Beruflich Dokumente
Kultur Dokumente
Girijesh Yadava
GE Healthcare, Waukesha, Wisconsin 53188
Simon N. Murphy
Carl E. Ravin Advanced Imaging Laboratories, Department of Radiology, Duke University,
Durham, North Carolina 27705
Ehsan Samei
Carl E. Ravin Advanced Imaging Laboratories, Department of Radiology, Duke University,
Durham, North Carolina 27705
(Received 21 July 2011; revised 15 May 2012; accepted for publication 16 May 2012; published 11
June 2012)
Purpose: To investigate a measurement method for evaluating the resolution properties of CT imag-
ing systems across reconstruction algorithms, dose, and contrast.
Methods: An algorithm was developed to extract the task-based modulation transfer function (MTF)
from disk images generated from the rod inserts in the ACR phantom (model 464 Gammex, WI).
These inserts are conventionally employed for HU accuracy assessment. The edge of the disk objects
was analyzed to determine the edge-spread function, which was differentiated to yield the line-spread
function and Fourier-transformed to generate the object-specific MTF for task-based assessment, de-
noted MTFTask . The proposed MTF measurement method was validated against the conventional wire
technique and further applied to measure the MTF of CT images reconstructed with an adaptive statis-
tical iterative algorithm (ASIR) and a model-based iterative (MBIR) algorithm. Results were further
compared to the standard filtered back projection (FBP) algorithm. Measurements were performed
and compared across different doses and contrast levels to ascertain the MTFTask dependencies on
those factors.
Results: For the FBP reconstructed images, the MTFTask measured with the inserts were the same as
the MTF measured from the wire-based method. For the ASIR and MBIR data, the MTFTask using the
high contrast insert was similar to the wire-based MTF and equal or superior to that of FBP. However,
results for the MTFTask measured using the low-contrast inserts, the MTFTask for ASIR and MBIR
data was lower than for the FBP, which was constant throughout all measurements. Similarly, as a
function of mA, the MTFTask for ASIR and MBIR varied as a function of noise—with MTFTask being
proportional to mA. Overall greater variability of MTFTask across dose and contrast was observed for
MBIR than for ASIR.
Conclusions: This approach provides a method for assessing the task-based MTF of a CT system
using conventional and iterative reconstructions. Results demonstrated that the object-specific MTF
can vary as a function of dose and contrast. The analysis highlighted the paradigm shift for iterative
reconstructions when compared to FBP, where iterative reconstructions generally offer superior noise
performance but with varying resolution as a function of dose and contrast. The MTFTask generated
by this method is expected to provide a more comprehensive assessment of image resolution across
different reconstruction algorithms and imaging tasks. © 2012 American Association of Physicists in
Medicine. [http://dx.doi.org/10.1118/1.4725171]
Key words: MTF, NPS, NEQ, detectability index, CT, iterative reconstruction, statistical reconstruc-
tion, quality assurance, image quality assessment
I. INTRODUCTION
an increasing challenge in our ability to assess their perfor-
The evaluation of spatial resolution of imaging systems has mance, including spatial resolution. For example, iterative
played a central role in imaging performance evaluation.1 and statistical reconstructions can exhibit nonlinear signal
The advent of more complex imaging systems has posed characteristics, which can affect system resolution properties
4115 Med. Phys. 39 (7), July 2012 0094-2405/2012/39(7)/4115/8/$30.00 © 2012 Am. Assoc. Phys. Med. 4115
4116 Richard et al.: Towards task-based assessment of CT performance 4116
differently than with standard reconstruction algorithms,
I (x, y) = f (x − x , y − y )PSF(x , y )dx dy , (1)
i.e., filtered back projection (FBP). One of the most
comprehensive metric used to measure and report spatial where PSF(x, y) is defined as the PSF such that
resolution of imaging systems is the modulation transfer func-
tion (MTF).2 The MTF provides a measure of how well the PSF(x, y)dxdy = 1. (2)
system transfers contrast across spatial-frequencies. The MTF
is also essential to evaluating other key imaging performance The Fourier representation of the PSF is the optical transfer
metrics such as the detective quantum efficiency (DQE), function (OTF),
noise equivalent quanta (NEQ), and detectability index.3 OTF(u, v) = F {PSF(x, y)}, (3)
Previous methods for MTF measurements have used a va-
riety of “known” objects to generate an input to measure the where F{} denotes the Fourier transform. The MTF is the
point spread function (PSF). Ideally, a delta-function-like ob- modulus of the OTF as
ject would be employed; however, this approach is gener- MTF(u, v) = |OTF(u, v)| , (4)
ally not practical due to its low-signal power. In CT, a wire
describing the magnitude of transferred signal at each spatial
method provides a pseudo-point object across multiple slices
frequency. Note that (1) Eq. (2) implies that MTF(0) = 1 (dc
and can provide a 2D in-slice MTF by averaging realization
signal magnitude is unchanged by the transfer function); and
across multiple slices. This approach has been investigated
(2) an ideal system has MTF = 1 at all spatial frequencies.
since the early years of CT.4, 5 Even so, most current qual-
The MTF can be employed to characterize the resolution of an
ity assurance (QA) methods have adopted line-pair pattern
imaging system in terms of MTF( f) or to reflect the resolution
methods due to their simpler and more intuitive approach.6, 7
in a scalar form, for example, the frequency ( f50 ) at which
In radiography, the most common method for measuring the
MTF( f) reduces to 0.50.
MTF is the angled edge technique, which provides an edge
Experimentally, it is impractical to measure the PSF di-
across multiple rows or columns that can generate a “super-
rectly. That would require imaging of an infinitesimally small
sampled” edge-spread function (ESF).8, 9 The ESF can be dif-
object, a 2D delta function, so that I(x,y) = PSF(x,y). If one
ferentiated to yield the line-spread function (LSF) from which
is to do so, the mean signal magnitude would be correspond-
the Fourier transform yields the MTF. This method has been
ingly small and the resulting image would be dominated by
adapted to CT using sharp objects or edges to measure in-slice
noise. More convenient impulse function techniques are typi-
resolution.10
cally used for such measurements, such as the LSF, which is
Both the wire and edge techniques have limited utility in
the Radon transform of the PSF,
advanced CT systems. These systems offer nonlinear image
processing and reconstruction algorithms which can lead to a LSF(x) = PSF(x, y)dy. (5)
contrast and dose dependency for the MTF. The conventional
wire and edge methods typically use highly dense materials The LSF can be measured from the image of a narrow
and very low noise to simulate an infinite impulse (i.e., sig- slit.11 Alternatively, the ESF, given by the antiderivative of
nal) to maximize CNR and improve measurement precision. the LSF, may be employed as
x
However, this corresponds to objects that are atypical in clin-
ESF(x) = LSF(x )dx , (6)
ical images. Therefore, for these CT systems, it is important −∞
to use test objects with (lower) contrast and (higher) noise
which can be measured from the image of a sharp edge. This
to yield MTFs relevant to realistic objects. This paper intro-
can be rewritten as
duces a novel method for evaluating the spatial resolution of
∂
CT systems using a conventional QA phantom. This work was LSF(x) = ESF(x). (7)
motivated by a desire to: (1) assess MTF at various contrast ∂x
and noise levels; (2) measure the MTF using a standard QA Hence, the PSF, LSF, ESF, and MTF are all related. For a ra-
phantom without any additional equipment; and (3) report on dially symmetric MTF, Eqs (2)–(7) together with the Fourier
the MTF performance of two new reconstruction algorithms slice theorem suggest that the one-dimensional (1D) MTF
in terms of their performance compared to standard recon- (i.e., a slice of the full 2D MTF) is given by
struction. ∂ ESF(x) exp(−2π if x)dx
∂x
MTF(f ) = ∂ , (8)
∂x
ESF(x)dx
II. METHODS AND MATERIAL where the denominator ensures that MTF(0) = 1 [equivalent
II.A. Theory to Eq. (2)]. Therefore, the MTF can be experimentally as-
sessed by imaging a sharp edge object to yield an estimate
Fourier based metrology is a prevalent approach to char- of the ESF, which in turn is related to the LSF (by derivative)
acterizing medical imaging performance. This includes the and to the MTF as in Eq. (8).
MTF, which characterizes the resolution of the imaging sys- The requirements for measuring the MTF of an imaging
tem. The output signal I(x,y) (i.e., image) is related to the in- system include: (1) a linear signal response and (2) a shift-
put signal f(x,y) (i.e., object) by the impulse response function invariant signal (i.e., signal independent of location in the im-
as age). However, in the strictest sense of the definition most
F IG . 1. (a) Axial slice of the ACR (model 464) phantom and (b) one of the insert zoomed in. (c) Image of the GE performance phantom (GEPP), which includes
a 50 μm tungsten wire, zoomed in (d).
imaging systems are nonlinear and shift-variant. Neverthe- II.B. MTF measurement method
less, under certain conditions, linearity and shift-invariance
can be a reasonable assumption. These conditions often in- An experimental methodology was developed to assess
clude low-contrast and localized measurements. As a result, the MTF from the image of a 2D circular object as illustrated
it is often necessary to qualify under which conditions the in Fig. 1. The method was largely adapted from the edge tech-
MTF results are valid. Measurement of MTF for nonlinear nique developed for radiography but with incorporation of a
system has been the focus of past research. For example, the circular object.9 The ACR phantom was used as the basis of
measurements of MTF on cathode ray tube (CRT) display the methodology. As described above, a motivation of this
as discussed in the AAPM task group 18,12 which employed work was the fact that nonlinear reconstruction algorithms in
low-contrast edge object for measuring the MTF. The MTF CT may display spatial resolution dependence on contrast and
for object-specific assessment was also essential in predict- noise level in the image. The circular objects of varying con-
ing image performance in dual-energy imaging using noise trast present in the ACR phantom (Fig. 1) enables the eval-
reduction algorithms. In that work, the MTF was measured uation of that dependence. The dependence of the MTF on
using materials that mimicked the contrast of various types of signal and noise represents a different type of measurement
tissues (i.e., soft tissue or bone; Ref. 13) and was employed to than those made on linear systems. Therefore, to highlight this
fairly accurately predict the detectability index for detection important difference between the conventional MTF and the
and discrimination tasks.14 Similarly, in this work, the MTF MTF measured for the purpose a task-based assessment on
is measured under predefined conditions such as contrast and nonlinear systems, we introduce a notation change to the MTF
noise levels. This yields a specific task-based transfer function as MTFTask to denote this distinction. In this sense, the object
that can be used to compute performance assessment metrics used to measure MTFTask essentially mimics the contrast and
such as the detectability index. This work is further justified edge of the detected signal as perceived by the observer. This
by the need to define the MTF on a locally linearizable region MTFTask is quite different from what is conventionally known
of the object to assess the performance of task-specific perfor- as the system MTF, rather MTFTask refers to an object depen-
mance in the presence of nonlinear reconstruction algorithms. dent measure of sharpness in the Fourier domain analogous to
Wire
MTFTask
MTFTask
MTFTask
0.6 0.6 0.6
F IG . 3. Comparison of the disk edge and wire based method for measuring the MTFTask for images reconstructed with (a) FBP, (b) ASIR, and (c) MBIR. Data
were acquired at 120 kVp and 200 mA.
Figure 4 demonstrates the MTFTask for FBP, ASIR, and bone) and across five different tube currents. The data dis-
MBIR reconstructions acquired at the highest mA level play a modest dependence on tube current for all contrast lev-
(200 mA) for various contrast inserts. The error bars display els. The MTFTask is highest for the highest contrast object and
the variability in the measured due to both stochastic sources lowest noise setting. Similarly, Fig. 6 shows the MTFTask mea-
such as image noise but also due to more deterministic sured using MBIR reconstructed data across different materi-
sources such as the variability and nonuniformity in the als and tube currents. The data display a strong dependence on
reconstruction. For the FBP measurements, the lower contrast contrast, as noted above, but even a more pronounced depen-
insert (i.e., acrylic and PE) provide low signal-to-noise ratio dence on tube current, where the MTFTask seems to increase
resulting in noisier MTFTask measurements. Nonetheless, with tube current across all materials. The highest MTFTask
the FBP-based MTFTask results agree across the different are noted for the highest tube current, but they appear to
inserts; they are well within the measurement error. This is an plateau towards highest mA settings. The data show clearly
expected result as the FBP is a linear reconstruction technique that a measurement of MTFTask at high dose and high contrast
and as such the MTFTask is expected to be independent of would provide a “best case” representation of the system’s
contrast or noise. The ASIR results exhibit similar loss in resolution.
MTFTask precision as a function of object contrast. But most Figure 7 summarizes the results by plotting the spa-
importantly, they further exhibit a dependency on contrast tial frequency at which the MTFTask reduces to 50% ( f50 ).
with higher resolution noted for the highest contrast object. Figures 7(a)–7(c) plot f50 as a function of tube current (mA)
The MBIR exhibits the largest dependence on object contrast, for the MTFTask measurements across the different inserts and
where the MTF is dramatically enhanced for the bone object reconstruction algorithms. Results for FBP show no depen-
while the MTFTask for the PE object is comparable to the dence with f50 fixed at 0.35 mm−1 . ASIR50 , in contrast, dis-
FBP-based MTFTask . These results illustrate that for iterative plays a small dependence with f50 ranging from 0.22 at very
reconstructions, the MTF needs to be determined as an low tube currents and low contrast (i.e., PE at 10 mA) to
object-based MTFTask rather that a system-based MTF. 0.35 mm−1 for high contrast and high tube current (i.e., bone
Figure 5 plots the MTFTask measured using ASIR50 re- at 200 mA). Results for MBIR display the largest dependence,
constructed data across different materials (PE, acrylic, and with f50 ranging from 0.21 to 0.6 mm−1 . For comparative
1 1 1
0 0 0
0 0.2 0.4 0.6 0.8 0 0.2 0.4 0.6 0.8 0 0.2 0.4 0.6 0.8
Spatial Frequency (cycles/mm) Spatial Frequency (cycles/mm) Spatial Frequency (cycles/mm)
F IG . 4. Comparison of MTFTask measured across the different inserts (bone, acrylic, and PE) for images reconstructed with (a) FBP, (b) ASIR50 , and (c) MBIR.
Data were acquired at 120 kVp and 200 mA.
1 1 1
MTFTask
MTFTask
MTFTask
0 0 0
0 0.2 0.4 0.6 0.8 0 0.2 0.4 0.6 0.8 0 0.2 0.4 0.6 0.8
Spatial Frequency (cycles/mm) Spatial Frequency (cycles/mm) Spatial Frequency (cycles/mm)
F IG . 5. Plots of MTFTask for ASIR50 at five tube current levels (25, 50, 75, 100, and 200 mA) and contrast inserts (a) PE, (b) acrylic, and (c) bone. Data were
acquired at 120 kVp.
purposes, Fig. 7(d) plots the dependency of noise on mA. The algorithms and across a range of noise and contrast levels.
significant correlation between mA and resolution for the it- In the case of FBP, no dependence was observed due to the
erative reconstructions, there is a clear paradigm shift with system’s linearity (as expected). For iterative reconstructions,
advanced reconstructions, where compared to standard FBP, the MTFTask was found to be dependent on both object
both noise and resolution vary as a function of tube current contrast and image noise. This makes the comparison of
and dose. MTF between CT systems and protocols more challenging.
The contrast dependence of the task-based MTF further
implies that the results presented above depend not only on
IV. DISCUSSION the material used for the MTF measurement but also the
surrounding material. Therefore, the MTFTask results shown
Advances in CT, including iterative algorithms, are in this work are representative only of the conditions used
expected to require more comprehensive performance as- in this study. Furthermore, the shape of the object may also
sessment methods. Spatial resolution will be central to this impact the MTFTask but was beyond the scope of this work.
evaluation and the MTF is essential for assessing Fourier- This points toward the need to define specific guideline for
based system performance metrics such as the NEQ and measuring the MTFTask for iterative algorithms, for example,
the detectability index.16, 17 The presented results highlight by specifying the noise or contrast at which the MTFTask is
an important paradigm shift in the clinic with iterative being evaluated. It is essential to establish that dependence
reconstructions—as the pixel noise can be constrained by as otherwise traditional wire-based MTF techniques would
the reconstruction algorithm, this causes distortions in the only provide the best case scenario in most cases. The abil-
image at low x-ray signal and lowers the spatial resolution ity to provide an object-based MTF also provides a more ac-
as characterized by the MTFTask . Current MTF assessment curate description of the imaging task when computing the
methods only focus on a single contrast and noise level.4, 5 detectability index.18 For example, a large low-contrast ob-
As a result, this work presented a method for measuring ject and a small high contrast object would have different
the MTFTask for CT systems for different reconstruction representative MTFs. This work provides a methodology for
1 1 1
MTFTask
MTFTask
0 0 0
0 0.2 0.4 0.6 0.8 0 0.2 0.4 0.6 0.8 0 0.2 0.4 0.6 0.8
Spatial Frequency (cycles/mm) Spatial Frequency (cycles/mm) Spatial Frequency (cycles/mm)
F IG . 6. Plots of MTFTask for MBIR at five tube current levels (25, 50, 75, 100, and 200 mA) and contrast inserts (a) PE, (b) acrylic, and (c) bone. Data were
acquired at 120 kVp.
1 1
0.8 0.8
(cycles/mm)
(cycles/mm)
0.6 0.6
0.4 0.4
0.2 0.2
0 0
0 50 100 150 200 0 50 100 150 200
Tube current (mA) Tube current (mA)
1
100
0.8
(cycles/mm)
80
Noise (HU)
0.6
60
0.4
40
0.2
20
0
0 50 100 150 200 0
0 50 100 150 200
Tube current (mA) Tube current (mA)
F IG . 7. Plots of f50 for MTFTask for FBP, ASIR50 , and MBIR at across tube current for the (a) PE, (c) acrylic, and (c) bone insert. (d) Plots of the corresponding
pixel noise measured in the background. Data were acquired at 120 kVp.
computing task-based MTF for evaluating observer model fig- fying information such as, but not limited to, the noise and
ure of merits for CT performance assessment using iterative contrast level at which the MTFTask is measured, they can pro-
reconstruction as presented in works by Zeng and Myers19 vide some reflection of local resolution for clinically relevant
and Yendiki and Fessler,20 which employed mathematical ob- tasks. Note that this is analogous to the MTF of dual-energy
servers to assess statistically reconstructed images. images using noise reduction where the MTF is dependent on
Applications for MTFTask naturally include QA. The ACR the contrast of the object.13 In that case, the proper descrip-
phantom used in this study is currently available in many clin- tion of the MTF was essential to accurately model and predict
ics as a QA phantom. Extending QA evaluations with a task- system and human performance.14 Another limitation of this
based MTF measurements could be automated, as is currently work is that the MTF was radially averaged. This was per-
underway at our institution. In addition to evaluating the rel- formed to minimize the noise in the measurements, especially
ative reliability of the MTF over time for QA purposes, the those made at a low mA. Therefore, the result ignores any
interpretation of the MTFTask in more absolute terms may also anisotropy and effectively averages out the MTF in the axial
be desirable. This can be achieved by combining the MTFTask plane. One way to overcome this limitation is to only average
with a task function to generate a surrogate of image quality, a certain angular extent, which is certainly possible within the
such as the detectability index,3 which could provide a useful framework presented but was outside the scope of this work. It
approach for comparing and optimizing image reconstruction should also be noted that the error bars describe the stochas-
performance, albeit under very specific imaging conditions. tic variability as well as the deterministic variability associ-
Notwithstanding the findings of the study, certain limita- ated with the nonuniformity of the MTF across realizations.
tions need to be acknowledged. First, in this work, the MTF While noise and contrast strongly affect the MTFTask , other
was applied not as a generic reflection of system resolution factors could be analyzed. For example, the MTFTask could
but rather as an object-specific quantity. This was done so to be measured as a function of other parameters such as the
accommodate the nonlinearity of the iterative reconstruction contrast-to-noise ratio. Furthermore, the study is related to its
methods and is reflected by the notation change—MTFTask . application to only one CT system; therefore, the MTF was
However, provided that the MTFs are annotated with quali- measured for only two types of iterative reconstructions—
there are many parameters in iterative reconstructions, such 6 E. L. Nickoloff and R. Riley, “A simplified approach for modulation trans-
as the regularizer and edge preservers, that could affect the fer function determinations in computed tomography,” Med. Phys. 12(4),
437–442 (1985).
MTFTask —future studies are warranted to include other CT 7 S. Rathee, B. G. Fallone, and D. Robinson, “Modulation transfer func-
system and reconstruction algorithms. tion of digitally reconstructed radiographs using helical computed tomog-
raphy,” Med. Phys. 29(1), 86–89 (2002).
8 E. Samei, M. J. Flynn, and D. A. Reimann, “A method for measuring the
V. CONCLUSION presampled MTF of digital radiographic systems using an edge test de-
vice,” Med. Phys. 25(1), 102–113 (1998).
In summary, the method described in this work provides 9 H. Fujita et al., “A simple method for determining the modulation trans-
a methodology for evaluation of the task-based MTF across fer function in digital radiography,” IEEE Trans. Med. Imaging 11(1),
various CT reconstruction algorithms. The method employs a 34–39.(1992).
10 I. Mori and Y. Machida, “Deriving the modulation transfer function of CT
conventional image quality phantom for CT with an adapted from extremely noisy edge profiles,” Radiol. Phys. Technol. 2(1), 22–32
circular edge technique to generate the ESF across vari- (2009).
11 I. A. Cunningham and B. K. Reid, “Signal and noise in modulation transfer
ous contrast and noise levels. The method provides a robust
function determinations using the slit, wire, and edge techniques,” Med.
method for characterizing resolution of different types of re-
Phys. 19(4), 1037–1044 (1992).
constructed objects. The findings show how iterative recon- 12 E. Samei, A. Badano, D. Chakraborty, K. Compton, C. Cornelius,
structions apply different tradeoffs across noise and resolution K. Corrigan, M. J. Flynn, B. Hemminger, N. Hangiandreou, J. Johnson,
as a function of tube current. In all cases, the noise was lower M. Moxley, W. Pavlicek, H. Roehrig, L. Rutz, J. Shepard, R. Uzenoff,
J. Wang, and C. Willis, “Assessment of display performance for medical
with ASIR and MBIR than with FBP. Nevertheless, the task-
imaging systems,” Report of the American Association of Physicists in
based MTF for MBIR was greater than for FBP at high-dose Medicine (AAPM) Task Group 18, 2005.
and high-contrast objects, however, the performance dimin- 13 S. Richard and J. H. Siewerdsen, “Cascaded systems analysis of noise re-
ished for low-dose, low-contrast objects. In contrast, MTFTask duction algorithms in dual-energy imaging,” Med. Phys. 35(2), 586–601
(2008).
for ASIR remained equal or lower than FBP in the set of imag- 14 S. Richard and J. H. Siewerdsen, “Comparison of model and human ob-
ing conditions investigated in this study. As a result, the MTF server performance for detection and discrimination tasks using dual-
assessment presented in this work begins to provide a method energy x-ray images,” Med. Phys. 35(11), 5043–5053 (2008).
15 J. B. Thibault et al., “A three-dimensional statistical approach to improved
for evaluating task-based resolution metrics for system opti-
image quality for multislice helical CT,” Med. Phys. 34(11), 4526–4544
mization and comparison across standard and iterative recon- (2007).
struction algorithms. 16 J. H. Siewerdsen, I. A. Cunningham, and D. A. Jaffray, “A framework for