Magnetic Resonance Imaging 24 (2006) 123 – 132

Application of perceptual difference model on regularization

techniques of parallel MR imaging
Donglai Huo a, Dan Xu b, Zhi-Pei Liang b, David Wilson a,c,4
a
Department of Biomedical Engineering, Case Western Reserve University, Cleveland, OH 44106, USA
b
Department of Electrical and Computer Engineering, University of Illinois, Urbana, IL 61801, USA
c
Department of Radiology, University Hospitals of Cleveland, Cleveland, OH 44106, USA
Received 4 August 2005; accepted 21 October 2005

Abstract
Parallel magnetic resonance imaging through sensitivity encoding using multiple receiver coils has emerged as an effective tool to reduce
imaging time or to improve image SNR. The quality of reconstructed images is limited by the inaccurate estimation of the sensitivity map,
noise in the acquired k-space data and the ill-conditioned nature of the coefficient matrix. Tikhonov regularization is a popular method to
reduce or eliminate the ill-conditioned nature of the problem. In this approach, selection of the regularization map and the regularization
parameter is very important. Perceptual difference model (PDM) is a quantitative image quality evaluation tool that has been successfully
applied to varieties of MR applications. High correlation between the human rating and PDM score shows that PDM should be suitable to
evaluate image quality in parallel MR imaging. By applying PDM, we compared four methods of selecting the regularization map and four
methods of selecting the regularization parameter. We found that a regularization map obtained using generalized series (GS) together with a
spatially adaptive regularization parameter gave the best reconstructions. PDM was also used as an objective function for optimizing two
important parameters in the spatially adaptive method. We conclude that PDM enables one to do comprehensive experiments and that it is an
effective tool for designing and optimizing reconstruction methods in parallel MR imaging.
D 2006 Elsevier Inc. All rights reserved.
Keywords: Perceptual difference model; Parallel MRI; Image quality; SENSE; Tikhonov regularization

1. Introduction artifacts. When the sensitivity information of the receiver

coils is available, SENSE is preferred [7].
Parallel MR imaging enables one to speed data acquisi-
The quality of the SENSE reconstruction is not only
tion or improve image signal-to-noise ratio (SNR). In
limited by the intrinsic nature of parallel imaging, but also
parallel imaging, multiple receiver coils with different local
deteriorated by the ill-conditioned inverse problem. The
sensitivities are used to acquire k-space signals. One could
problem can be solved by optimizing the geometry of the
reduce the number of Fourier encoding steps with sparser
coils [8,9] or by using advanced reconstruction techniques
sampling of k-space to save acquisition time or reconstruct
to improve the condition of the inverse matrix. Tikhonov
images with full sampling data to improve image quality.
regularization is a popular choice for improving reconstruc-
The concept of parallel imaging has been proposed since the
tion in parallel imaging [10,11]. Important design issues in
late 1980s and early 1990s [1 – 4]; however, practical
Tikhonov regularization are the selection of the regulariza-
parallel imaging was not developed until SMASH (Simul-
tion parameter and the prior information incorporated into
taneous Acquisition of Spatial Harmonics) [5] and SENSE
the regularization, or the regularization image.
(SENSitivity Encoding) [6] were proposed in the late 1990s.
In this paper, we investigate the optimization of Tikhonov
With sparse sampling of k-space, SENSE and SMASH are
regularization construction using quantitative image quality
two ways to reconstruct images while avoiding aliasing
measures. We use a perceptual difference model (PDM), a
computerized human vision model that calculates the visual
4 Corresponding author. Department of Biomedical Engineering, Case
Western Reserve University, Cleveland, OH 44106, USA. Tel.: +216
difference between a btest image Q and a bgold standard
3684099; fax: +216 3684969. image,Q which in our parallel imaging experiments corre-
E-mail address: dlw@po.cwru.edu (D. Wilson ). spond to fast-acquired images obtained with k-space
0730-725X/$ – see front matter D 2006 Elsevier Inc. All rights reserved.
doi:10.1016/j.mri.2005.10.018
124 D. Huo et al. / Magnetic Resonance Imaging 24 (2006) 123 – 132
subsampling and slow-acquired images obtained with full
sampling, respectively. PDM has been shown to correlate
well with human observers in a variety of MR experiments,
including spiral imaging and keyhole imaging [12,13]. In
computer experiments, we use PDM to investigate four
methods of selecting the regularization parameter and four
methods of selecting the regularization image. Once we have
identified parameters of interest, we use an optimization
algorithm and a PDM objective function to further optimize
Tikhonov regularization. Noise effects on regularization are
also investigated. In this study covering multiple image
reconstruction methods, continuous reconstruction parame-
ters, multiple test images, and variable noise, thousands of
images were generated. It would be impossible for human
observers to review this number of images, and, if they did,
given the inter- and intra-observer variability, uncertain
results would probably be obtained. The PDM makes such
experiments possible.
2. Methods
2.1. Perceptual difference model
The quality of the reconstructed images was evaluated
with the PDM designed in our laboratory. The model was
inspired by the visible differences predictor model reported
by Daly [14], but we have made multiple modifications over
the years. It has been validated for the evaluation of other fast
MRI applications and described in detail in previous papers
[12,13,15]. It contains components that model the nonline-
arity in the sensitivity of the retina [16,17], the contrast
sensitivity function [16], and the channels of spatial
frequency found in the visual cortex [18], as well as other
Fig. 1. A block diagram of the perceptual difference model (PDM). The output is a map showing the likelihood of a perceptual difference between the two input
images. The summation of this map over a region of interest gives a PDM error score. The gold standard image is acquired with a full k-space acquisition and
reconstructed using the total least square method. Subsequent test images are reconstructed from down-sampled k-space data reconstructed with different
regularization methods or regular SENSE method.
features including a measure of contrast [14], and visual de-
tection threshold [19]. The PDM structure is shown in Fig. 1.
Inputs to the PDM are an ideal reference image obtained from
the full-sampled reconstructed image and a test image of one
of the down-sampled images. Images are windowed to
maximize overall image contrast, and this same windowing
is maintained during evaluation by human observers de-
scribed later. The regular output of the PDM is a spatial map
representing the magnitude of differences that a human
observer would perceive between the two images. This map
can be summed over a region of interest (ROI), which is
defined manually to include relevant anatomy, to give a scalar
PDM error. In this paper, we used an ROI consisting of a
manually defined ellipse encompassing the whole head area.
2.2. PDM validation
The correspondence between human observer scoring of
image quality and PDM scoring was demonstrated through
experiments performed in a manner similar to that
previously reported by Salem et al. [12] and by Martens
and Meesters [20] on a similar model. To test the full range
of image quality in our simulated images, we selected 40
test images with PDM scores linearly spread between best
to worst. In the experiment, the 40 images were presented to
four subjects, one of the authors and three experienced
radiologists. The image was displayed following gray scale
windowing as reported previously. The region outside of
the region of interest was set to zero value (black). The
evaluation experiment was carried on a Matlab (The
MathWorks, Natick, MA) GUI program, and all the results
were automatically recorded. Each presentation consisted of
a two-panel display (Fig. 2A), with the high-quality re-
ference image and a randomly selected test image on the left
 D. Huo et al. / Magnetic Resonance Imaging 24 (2006) 123 – 132
Fig. 2. Human observer experiments and comparison to PDM. A screenshot
of the image display program (A) shows both reference and test image, and
contains sliders for entering IQ scores (see text). Forty images were
evaluated that covered the full range of image quality, and human observer
ratings were averaged over two observations before normalization as
described in the text. In (B), the average normalized ratings for each
observer and the corresponding PDM error score are plotted as a function of
the 40 test images, ordered according to PDM score. The linear relationship
between raw PDM score and human scores averaged over the four subjects
is shown in (C). A high correlation (R = 0.9388) exists between the human
score and PDM score. The intercept in x-axis (1.13) corresponds to
bnonperceptible differenceQ.
125
and right, respectively. Observers were instructed to score
the quality of the test image on a scale of 100 to 0, with 0
being the best quality and 100 being the worst quality, by
sliding a slider with a mouse or keyboard. We made
observers aware that we considered the reference image to
be bbestQ and that they should consider it to have a score of
0. Three ratings were obtained for each image pair. That is,
we asked observers to rate the test image on (1) overall
image quality, (2) b noise effects,Q and (3) aliasing and other
reconstruction errors. In a training session, naRve observers
were shown a wide range of images and instructed as to
what we considered b noiseQ (high-frequency, relatively
uncorrelated noise) and b aliasing and other reconstruction
errors.Q Following this discussion on at least 10 images,
observers performed a training session on at least 30 images
spanning a wide range of image quality, so as to help
calibrate them for the experiment. During this time, subjects
were free to ask questions of the first author. Test session
began right after the training session. To account for intra-
observer differences, each of the 40 test images was
displayed and evaluated twice. The experiment was
carried out in a darkened room and normally took 1 h. A
perceptually linearized, high-quality gray scale monitor was
used. There was no time limitation in the experiment, and
subjects were allowed to revise their results, including
backtracking, at any time.
2.3. Data generation and simulation
The head image of a healthy subject was acquired from a
four-channel surface coil and a body coil. A fast scan from
the body coil and from the surface coils provided the image
data for estimating the complex sensitivity map of the
surface coils. A slow scan from the surface coils provided
the data for actual reconstruction. The full sampled data
together with the sensitivity map information were used to
reconstruct a high-quality image as the reference image for
the PDM. Data were down-sampled and reconstructed to
create the test images.
To investigate noise effects, high-quality acquired k-space
data were corrupted by correlated MRI noises [21]. Gaussian
distributed noise was added to both real and imaginary
channel data for each coil. In order to investigate the noise
effects in both the acquired data and the calibrating data
(prescan), we generated data sets giving five different k-
space acquired data SNRs (20, 25, 30, 50 and 10,000 dB)
and five prescan data SNRs (20, 25, 30, 50 and 10,000 dB).
For each noise data set, 30 realizations were calculated, and
the average PDM value was used as the image quality value.
2.4. SENSE reconstruction
In general, the equation for SENSE image reconstruction
can be written as
Sq ¼ d ð1Þ
in which q contains the desired pixel values, d contains all
the measured k-space data and S is the estimated sensitivity
126 D. Huo et al. / Magnetic Resonance Imaging 24 (2006) 123 – 132

map. In practice, there are errors in both the estimation of image is increased. We will also consider four methods of
sensitivity maps S and the measured data S. Eq. (1) becomes selecting k:
ðS0 þ DS Þðq0 þ DqÞ ¼ d0 þ Dd ð2Þ ! Set k as a constant over the entire image. This is
the method first proposed for Tikhonov regulariza-
When ||S 0+||||DS|| b1,

tion [22].
Dq j Dd DS ! L-Curve method. The L-curve is a parametric plot of
V þ ð3Þ
q0 1  jODSO=S0 d0 S0 the data fitting error and the deviation from the
regularization image. The idea is that the optimal
where j = ||S 0 ||||S 0+ || is the condition number of S 0 . regularization parameter is to minimize and balance
For nonlocalized coils with large acceleration factors, S 0 the two error terms and it corresponds to the elbow of
tends to be ill-conditioned and leads to a significant recon- the L-curve [25]. This method has been used in
struction error. parallel imaging with some success [10].
2.5. Tikhonov regularization of SENSE reconstruction ! Generalized cross-validation (GCV) [26]. The basic
idea of GCV is that if some data are left out and
Tikhonov regularization is perhaps the most common the solution is computed from the reduced problem,
method to solve the ill-conditioned problem. A weighted then the solution is predicted from the observed data
sum of the data misfit term ||Sqd||2 and a regularization and the choice of the regularization parameter should
term ||qq r||2 with weighting factor k 2 are formed, and the be independent of an orthogonal transformation of
solution is the one that minimizes the sum, the data.
n o ! A spatially adaptive method was proposed [27] to
q4 ¼ arg min OSq  dO2 þ k2 Oq  qr O2 ð4Þ
q choose k in a spatially dependent fashion, considering
the conditioning of all the equations. Specifically, the
A closed form of the solution could be expressed as
proposed algorithm first sets k to [kmin, kmax] and
 1 then selects the k point-by-point in a spatially
q4 ¼ qr þ S H S þ k2 I S H ðd  Sqr Þ ð5Þ
adaptive manner. More specifically, k(x) is set to be
The selection of regularization parameter k and regular- a linear function of the local condition number of S;
ization image q r will greatly affect the final reconstructed that is,
image quality.
jðS Þ  jmin ðS Þ
kð xÞ¼ ðkmax  kmin Þ þ kmin ð6Þ
2.5.1. Selection of regularization image jmax ðS Þ  jmin ðS Þ
Researchers have proposed various methods for estimat-
ing the regularization image. We will compare four methods where 8 9
in this paper. >
> >
>
>
< >
=
maxi ri
! Setting q r = 0. This was used in Ref. [22] with kmin ¼ arg min 
bK
k >
>
> k2 >
>
>
some success. : min i ri þ ;
! Median filter method. The conventional SESNE ri
algorithm is used to obtain an initial reconstruction, ( )
X
which is then filtered by a 55 median filter to kmax ¼ arg max OSqreg ðkÞ  dOb E ð7Þ
suppress the residual aliasing artifacts [11]. k x
! Autocalibration method. In this method, additional
k-space center lines at the Nyquist rate is acquired and
Here r i is the ith singular value of sensitivity map
autocalibrated to reconstruct a low-resolution regu-
S. K and E are user-defined constants. In our evaluation
larization image [23].
experiment, they were set as K =10 and E =0.5; they were
! Generalized-series (GS) model method. We will also
further optimized in an optimization experiment.
acquire the additional center k-space lines as we did
in the autocalibration method. An image is then
reconstructed from the k-space data in each coil using 3. Experiments and results
the GS model [24] whose basis functions are formed
3.1. Validation of PDM
from the reference data collected at the Nyquist rate.
Raw data in the PDM validation experiment were
2.5.2. Selection of regularization parameter processed. First, two scores given for the same test image
The choice of the regularization parameter k is also from the same subject were averaged to reduce the intra-
very important to successful regularization. As k increases, observer difference. To compensate for the scale boundary
the contribution to the reconstruction from the data effects, we used a nonlinear scale transformation method
themselves is reduced and that from the regularization recommended by the International Telecommunication
 D. Huo et al. / Magnetic Resonance Imaging 24 (2006) 123 – 132 127

Table 1
Correlation coefficient between human subjects, PDM and MSE scores
Subject 1 Subject 2 Subject 3 Subject 4 PDM MSE
Subject 1 0.9707 0.9541 0.9557 0.9626 0.9120 0.4218
Subject 2 0.9541 0.9719 0.9509 0.9529 0.8951 0.4926
Subject 3 0.9557 0.9509 0.9616 0.9738 0.9506 0.5113
Subject 4 0.9626 0.9529 0.9738 0.9855 0.9394 0.4486
PDM 0.9120 0.8951 0.9506 0.9394 1 0.4399
MSE 0.4218 0.4926 0.5113 0.4486 0.4399 1
Correlation coefficients were calculated for all possible combinations of subjects, PDM scores and MSE scores. Note that PDM correlates nearly as well as the
correlation between subjects. The performance of PDM is much better than the performance of MSE.

Union in their report on methods for assessing television
images [28] so as to normalize the human rating data. The
mean normalized human observer scores were graphically
and statistically compared to the PDM scores on the same
test images.
Human observer scoring of image quality was highly
correlated with PDM scoring of image differences. In
Fig. 2(B), the average normalized ratings for each observer
and the corresponding PDM error score are plotted as a
function of the 40 test images. We could see that the
performance of PDM is just like a normal human subject.
There is a linear relationship between the human scores
averaged over the four subjects and PDM score, as shown in
Fig. 2(C). The x-intercept (1.13) corresponds to a

Fig. 3. For reduction factor R = 2, PDM score for various q r and k selection combinations. The best solution (lowest PDM score) is shown in (B), corresponding
to the spatially adaptive method to select k and the GS method to select q r. Image reconstructed with full k-space data, or the reference image, is shown
in (C). The image reconstructed with regular SENSE without regularization is shown in (D) and the corresponding PDM score (4.33) is shown as the horizontal
bar in chart (A).
128 D. Huo et al. / Magnetic Resonance Imaging 24 (2006) 123 – 132

b nonperceptible difference.Q Data are further analyzed in
Table 1. To show the intersubject difference, we use a
simple linear regression model y =mx+b to fit any two
subject readings, and the correlation coefficient is shown in
Table 1. The same procedure applies to calculation of the
correlation coefficient between human subject rating and
PDM [or mean square error (MSE)]. The intrasubject
difference is reflected by the correlation coefficient between
two measurements from the same subject. And all these
correlation coefficients are listed in Table 1. The PDM
accurately reflects human observer ratings of image quality.
As a comparison, another image quality index, the MSE,
correlated less well with human observer ratings.
PDM score is not the only possible output format of
PDM model. One could also get the PDM value in different
spatial frequencies. As seen in Fig. 2(A), human scoring of
noise, aliasing, and other effects was also collected in the
experiments. Our analysis (results not shown here) shows
that a good correlation exists between the high-frequency
output of the PDM model and human scoring of the noise
effects. Accordingly, low-frequency output of the PDM
score correlates better with aliasing and other effects.
Although further investigation is still needed, PDM provides
a possible way to separate the different effects that degrade
the image quality.
3.2. Comparison of regularization methods
The PDM was first used to select the best regularization
methods. The test images were reconstructed using reduction
factors R =2 or 4 (down-sampled by 2 or 4), four methods of
selecting the regularization image and four methods of
selecting the regularization parameter. Images are also
reconstructed by the regular SENSE method without regu-
larization from the two down-sampled data sets. Image
quality of all the reconstructed images is evaluated by PDM.
With R = 2, the impact of the reconstruction methods on
the image quality is shown in Fig. 3. Four methods of
selecting the regularization parameter k and four methods of
selecting regularization image q r generate 16 combinations,
and from the figure we can see that the GS method
combined with the spatially adaptive method gives the best
solution. The corresponding image is shown in Fig. 3(B).
The reference image is shown in Fig. 3(C) and the image
reconstructed with regular SENSE (no regularization) is

Fig. 4. For reduction factor R = 4, PDM score for various q r and k selection combinations. The best solution (lowest PDM score) is shown in (B), corresponding
to the spatially adaptive method to select k and the GS method to select q r. The image reconstructed with regular SENSE without regularization is shown in (C)
and the corresponding PDM score (31.46) is shown as the horizontal bar in chart (A).
 D. Huo et al. / Magnetic Resonance Imaging 24 (2006) 123 – 132 129

shown in Fig. 3(D). The PDM score of Fig. 3(D) is shown
by the horizontal bar in Fig. 3(A). As expected, down
sampling with R = 2 degrades image quality as compared to
the reference image. As we can see, the unregularized image
gives a PDM score of 4.33, somewhat higher than the
bimperceptibleQ threshold. In this case, the application of
regularization helps improve the image quality to some
extent, but the improvement is limited, as observed by the
PDM score and the visual inspection.
Results are shown in Fig. 4 for R = 4. Again, we can see
that the GS method to select the regularization image
combined with the spatially adaptive method offers the best
solution. The corresponding image is shown in Fig. 4(B),
and the regular SENSE method reconstructed image is
shown in Fig. 4(C). In this case, we can see that the regular
SENSE reconstruction generates images with significant
errors when the reduction factor is high, and regularization
improves image quality considerably.
3.3. Parameter optimization in the spatially
adaptive method
From results in the last section, the GS method of
selecting q r and the spatially adaptive method of selecting k
were best. However, the spatially adaptive method needs
two free parameters, K and E. In the previous evaluation, we
set them to be 10 and 0.5, based on our previous anecdotal
experience. PDM provides a means of optimizing these free
parameters. We used a simplex search algorithm that was
included in Matlab to find the optimal point. During the
optimization process, about 50 images were evaluated, and
this number should change with different initial conditions
and stopping criteria. In Fig. 5, we see how the recon-
structed image quality changes as a function of K and E
for R = 4. After searching through the 2-D surface, we can
easily find the optimized values K = 5.91 and E =1.04 cor-
responding to the lowest PDM score. The optimal image is
shown in Fig. 5(B) corresponding to point A in the image.

Fig. 5. Use PDM score as object function to optimize the free parameters K and E. The surface plot in (A) shows how the PDM score changes with the changes
of K and E. Point A, in which K = 5.91 and E = 1.04, represents the best regularization image [shown in (B)] we can get. Image corresponding to unoptimized
point B, with K = 17 and E = 0.8, is shown in (C).
130 D. Huo et al. / Magnetic Resonance Imaging 24 (2006) 123 – 132
Fig. 6. The PDM score change with different prescan data SNRs and
acquisition data SNRs. The figure reflects the image quality reconstructed
with the Tikhonov regularization method in which we use the GS method
to select the regularization image and the spatially adaptive method to
select the regularization parameter. In the spatially adaptive method, free
parameter K was set to 5.91 and E was set to 1.09. The error bar reflects
the PDM score standard error distribution for different noise realizations.
Image quality was degraded with the increase of noise levels, and no
significant difference is observed between the 50- and 10,000-dB data
SNR (overlapped).
For comparison, a less optimal image from point B is shown
in Fig. 5(C).
3.4. Noise effects
Data sets with R = 4 and added noise lead to very noisy,
sometimes unacceptable reconstructed images with the
normal SENSE reconstruction. Regularization is essential
here. We tested the performance of images reconstructed
under noise conditions, with GS for q r and spatially
adaptive for k. A preselected K = 5.91 and E =1.04 were
used in the reconstruction. We found that the optimal K and
E are reasonably insensitive to added noise. For example, in
the case of 30-dB SNR for both prescan data and acquired
data, the optimal K and E are very similar to optimal values
with no added noise. The change in PDM scores (7.48 with
preselected value to 7.35 with optimal value) was very
small, and we could detect no visible change in the images.
PDM scores of regularized images under different noise
SNRs are shown in Fig. 6. The quality of the reconstructed
images is degraded significantly both by noise in the
acquired data and by the sensitivity map. As expected,
image quality improved with increasing SNRs. Notice that
the difference between 50 - and 10,000 - dB data SNR is very
small and the corresponding points almost overlap.
4. Discussion
In general, regularization greatly improves reconstruc-
tions as compared to normal SENSE when a high reduction
factor (R = 4) is used, especially in the presence of noise.
Much less improvement is obtained with a lower reduction
factor (R = 2). For all kinds of regularization reconstruction
methods, our results show that at both high and low
reduction factor conditions, the spatially adaptive method is
the best method for selecting the regularization parameter k
and GS model is the best method for selecting the
regularization image q r. In the spatially adaptive method,
we need to set two free parameters, K and E. PDM has the
ability to optimize these parameters and optimized results
can significantly improve the regularized image quality.
Our results have shown that PDM could be very suitable
for evaluating the image quality in parallel MR imaging. A
lot of factors, including noise, blur and other artifacts, will
degrade the image quality. PDM can correctly consider all
of these factors and give a score close to human rating.
Based on our results, the MSE is less accurate in evaluating
the image quality. Our previous experiments [12] also show
that MSE is accurate in evaluating noise effects, but fails on
blur and other artifacts. Another good example of PDM’s
superiority over MSE is the processing of mean background
difference. For the most part, a mean background change
would not affect one’s ability to interpret the image. MSE
would give a very large error value because every pixel will
contribute to the error. As for PDM, since we know that the
mean background corresponds to the zero frequency
channel of the contrast sensitivity function included in the
PDM, this channel is weighted less in the whole calculation.
As a result, PDM gives a much more reasonable error
score than MSE.
Both previous [12,13,15] and present results show the
capability of using PDM to evaluate MR imaging recon-
struction methods and optimize important free parameters. It
makes possible experiments with a large number of
parameters and noise conditions to evaluate. For instance,
with regard to this study on parallel imaging, we have
evaluated at least 2000 images. It is impossible for humans
to systematically evaluate such a large number. In addition,
any bfast Q human evaluation method will probably be very
subjective and subject to bdriftQ of image quality criteria.
Currently, for each 256256 test image, the time to
calculate the PDM score is about 30 s on our Pentium IV
3.0-GHz Dell computer (Dell, USA). Calculation time could
be reduced in several ways. In the condition of comparing a
lot of images or optimizing important parameters, we could
store the information of all channels of the reference image.
We could also store some other information when the
viewing conditions (luminance, distance, resolution, etc.)
are fixed. Therefore, the calculation time can be greatly
reduced and the calculation might be live with the data
acquisition process.
Our results show that regularization techniques are very
helpful on improving the image quality of SENSE
reconstruction, especially when the reduction factor is
high. Our results show that the best regularization
method will be the GS method to select q r and the spatially
adaptive method to select k. Besides the best solution, in
our calculation, setting regularization parameter as a
constant and using the median filter method to select the
 D. Huo et al. / Magnetic Resonance Imaging 24 (2006) 123 – 132
regularization image are another quick and acceptable so-
lution. In the selection of regularization parameter, L-curve
and GCV method did not give very good results, and they
are very computationally intensive. We do not recommend
these two methods based on our experiment results.
It is exciting that we can use the PDM to optimize free
parameters in the reconstruction. Optimization is advisable,
as shown by the large difference between the optimized and
unoptimized images in Fig. 5. The surface in Fig. 5(A) is
very smooth, indicating that one need not be overly
concerned with local minima. Indeed, we tested several
different initial values, and each converged to the same
result. In addition, as described previously, the optimum was
reasonably insensitive to added noise.
In the reconstruction optimization process, we had to
define a stopping criterion. Experimentally, we determined
that a tolerance of 1% gave nearly optimal results and saved
considerable time as compared to a much tighter tolerance.
PDM allows the MR-imaging engineer to make quanti-
tative trade-offs, and the experiments with added noise give
an excellent example. We can clearly see the effect of added
noise on image quality in Fig. 6. Image quality is improved
with increasing SNR of acquired data, but the improvement
is limited above 50 dB because the image quality is now
limited by the inaccurate estimation of the sensitivity map
and the ill condition of the inverse problem. The 50- dB
value thus provides a guideline; one need not design a
system with SNR exceeding much beyond this, possibly
saving acquisition time. We can also see the importance of
accurate estimation of the sensitivity map. Often, low-
resolution, fast sequences are used to get the sensitivity
map. From Fig. 6, prescan SNR can degrade image quality
below 30 dB. One can also compare image quality with
regularization (Fig. 6) to normal SENSE reconstruction
(Fig. 4), for R =4. Even at the lowest SNR (20 dB for
both acquisition and prescan), image quality following
regularization exceeded that for regular SENSE with no
added noise. This indicates the power of improved
reconstruction techniques.
Although results presented here were generated from a
single MR data set, we have spot tested results on other
brain data sets having different acquisition sequences and
different methods to estimate the sensitivity map, and results
are similar. The best regularization method is still the GS
method to select q r and the spatially adaptive method to
select k. Furthermore, using K = 5.91 and E = 1.09 generated
near optimal PDM score (within 1% difference) in other
cases. Therefore, we can reasonably assume that results are
generalizable to some other conditions.
5. Conclusion
PDM has been successfully applied to evaluate the
regularization methods in parallel MR imaging and to
optimize the important free parameters. Tikhonov regular-
ization is very important to improve the image quality when
131
the reduction factor is equal or close to the number of
coils. The GS method to select the regularization image
combined with the spatially adaptive method to select the
regularization parameter gives the best solution for Tikho-
nov regularization.
Acknowledgments
This work was supported under NIH grant R01
EB004070 and NSF grant NSF-BES-0201876. We thank
for the contribution from Leslie Ying in University of
Wisconsin, Milwaukee, WI, and the subjects for participating
in the experiments.
References
[1] Roemer PB, Edelstein WA, Hayes CE, Souza SP, Mueller OM. The
NMR phased-array. Magn Reson Med 1990;16:192 – 225.
[2] Hutchinson M, Raff U. Fast MRI data acquisition using multiple
detectors. Magn Reson Med 1988;6:87 – 91.
[3] Kwiat D, Einav S, Navon G. A decoupled coil detector array for fast
image acquisition in magnetic-resonance-imaging. Med Phys 1991;
18:251 – 65.
[4] Ra JB, Rim CY. Fast imaging using subencoding data sets from
multiple detectors. Magn Reson Med 1993;30:142 – 5.
[5] Sodickson DK, Manning WJ. Simultaneous acquisition of spatial
harmonics (SMASH): fast imaging with radiofrequency coil arrays.
Magn Reson Med 1997;38:591 – 603.
[6] Pruessmann KP, Weiger M, Scheidegger MB, Boesiger P. SENSE:
sensitivity encoding for fast MRI. Magn Reson Med 1999;42:
952 – 62.
[7] Griswold MA, Jakob PM, Heidemann RM, Nittka M, Jellus V, Wang
JM, et al. Generalized Autocalibrating Partially Parallel Acquisitions
(GRAPPA). Magn Reson Med 2002;47:1202 – 10.
[8] de Zwart JA, Ledden PJ, Kellman P, van Gelderen P, Duyn JH. Design
of a SENSE-optimized high-sensitivity MRI receive coil for brain
imaging. Magn Reson Med 2002;47:1218 – 27.
[9] Weiger M, Pruessmann KP, Leussler C, Roschmann P, Boesiger P.
Specific coil design for SENSE: a six-element cardiac array. Magn
Reson Med 2001;45:495 – 504.
[10] Lin FH, Kwong KK, Belliveau JW, Wald LL. Parallel imaging re-
construction using automatic regularization. Magn Reson Med 2004;
51:559 – 67.
[11] Liang ZP, Bammer R, Ji J, Pelc N, Glover G. Making better SENSE:
wavelet denoising, Tikhnov regularization, and total least squares.
Proceedings of 10th Annual Meeting of ISMRM, 2002.
[12] Salem KA, Lewin JS, Aschoff AJ, Duerk JL, Wilson DL. Validation
of a human vision model for image quality evaluation of fast
interventional magnetic resonance imaging. J Electron Imaging 2002;
11:224 – 35.
[13] Wilson DL, Salem KA, Huo D, Duerk JL. Perception difference
paradigm for analyzing image quality of fast MRI techniques.
Proceedings of SPIE Medical Imaging on Image Perception, Observer
Performance, and Technology, 2003.
[14] Daly S. The visual differences predictor: an algorithm for the
assessment of image fidelity. In: Watson AB, editor. Digital images
and human vision. Cambridge (MA)7 MIT Press; 1993. p. 179 – 206.
[15] Flask CA, Salem KA, Lewin JS, Duerk JL. Keyhole Dixon method for
faster fat suppression with perceptual equivalence. Radiology 2002;
225:314.
[16] Mannos JL, Sakrison DJ. The effects of a visual fidelity criterion on
the encoding of images. IEEE Trans Info Theory 1974;IT-20:525 – 36.
[17] McCann JJ, McKee SP, Taylor TH. Quantitative studies in retinex
theory: a comparison between theoretical predictions and observer
132 D. Huo et al. / Magnetic Resonance Imaging 24 (2006) 123 – 132
responses to the bColor MondrianQ experiments. Vision Res 1976;
16:445 – 58.
[18] Watson AB. The cortex transform — rapid computation of simu-
lated neural images. Comput Vis Graph Image Process 1987;39:
311 – 27.
[19] Normann RA, Baxter BS, Ravindra H, Anderton PJ. Photoreceptor
contributions to contrast sensitivity — applications in radiological-
diagnosis. IEEE Trans Syst Man Cybern 1983;13:944 – 53.
[20] Martens JB, Meesters L. Image dissimilarity. Signal Process 1998;
70:155 – 76.
[21] Macovski A. Noise in MRI. Magn Reson Med 1996;36:494 – 7.
[22] King KF, Angelos L. SENSE image quality improvement using
matrix regularization. Proceedings of 9th Annual Meeting of
ISMRM, 2001.
[23] Jakob PM, Griswold MA, Edelman RR, Sodickson DK. AUTO-
SMASH: a self-calibrating technique for SMASH imaging. Si-
Multaneous Acquisition of Spatial Harmonics. MAGMA 1998;7:
42 – 54.
[24] Liang ZP, Stenger A, Ji JX, Ma J, Boada F. Parallel generalized series
imaging. Proceedings of 11th Annual Meeting of ISMRM, 2003.
[25] Hansen PC. Analysis of discrete ill-posed problems by means of the
L-curve. Siam Rev 1992;34:561 – 80.
[26] Wahba G, Wang YH. When is the optimal regularization parameter
insensitive to the choice of the loss function. Commun Stat Theory
Methods 1990;19:1685 – 700.
[27] Ying L, Xu D, Liang ZP. On Tikhonov regularization for image
reconstruction in parallel MRI. Proceedings of 26th Annual Interna-
tional Conference of Engineering in Medicine and Biology Society,
2004;1:1056 – 9.
[28] International Telecommunication Union. ITU-R BT.500-11 Method-
ology for the subjective assessment of the quality of television
pictures; 2002. Available at http://www.itu.int.