the targeted tissue type (soft tis-
sue). The histogram comprised two principal cat-
egories: homogeneous areas and transitional areas.
Transitional areas, such as the interface between
soft-tissue and bone, constituted a surprisingly small
portion of the image. Furthermore, because they
were at the boundary of multiple tissue types, they
exhibited higher and more variable ranges of SDs
compared with those of homogeneous areas. As a
result, the histogram had a characteristic high nar-
row peak corresponding to homogeneous areas and
a short broad tail corresponding to transitional ar-
eas. The global noise level was determined by iden-
tifying the mode of the histogram peak correspond-
ing to homogeneous tissue.
Validation: Phantom Images
A phantom study was conducted to perform the
initial validation of the global noise level in repre-
senting image noise. Although several anthropo-
morphic phantoms were commercially available,
none possessed the same level of inhomogeneity
present in patient images. To better evaluate the
global noise level algorithm, a test object more
similar in complexity to human abdominal and
TABLE 2: Abdominopelvic Protocols
Reconstruction Slice
Scanner Model Name kVp Thickness (mm)
LightSpeed VCT 120
Discovery CT 750 HD 120
Somatom Definition Flash 120
Note—FBP = filtered back projection, ASIR = adaptive statistical iterative reconstruction. LightSpeed VCT and Discovery CT 750 HD manufactured by GE Healthcare,
Somatom Definition Flash manufactured by Siemens Healthcare. SAFIRE (Sinogram Affirmed Iterative Reconstruction) 3 is a product of Siemens Healthcare.
AJR:205, July 2015 W95
Noise Measurement in Clinical CT
Settings
100, 33, and 17
5 and 1
B31f, B35f, B71f, I31fa, I36fa, I75fa
thoracic cavities was needed. Therefore, a cadaver
of Meleagris gallopavo (turkey) was selected for
the validation (Fig. 2). This test object combined
homogeneous tissue regions with inhomogeneous
areas, air cavities, bones, and fluid recesses, mak-
ing it an effective platform on which to validate
the global noise level algorithm.
The phantom was scanned on a commercial
CT scanner (Somatom Definition Flash, Siemens
Healthcare) using two slice thicknesses, three dose
levels, and six reconstruction kernels (Table 1). For
each set of scanning parameters, the phantom was
scanned twice; the repeated image series were sub-
tracted to remove the anatomic structures and iso-
late the quantum noise in the image. Six ROIs were
uniformly distributed throughout the soft tissue in
a representative slice of the phantom (Fig. 2). The
noise was averaged across the six ROIs and com-
pared with the global noise level. The Spearman
rank correlation coefficient was calculated between
the two datasets and used to evaluate the differ­
ences between the two datasets, with a p value less
than 0.05 indicating a significant difference.
Validation: Patient Images
An observer study was performed to further
validate that the global noise level is indicative
of image noise in patient images. Six clinical pa-
tient images, three abdominal and three thoracic,
were randomly selected from our clinical dataset.
All six images were acquired on a commercial CT
scanner (CT 750 HD, GE Healthcare) using clin-
ical abdominal and chest imaging protocols. All
images were acquired using the standard clinical
protocol; the abdominal images were contrast en-
hanced. Four physicists took part in the observer
study. Previous studies have shown that observ-
ers can accurately measure noise in patient im-
ages [8]. Each observer was instructed to place
five ROIs in areas of soft tissue in a way that these
Noise Index or Quality
Pitch Reference mAs
5 1.375 15
5 1.375 22
5 0.8 200
ROIs did not overlap with any anatomic structure.
The ROIs were made as large as possible without
including adjacent anatomic structures and were
spaced as evenly as possible throughout the pa-
tient. The noise in the image was determined by
computing the average SD across the ROIs. Fi-
nally, image noise measured by the observers was
compared with the global noise level. The Pearson
correlation coefficient was calculated between the
two datasets and used to evaluate the differences
between the two datasets, with a p value less than
0.05 indicating a significant difference.
Application: Variability in Noise Across
Scanner Models
To illustrate the clinical utility of the global
noise level, the global noise level was calculated
for 2358 patients imaged on three commercial CT
scanner models (Discovery CT 750HD [n = 512]
and LightSpeed VCT [n = 1644], both GE Health-
care and Somatom Definition Flash [n = 202], Sie-
mens Healthcare) using the abdominopelvic pro-
tocol (Table 2). To minimize processing time, 10
evenly spaced images were selected from each
study for analysis. The global noise level was aver-
aged across those 10 images to provide a represen-
tative noise level for the examination as a whole.
For comparison purposes, the patient effective di-
ameters and size-specific dose estimates (SSDEs)
were determined using a previously published
dose monitoring program [6]. ANOVA was used
to determine the significance of the difference in
effective diameter, SSDE, and global noise level
between scanner models (p < 0.05 indicated sig-
nificance). Because of the large number of patient
cases, even very small differences across scanner
models may appear significant. Therefore, in ad-
dition to significance testing, the Cohen f statistic
was used to investigate the effect size across scan-
ner models for effective diameter, SSDE, and glob-
al noise level with the following evaluation criteria:
f > 0.10 is a small effect, > 0.25 is a medium effect,
and > 0.40 is a large effect [12].
Results
Stability of the Global Noise Level With
Kernel Size
As noted, first the sensitivity of the global
noise level to kernel size was investigated by
Gantry Rotation Reconstruction
Time (s) Kernel
0.5 FBP, standard
0.5 ASIR 40%, standard
0.5 SAFIRE 3, I31f
 Christianson et al.

tively. Similarly, changing the slice thickness

70 70
from 5 to 1 mm was expected to increase the
60 60 noise by the square root of five (123% in-
Global Noise Level (HU)

Global Noise Level (HU)

crease in noise). The image noise and global
50 50
noise level, however, only increased by 97%
40 40 and 84%, respectively.
Downloaded from www.ajronline.org by 200.28.58.4 on 05/09/22 from IP address 200.28.58.4. Copyright ARRS. For personal use only; all rights reserved

30
20
10
0 0
0 10 20
Kernel Size (mm)
Fig. 3—Sensitivity of global noise level to ROI size.
A and B, Graphs show how global noise level varies with ROI size for abdominal image (A) and thoracic image
(B). In both cases, there is stable region from around 6–20 mm.
applying the noise detection algorithm to ab-
dominal and thoracic patient CT images using
a range of ROI sizes. In general, the measured
noise level increased with kernel size up to 6
mm (Fig. 3). From 6 to 20 mm, the measured
noise level was relatively constant. Beyond
20 mm, the measured noise level increased
abruptly as the number of ROIs composed
completely of homogeneous tissue dropped to
a critically low level. Beyond this threshold,
the measured noise levels were corrupted with
the inability of large kernels to avoid anatom-
ic transitions. On the basis of these findings, a
6-mm kernel size was determined as ideal for
the global noise level measurement.
Validation: Phantom Images
The global noise level was compared with
image noise measured using subtraction
techniques in a phantom consisting of hu-
30
20
10
30 0 10 20 30
Kernel Size (mm)
A B
manlike tissues and organs (Fig. 4). Across
all image acquisition parameters, the average
absolute difference between the global noise
level and image noise was 3.4%. The larg-
est difference between the two measures was
12.0%, which occurred at the highest dose
level using iterative reconstruction resulting
in the image with the lowest overall noise
magnitude. In this case, the global noise lev-
el differed from the image noise by only 0.7
HU. The correlation between the two data­
sets was significant (ρ = 0.9991 and p < 0.01).
Although the global noise level and image
noise agreed well, they did not always match
the expected noise based on slice thickness
and dose level. For example, decreasing the
dose to 50% was expected to increase the
noise by the square root of two (41% increase
in noise). The global noise level and image
noise increased by 36% and 34%, respec-
Validation: Patient Images
The global noise level was compared with
image noise determined from an observer
study in six clinical patient images (Fig. 5).
Across the six images, the average difference
between the global noise level and observ-
er noise measurement was 4.7%, which was
less than the intraobserver variability of 6.2%.
The agreement between the global noise level
and observer measured noise was slightly bet-
ter in the abdominal images than in the tho-
racic images (3.9% and 5.4% for the abdom-
inal and thoracic images, respectively). The
correlation between the two datasets was sig-
nificant (ρ = 0.9979 and p = 0.001).
Application: Variability in Noise Across
Scanner Models
Effective diameter, SSDE and global noise
level were determined for 2358 patients im-
aged on three CT scanner models using the
abdominopelvic protocol (Fig. 6). The median
effective diameter differed by 2–8% between
the scanner models. Although the difference
in effective diameter across scanner models
was significant (p < 0.001), the effect size was
small (f = 0.129) indicating minimal practi-
cal significance in terms of patient diameters
across scanner models. The median SSDE
differed by 9–33% between scanner models.
The difference in SSDE across scanner mod-
els was significant (p = < 0.001) and the effect

90 20 16
80 18 14
70 16
12
60 14
Noise (HU)

10
Noise (HU)

Noise (HU)

12
50
10 8
40
8
30 6
6
20 4
4
10 2 2
0 0 0
B31f B35f B75f l31f l36f l70f 100%
50% 17% 5 mm 1 mm
Reconstruction Kernel Dose Level Slice Thickness
Global Noise Level Subtraction Method Global Noise Level Subtraction Method Global Noise Level Subtraction Method
A B C
Fig. 4—Phantom validation.
A–C, Graphs show comparison of global noise level with noise determined from subtracted images across reconstruction kernel (A), dose level (B), and slice thickness (C).
(Fig. 4 continues on next page)

W96 AJR:205, July 2015

 Noise Measurement in Clinical CT

100
Fig. 4 (continued)— ity control program. Finally, such a quality
Phantom validation. control program facilitates characterization
D, Graph shows global
noise level versus of consistency across multiple platforms, op-
noise measured from erators, shifts, and institutions.
80 subtracted images for all Despite the clinical utility of the global
scanning combinations.
Each point corresponds
noise level and its strong agreement with noise
to noise measured from measured in phantom and clinical patient im-
Global Noise Level (HU)
Downloaded from www.ajronline.org by 200.28.58.4 on 05/09/22 from IP address 200.28.58.4. Copyright ARRS. For personal use only; all rights reserved

60
40
20
0
0 20
Subtraction Method (HU)
size was moderate (f = 0.265), reflective of ac-
tual differences between the dosing protocols
across scanners. The median global noise lev-
el differed by 15–35% between scanner mod-
els. The difference in global noise level across
scanner models was significant (p = < 0.001)
and the effect size was large (f = 0.463), indi-
cating a high degree of practical significance
due to differences in both the imaging proto-
cols and system attributes.
Discussion
We developed an automated computa-
tionally efficient method to measure the
global noise level from patient CT images.
This method segments the image into tissue
types, generates an SD map, and uses histo-
gram analysis to determine the global noise
level in homogeneous regions of soft tissue.
We found that the global noise level mea-
sured using this method agreed to within 5%
unique set of scanning
parameters and solid
line represents identity
line.
40 60 80 100
D representative of the spectrum of patient size
new paradigm in quality metrology that can
be used toward a variety of goals. First, it can
be used for granular estimation of quality on
an image-by-image basis. For example, this
method enables characterization of image
noise across slices of a CT series. Second,
this approach enables individualized retro-
spective assessment of the diagnostic opera-
tion, whereas other techniques treat all pa-
tients similarly. Third, patient image quality
data facilitate prospective definition of pro-
tocols to achieve the desired diagnostic accu-
racy. Fourth, because image quality and ra-
diation dose are inversely related, measuring
both is necessary for a comprehensive qual-
20
18
16
ages, there are several limitations of this ap-
proach. First, although large-scale anatomic
structures are avoided by the histogram anal-
ysis, small-scale anatomic texture may con-
tribute to the global noise level. The strong
agreement between the global noise level and
the noise measured in the phantom using the
subtraction technique, which should remove
most anatomic contributions, suggests that
the amount of anatomic structure was small.
On the other hand, the amount of anatom-
ic structure likely varies between individu-
als, and the phantom used in this study is not
and tissue distributions seen clinically. Fur-
thermore, the observer study showed a strong
agreement between the global noise level and
observer measurements even in the presence
of contrast enhancement, which should em-
phasize the small-scale anatomic structure.
Conversely, anatomic structure contamina-
tion at low noise levels could explain the find-
ing that the global noise level did not increase
as expected with changes in slice thickness
and dose levels. More work is needed to de-
termine the full impact of small-scale ana-
tomic structures.
Second, this technique measures noise
magnitude but ignores the spatial correla-
tions of the noise. In general, low-frequen-

of noise measurements made in a complex 14

Global Noise Level (HU)

anthropomorphic phantom as well as those

from observer assessments in clinical CT ex- 12
aminations. Further, we implemented this
10
algorithm in routine quality control practic- Fig. 5—Observer
es. Sampling 10 images from each series for validation. On graph
of global noise level 8
global noise level measurement kept process- versus noise determined
ing time under 1 second for each acquisition, by human observers, 6
enabling this method to be applied to every each point represents
average across all 4
patient undergoing a CT examination. observers for given
A notable advantage of our findings is that image, and error 2
image quality can be evaluated in terms of bars correspond to
the actual attributes of clinical images rather SD across observer 0
measurements. Solid 0 5 10 15 20
than idealized phantoms. Measuring image line represents identity Observer Noise Measurement (HU)
quality using actual clinical images offers a line.

AJR:205, July 2015 W97

 Christianson et al.

45 45
+ +
+
+
+
40 +
+
+
+
40 + +
+
35 +
+
+
+
+ +
+ +
+
Effective Diameter (cm)

30 +
+ +
+
+ +
+
+
Downloaded from www.ajronline.org by 200.28.58.4 on 05/09/22 from IP address 200.28.58.4. Copyright ARRS. For personal use only; all rights reserved

35

SSDE (mGy)
25
+
20 +
30 +
15

10
25

+
+ 5
+ +
20 0
Discovery Lightspeed Somatom Discovery Lightspeed Somatom
CT 750 HD VCT Definition CT 750 HD VCT Definition
Flash Flash

A B
Fig. 6—Variability in surance of resulting clinical images that can
35 + effective diameter, size-
+
+
aid in protocol optimization. The global noise
+
+
specific dose estimate
level also offers a simple retrospective tool to
+
+ (SSDE), and global noise
30 +
+ level across scanner compare image noise in patient images across
+
+ models.
+
+ +
+
+
scanner models. As expected, the SSDE var-
+ +
+ A–C, Box-and-whisker
ied with scanner model and reconstruction al-
25 +
+ + plots show median
+
Global Noise Level (HU)

20
15
10
5 VCT manufactured by
0 Flash manufactured by
Discovery
CT 750 HD
cy noise content, resulting in a coarse lumpy
appearance, is more detrimental to diagnos-
tic performance than high-frequency noise
content, resulting in a fine sandy appearance
[13–17]. More work is needed to develop a
robust technique to quantify the noise tex-
ture present in patient CT images. In the in-
terim, the noise texture measured in a uni-
form background can be combined with the
global noise level to fully characterize the
noise present in the image.
Third, many factors other than noise con-
tribute to image appearance, most notably,
resolution and contrast. It may be possible to
and 25th and 75th gorithm, thereby making the SSDE difficult
percentiles of effective to use as a primary quality control metric. Be-
diameter (A), SSDE (B), cause the noise required for diagnostic image
+
+ and global noise level (C)
for each scanner model. quality depends to a much lesser extent on
Outliers (indicated scanner model and reconstruction algorithm,
with + marker) were the global noise level is better suited for estab-
those that exceeded
lishing diagnostic reference levels [20].
+
+ 1.5 times interquartile
length. Discovery CT The benefits of the global noise level
+ 750 HD and LightSpeed method are exemplified when comparing
GE Healthcare, and
protocols across scanner models. The global
Somatom Definition noise level varied by 15–35% between scan-
ner models in this study, indicating the possi-
Lightspeed Somatom Siemens Healthcare. bility of improving the consistency of image
VCT Definition
Flash quality and reducing patient dose. Assuming
the images from the LightSpeed VCT scan-
C
ner are of diagnostic quality, the noise could
derive some insight into these attributes from be increased on the CT 750 HD and the So-
patient CT images; however, methods to do matom Definition Flash scanners according-
so have yet to be developed. Currently, these ly. Adjusting the protocols to achieve con-
image attributes must be measured in phan- sistent noise across scanner models would
toms and extrapolated to patient images. decrease the SSDE by 27% and 45% on the
They can then be combined with the global CT 750 HD and Somatom Definition Flash
noise level through the use of computational scanners, respectively. This newfound po-
observer models to provide a meaningful de- tential for dose reduction shows the potential
tectability index that is related to the likeli- clinical impact of an automated image qual-
hood of an observer being able to detect the ity monitoring program.
presence of a lesion [18, 19].
Even with these limitations, the benefits of Conclusion
using the global noise level are substantial. This study involved the development of an
This innovative technique permits quality as- automated computationally-efficient method

W98 AJR:205, July 2015


to measure the global noise level for CT im-
ages. The global noise level measured using
this new method agreed well with noise mea-
surements made using an anthropomorphic
phantom containing humanlike tissues and
organs as well as those from patient CT im-
ages. Furthermore, we showed how the glob-
Downloaded from www.ajronline.org by 200.28.58.4 on 05/09/22 from IP address 200.28.58.4. Copyright ARRS. For personal use only; all rights reserved
al noise level can be applied to a large num-
ber of patients to analyze the consistency of
image noise between scanner models. The
automated technique to measure CT noise
in patient images offers a promising new ap-
proach for monitoring image noise on a per-
patient basis, providing a platform to facil-
itate standardization of image quality and
optimization of imaging protocols.
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 do-
ing less with more, and is there a role for health
physicists? Health Phys 2009; 97:1–5
3. Pearce MS, Salotti JA, Little MP, et al. Radiation
exposure from CT scans in childhood and subse-
quent risk of leukaemia and brain tumours: a retro-
spective cohort study. Lancet 2012; 380:499–505
4. Mathews JD, Forsythe AV, Brady Z, et al. Cancer
risk in 680,000 people exposed to computed to-
AJR:205, July 2015 W99
Noise Measurement in Clinical CT
mography scans in childhood or adolescence:
data linkage study of 11 million Australians. BMJ
2013; 346:f2360
5. Joint Commission. New/revised diagnostic imag-
ing standards from Joint Commission. J Nucl Med
2014; 55:13N
6. Christianson O, Li X, Frush D, Samei E. Auto-
mated size-specific CT dose monitoring program:
assessing variability in CT dose. Med Phys 2012;
39:7131–7139
7. Goldman LW. Principles of CT: radiation dose
and image quality. J Nucl Med Technol 2007;
35:213–225
8. Kalra MK, Maher MM, Kamath RS, et al. Six-
teen-detector row CT of abdomen and pelvis:
study for optimization of z-axis modulation tech-
nique performed in 153 patients. Radiology 2004;
233:241–249
9. Larson DB, Malarik RJ, Hall SM, Podberesky DJ.
System for verifiable CT radiation dose optimiza-
tion based on image quality. Part II. Process con-
trol system. Radiology 2013; 269:177–185
10. Solomon JB, Samei E. Are uniform phantoms
sufficient to characterize the performance of itera-
tive reconstruction in CT? (abstr) SPIE website.
proceedings.spiedigitallibrary.org/proceeding.aspx?
articleid=1662019. Published March 13, 2013.
Accessed March 24, 2015
11. Lin Y, Luo H, Dobbins JT, et al. An image-based
technique to assess the perceptual quality of clinical
chest radiographs. Med Phys 2012; 39:7019–7031
12. Kirk RE. Practical significance: a concept whose
time has come. Educ Psychol Meas 1996; 56:746–759
13. Boedeker KL, Cooper VN, McNitt-Gray MF. Ap-
plication of the noise power spectrum in modern
diagnostic MDCT. Part I. Measurement of noise
power spectra and noise equivalent quanta. Phys
Med Biol 2007; 52:4027–4046
14. Chew E, Weiss GH, Brooks RA, Di Chiro G. Ef-
fect of CT noise on detectability of test objects.
AJR 1978; 131:681–685
15. Burgess AE. Statistical efficiency of perceptual
decisions. (abstr) SPIE website. proceedings.
spiedigitallibrary.org/proceeding.aspx?articleid=
1236924, Published June 15, 1984. Accessed
March 24, 2015
16. Hanson KM. Detectability in computed tomo-
graphic images. Med Phys 1979; 6:441–451
17. Solomon JB Bochud F, Samei E. Design of an-
thropomorphic textured phantoms for CT perfor-
mance evaluation. (abstr) spie.org/Publications/
Proceedings/Paper/10.1117/12.2043555. Pub-
lished March 19, 2014. Accessed March 24, 2015
18. Burgess AE, Jacobson FL, Judy PF. Human ob-
server detection experiments with mammograms
and power-law noise. Med Phys 2001; 28:419–437
19. Gang GJ, Lee J, Stayman JW, et al. Analysis of
Fourier-domain task-based detectability index in
tomosynthesis and cone-beam CT in relation to
human observer performance. Med Phys 2011;
38:1754–1768
20. [No authors listed]. 1990 Recommendations of the
International Commission on Radiological Protec-
tion (Report 60). Ann ICRP 1991; 21:1–2001