Accelerated Proton Echo Planar Spectroscopic Imaging
(PEPSI) Using GRAPPA with a 32-Channel Phased-Array
Shang-Yueh Tsai,1Ricardo Otazo,2Stefan Posse,2,3Yi-Ru Lin,4Hsiao-Wen Chung,1
Lawrence L. Wald,5Graham C. Wiggins,5and Fa-Hsuan Lin5–7*
Parallel imaging has been demonstrated to reduce the encod-
ing time of MR spectroscopic imaging (MRSI). Here we inves-
tigate up to 5-fold acceleration of 2D proton echo planar spec-
troscopic imaging (PEPSI) at 3T using generalized autocalibrat-
ing partial parallel acquisition (GRAPPA) with a 32-channel coil
array, 1.5 cm3voxel size, TR/TE of 15/2000 ms, and 2.1 Hz
spectral resolution. Compared to an 8-channel array, the
smaller RF coil elements in this 32-channel array provided a
3.1-fold and 2.8-fold increase in signal-to-noise ratio (SNR) in
the peripheral region and the central region, respectively, and
more spatial modulated information. Comparison of sensitivity-
encoding (SENSE) and GRAPPA reconstruction using an
8-channel array showed that both methods yielded similar
quantitative metabolite measures (P > 0.1). Concentration val-
ues of N-acetyl-aspartate (NAA), total creatine (tCr), choline
(Cho), myo-inositol (mI), and the sum of glutamate and glu-
tamine (Glx) for both methods were consistent with previous
studies. Using the 32-channel array coil the mean Cramer–Rao
lower bounds (CRLB) were less than 8% for NAA, tCr, and Cho
and less than 15% for mI and Glx at 2-fold acceleration. At
4-fold acceleration the mean CRLB for NAA, tCr, and Cho was
less than 11%. In conclusion, the use of a 32-channel coil array
and GRAPPA reconstruction can significantly reduce the mea-
surement time for mapping brain metabolites.
Med 59:989–998, 2008. © 2008 Wiley-Liss, Inc.
Key words: proton echo planar spectroscopic imaging; PEPSI;
MR spectroscopic imaging; parallel MRI; 32-channel phase ar-
MR spectroscopic imaging (MRSI) plays important roles in
both clinical diagnosis and biomedical research. One of
the main challenges of the conventional MRSI techniques
is the lengthy data acquisition time, a result of the many
phase-encoding steps required for complete spatial encod-
ing. Several methods have been proposed to reduce scan-
ning time using reduced or weighted k-space acquisition
(1). Other methods acquire multiple (typically two to four)
individually phase-encoded spin echoes within a single
RF excitation to reduce encoding time (2). However, be-
cause the acquisition of multiple echoes requires short-
ened echo spacing, such a method is characterized by a
limited spectral resolution. Alternatively, it is possible to
acquire all the spatial information in a single shot using
fast imaging readout modules, and to encode the spectral
information by incrementing the spectral evolution time in
separate RF excitations (3–6). In this way, spatial resolu-
tion is independent of scanning time, thus high spatial
resolution can be achieved. However, the disadvantage of
such approaches is the time-consuming spectral encoding
process required to achieve high spectral resolution and
bandwidth. Proton echo planar spectroscopic imaging
(PEPSI) (7–9) uses an oscillating readout gradient to simul-
taneously acquire spatial and spectral information in a
single RF excitation. PEPSI yields spectral resolution that
approximates that of conventional MRSI and enables a
reduction in the encoding time by a factor of more than an
order of magnitude. PEPSI was developed for clinical MR
3D metabolite distributions in several minutes (10,11); the
technique has been employed in clinical studies (12,13).
The development of parallel MRI techniques (14–17)
has enabled a significant reduction in imaging time, as
phase-encoding steps can be partially replaced by the spa-
tial information inherent in a multiple-channel receiver
coil array with nonuniform spatial sensitivities. Parallel
MRI techniques can be used in combination with either
standard MRSI (18,19) or fast MRSI (20) to reduce data
acquisition time. We have recently demonstrated that the
use of a combined PEPSI and a sensitivity-encoding
(SENSE) technique with an 8-channel coil array achieves
sub-minute MRSI data acquisition at the cost of reduced
signal-to-noise ratio (SNR) (21). Two major factors limit
the acceleration factor attainable for parallel MRI: the
number of receiver channel coils and the SNR of the MR
acquisitions. By using a coil array with a large number of
channels, we may further improve the temporal resolution
and the SNR of PEPSI in a parallel MRI strategy. The
enhanced SNR made possible by an array of smaller RF
coil elements enables tessellation of the cortical coverage,
1Department of Electrical Engineering, National Taiwan University, Taipei,
2Departments of Electrical & Computer Engineering, University of New Mex-
ico, Albuquerque, New Mexico, USA.
3Department of Psychiatry, University of New Mexico School of Medicine,
Albuquerque, New Mexico, USA.
4Department of Electronic Engineering, National Taiwan University of Science
and Technology, Taipei, Taiwan.
5MGH-HMS-MIT Athinoula A. Martinos Center for Biomedical Imaging,
Charlestown, Massachusetts, USA.
6Department of Radiology, Massachusetts General Hospital, Boston, Massa-
7Institute of Biomedical Engineering, National Taiwan University, Taipei, Tai-
Grant sponsor: National Institutes of Health; Grant numbers: R01 HD040712,
R01 NS037462, R01 EB000790, P41 RR14075; Grant sponsor: Mental Illness
and Neuroscience Discovery (MIND) Institute; Grant sponsor: Ministry of
Education (international student exchange program to S.Y.T.).
*Correspondence to: Fa-Hsuan Lin, PhD, Athinoula A. Martinos Center for
Biomedical Imaging, Bldg. 149, 13th St., Mail code 149-2301, Charlestown,
MA, 02129. E-mail: email@example.com; Institute of Biomedical Engi-
neering, National Taiwan University, 1, Sec. 4, Roosevelt Rd., 106 Taipei,
Taiwan. E-mail: firstname.lastname@example.org.
Received 9 February 2007; revised 9 November 2007; accepted 8 December
Published online in Wiley InterScience (www.interscience.wiley.com).
Magnetic Resonance in Medicine 59:989–998 (2008)
© 2008 Wiley-Liss, Inc.
and the acquisition of more spatially disparate information
from different channels in the array improves image re-
construction of parallel MRI. The further reduction in scan
time reduces motion sensitivity and thus offers the poten-
tial benefit of reducing motion artifacts for patients who
are incapable of remaining still for lengthy scans. PEPSI
has been employed for time-resolved metabolic imaging to
characterize metabolic dysfunction during sodium-lactate
infusion in patients with panic disorder (12), and to detect
the elevated lactate level in children during an auditory-
based language task (22). Highly accelerated MRSI with
better temporal resolution may also facilitate monitoring
of a possible transient metabolic response to time-locked
stimuli using multiple trials.
Recently, coil arrays have been designed with as many
as 32 channels to accelerate structural and functional MRI
(23–25), and the 32-channel array has been demonstrated
to improve SNR by a factor of 3.5 in the lateral cortex and
a factor of 1.5 in the corpus callosum when compared with
an 8-channel head coil consisting of eight surface coils cov-
32-channel array can reduce noise amplification (g-factor) by
1.5-fold in 1D parallel MRI at 4-fold acceleration (25).
To date, the most widespread parallel MRI techniques
are the sensitivity encoding (SENSE), which works in the
image domain, and the generalized partial parallel acqui-
sition (GRAPPA), which works in k-space. Both methods
can provide good results with nearly identical reconstruc-
tion quality, and they are available for clinical routine (26).
We have demonstrated that SENSE can be applicable for
PEPSI acceleration (21) and we chose GRAPPA reconstruc-
tion in this study because an explicit estimation of coil
sensitivity profiles is not required.
Here we report the results of our in vivo experiments
employing a 32-channel head coil array at 3T (25) to in-
vestigate the feasibility of accelerating data acquisition up
to 5-fold along a single phase-encoding direction with the
use of PEPSI and GRAPPA reconstruction (16). We also
compared and evaluated the reconstruction performance
of GRAPPA and SENSE using an 8-channel coil array. This
comparison has not been performed yet for MRSI and
presents differences to structural image reconstruction due
to the low-resolution characteristics. The dominant cere-
bral metabolites, including singlets of N-acetyl-aspartate
(NAA), total creatine (tCr), including creatine (Cr) and
phosphocreatine (PCr), and choline (Cho), as well as com-
plex multiplets of myo-inositol (mI) and the combination
of glutamate and glutamine (Glx) were measured and
quantified in accelerated MRSI experiments at 3T.
MATERIALS AND METHODS
Experiments on both human subjects and a spectroscopy
phantom were performed using a 3T scanner (Siemens
MAGNETOM Trio, Siemens Medical Solutions, Erlangen,
Germany) equipped with an 8-channel head coil array
(Siemens Medical Solutions) that covers the entire head
circumferentially with eight surface coils, and a 32-chan-
nel head coil array (25). The 32-channel array consists of
small circular receiver coils overlapped to minimize the
mutual inductance between neighboring coils. Two types
of coil elements (8.5 cm and 6 cm diameters) were ar-
ranged evenly to wrap around the whole head on a close-
fitting fiberglass helmet modeled after the European stan-
dard head norm EN960/1994 for protective headgear (25).
The array is referred to as having a soccer ball geometry. In
this study the discussion of the 32-channel coil array is
specific to this coil geometry.
The PEPSI pulse sequence used in this study included
water suppression by chemical shift selective saturation
(CHESS) and a WET (water suppression enhanced through
T1effects) technique; eight slices of outer volume lipid
suppression were applied along the perimeter of the brain
(Fig. 1), with spin-echo excitation and fast spatial-spectral
encoding using an EPI readout gradient train along the
X-axis. We acquired data at a bandwidth of 83.33 kHz,
with 1024 gradient inversions and online regridding to
account for gradient ramp sampling. Even-echo and odd-
echo data were separately reconstructed as described pre-
viously (11). We applied a sinusoidal k-space filter to both
nonwater-suppressed (NWS) and water-suppressed (WS)
data to suppress side lobes of the point-spread function.
The spectral width after even/odd echo editing was
1087 Hz, with 512 complex points. Phase encoding gradi-
ents were applied along the Y-axis to obtain 2D spatial
encoding. In order to employ the GRAPPA technique for
acceleration we reduced the phase encoding steps along
the Y-axis by sampling one k-space line in a block of two,
three, four, or five consecutive k-space lines, allowing us
to achieve 2-, 3-, 4-, and 5-fold accelerations, respectively.
We acquired both WS and NWS data; NWS data were
collected without presaturation and were used for auto-
matic phase and frequency shift corrections.
We acquired fully sampled and accelerated single-aver-
age in vivo PEPSI data from a para-axial slice at the upper
edge of the ventricles using the following parameters:
TR ? 2 sec, TE ? 15 ms, 32 ? 32 spatial matrix, FOV ?
240 mm, slice thickness ? 15 mm, and voxel size ?
0.85 cm3(Fig. 1). After application of the spatial sinusoidal
filter, the voxel size was 1.9 cm3. The data acquisition time
was 64 sec for a fully sampled dataset; for the accelerated
in vivo data with 2-, 3-, 4-, and 5-fold GRAPPA, acquisi-
tion times were 32, 22, 16, and 12 sec, respectively. In
addition to scanning human subjects, we scanned a spher-
ical spectroscopic phantom containing a physiological
mixture of metabolite solutions (Cho, Cr, NAA, mI, gluta-
mate, and lactate) with concentrations similar to those
found in human brain in vivo. Only fully sampled data
were collected for the phantom, using the parameters:
TR/TE ? 2000/15 ms, matrix size ? 32 ? 32, FOV ?
180 mm, and slice thickness ? 15 mm. We performed
simulations of parallel imaging by discarding k-space data
in the KYdirection, allowing us to investigate SNR degra-
dation in the absence of possible complicating factors such
as gradient heating-induced frequency drifts. To compare
the SNRs of the 8-channel and 32-channel arrays, we also
measured the fully sampled PEPSI data using an 8-channel
head array provided by the manufacturer.
GRAPPA Reconstruction and Spectral Postprocessing
Accelerated PEPSI data were reconstructed using a stan-
dard GRAPPA algorithm (16). GRAPPA reconstruction is
990Tsai et al.
based on the generation of spatial harmonics along the
acceleration axis KYusing information from spatially lo-
calized coils. We generated uncombined images from each
channel of the coil array by applying the reconstruction
coefficients to the accelerated data in multiple block-wise
reconstructions. The reconstruction coefficients were esti-
mated from the reference data, termed the automatic cali-
bration signal (ACS) lines, which can be acquired from a
separate reference scan or from accelerated scans with
central k-space lines sampled at the Nyquist rate. Because
PEPSI requires NWS reference data for automatic phasing,
frequency shift, and eddy current corrections, fully sam-
pled NWS reference data were used as the ACS lines in
this study. All 32 phase-encoding steps in the NWS image
were employed as ACS lines to estimate the reconstruction
coefficients; the coefficients were then repetitively applied
to reconstruct the entire accelerated WS data using a re-
construction kernel in a 4 by 3 (KY by KX) block. The
reconstruction algorithm of SENSE has been described in a
previous study (21).
We separately applied automatic zeroth-order spectral
phase correction and eddy current correction based on the
NWS scan to the GRAPPA-reconstructed even and odd
data from the individual coils. We then combined both
echoes from each coil channel individually and averaged
all channels to produce the spectroscopic images. The
phase correction performed before combining data from
individual coils was expected to reduce possible artifacts
caused by partial phase cancellation and allowed phase-
coherent complex data combination (27). In vivo localized
spectra were quantified with LCModel-based spectral fit-
ting (28), the range of which was 0.5–4 ppm. LCModel
allowed analysis of the in vivo spectra as a linear combi-
nation of individual in vitro metabolite spectra, which
included simulated macromolecules and lipid compo-
nents. Metabolite concentrations of NAA, tCr, Cho, mI,
and Glu?Gln were obtained using the water-scaling
method (28). Metabolite concentration values were com-
puted without partial volume and relaxation correction,
and are thus higher than the actual concentration values.
All reconstructions were carried out on a regular personal
computer in the Matlab programming environment (Math-
Works, Natick, MA). The computation time of GRAPPA
reconstruction was 10 min for one acceleration rate and
spectral fitting can be finished in 2 min.
Spectral and Error Analysis
For the in vitro phantom experiments, the SNR of the NAA
peak was evaluated by applying the following formula to
both the fully sampled and the accelerated data.
FIG. 1. Localization of the PEPSI experiments and the spatial position for outer volume saturation bands. The central black box indicates
the location of the multivoxel spectra. The black squares indicate the locations of voxels for representative spectra shown in Fig. 3. 6 ? 6
multivoxel spectra were cut from the fully sampled data and the 2-, 3-, 4-, and 5-fold GRAPPA accelerated data.
32-Channel GRAPPA PEPSI991
Here, nsignaland nnoiserepresent the numbers of spectral
points in the metabolite and in the noise ranges, respec-
tively. For the metabolic signals the range was defined in
intervals of 0.1 ppm (symmetrical around the peak maxi-
mum), whereas the range for noise was from 7.5–8.5 ppm.
sj indicates the reconstructed spectrum with spectral in-
dex j. Normalized SNR (nSNR), the ratio between the SNR
of GRAPPA-accelerated data and the SNR of fully sampled
data listed below, was used to evaluate the reconstruction
The SNR and nSNR were calculated at two regions of
interest (ROIs) selected from the peripheral and central
regions in the phantom, respectively (Fig. 2f).
For in vivo data, the error metric used to quantify me-
tabolite concentration in LCModel was the Cramer–Rao
lower bound (CRLB), the lowest bound of the standard
deviation (SD) of the estimated metabolite concentration.
Expressed in concentration percentage, CRLB can function
as an indicator of the reliability for metabolic concentra-
tion quantification (28). The CRLB of each metabolite is
commonly used to quantify the goodness-of-fit of LCModel
(29). Typically, the metabolite concentrations quantified
by LCModel with a CRLB of less than 20% are considered
Reconstruction errors in the metabolite concentrations
were also evaluated in a pixel-by-pixel manner based on
the difference in concentrations between the fully sampled
and the accelerated data:
RMS error (%) ???
where Cfullis the concentration from the fully sampled
data, CRis the concentration from the accelerated data after
GRAPPA reconstruction, and N is the total number of
image voxels in the ROI. The subscript i indicates the
spatial location of the image voxel. To present the metab-
olite maps we used the following thresholds to reject spec-
troscopic image voxels that could not be fitted with satis-
factory accuracy by LCModel: 1) CRLB ?50% and 2) spec-
concentration maps were interpolated to a 128 ? 128
matrix using zero-filling to improve visualization. To in-
vestigate regional differences in the fitting performance,
two ROIs, one in the white matter and one in the gray
matter, were selected manually on the localization image.
In vitro phantom experiments using the 32-channel array
yielded SNR values 3.1- and 2.8-fold greater than those
obtained with the 8-channel array in the peripheral region
and the central region, respectively (Table 1). Normalized
SNR (nSNR) values demonstrated that the soccer ball coil
geometry of the 32-channel coil arrays used in our study
reduced SNR loss in GRAPPA reconstructions compared
with the 8-channel coil array, especially at high accelera-
tion rates. At 4-fold acceleration, nSNR with the 32-chan-
nel array was around 0.2 and with the 8-channel array
around 0.1 (Table1). We did observe higher SNR at the
peripheral region in the fully sampled data and the SNRs
generally reduced at higher acceleration rates. SNR be-
came similar in the peripheral and the central regions at
acceleration rates over 3 (Table 1). However, the SNR maps
did not show spatially related SNR reductions signifi-
cantly (Fig. 2).
In Vivo Experiments Using the 32-Channel Array Coils
Figure 1 shows the spectra from the 6 ? 6 grids at the
central region of the brain. The spectra at 2-fold and 3-fold
FIG. 2. SNR maps of (a) the fully sampled and the (b) 2-, (c) 3-, (d)
4-, and (e) 5-fold GRAPPA accelerated data in the phantom exper-
iments. f: The shaded areas in the mask depict the ROIs selected
from the peripheral and the central regions used in Table 1.
Averaged SNR and Normalized SNR (nSNR) of the Phantom Experiments Measured over Selected ROIs Using a 32-Channel Array and
an 8-Channel Array
Compared to the 8-channel array, the 32-channel array offers an approximate SNR gain by a factor of three. We found comparable SNR
between 4-fold acceleration using a 32-channel array and 2-fold acceleration using an 8-channel array. The number of voxels in the
peripheral and the central ROIs (see Fig. 2) are 165 and 108, respectively.
992Tsai et al.
accelerations were similar in spectral quality to the fully
sampled spectra. Noise levels increased noticeably at
4-fold acceleration, but the three major metabolite peaks
(NAA, tCr, and Cho) remained clearly detectable. At R ? 5
most spectra were contaminated by the enhanced noise
levels except several voxels in the central region. Figure 3
shows the reconstructed spectra of two representative vox-
els located in the white matter (WM) and in the gray matter
(GM), displaying the spectral ranges of the metabolites
(1–4 ppm) and the noise (6–8 ppm). The noise levels were
small at R ? 2 and R ? 3, indicating that spectra acquired
at R ? 2 and R ? 3 were not significantly different from the
fully sampled data. In addition, reconstruction artifacts in
these two voxels did not significantly skew the shapes of
the baselines. The noise level increased at 4-fold acceler-
ation and the fitting of the three major metabolite peaks
(NAA, tCr, and Cho) were affected, which may lead to
higher errors in the quantification of metabolite concentra-
tions. At 5-fold acceleration NAA can still be clearly de-
tected, but the further increase in the noise level impaired
reliable quantification of the metabolic peaks.
3T in vivo metabolite maps of NAA, tCr, Cho, mI, and
Glx with their corresponding CRLB maps are shown in Fig.
4. Whereas NAA, Cho, and mI showed relatively uniform
spatial distribution, tCr displayed slight gray-white matter
contrast and the Glx maps showed strong gray-white mat-
ter contrast. For tCr and Glx, higher concentrations were
found in GM than in WM (Table 2), which is consistent
with our expectation (30). The concentration ratios be-
tween GM and WM (GM/WM) were 1.37 for tCr and 1.67
for Glx, respectively. A comparison between the fully sam-
pled and the GRAPPA-reconstructed data showed that re-
construction artifacts increased noticeably at R ? 4 for
NAA, tCr, and Cho, but the CRLBs for NAA, tCr, and Cho
at this acceleration rate were still lower than 20% (Table 2)
for over 95% of the voxels in brain tissue (Fig. 4b). At
5-fold GRAPPA acceleration the concentration maps of
NAA, tCr, and Cho showed strong reconstruction artifacts,
indicating that such further acceleration failed to preserve
the spatial uniformity of the metabolite information. Re-
construction artifacts in the reconstructed mI and Glx me-
tabolite maps increased substantially at 3-fold acceleration
(Fig. 4a). As a result, the CRLBs increased significantly
(Table 2) and there were over 10% voxels with CRLBs
larger than 20% (Fig. 4b). The averaged SNRs for the entire
slice were 27.7 for the fully sampled data, 17.7 for R ? 2,
11.2 for R ? 3, 8.1 for R ? 4, and 4.1 for R ? 5. The
whole-slice-averaged linewidths were around 0.057 ppm
for all acceleration rates, but the SDs of the linewidths
were 0.015 ppm for the fully sampled data, 0.017 for R ?
2, 0.017 for R ? 3, 0.018 for R ? 4, and 0.024 for R ? 5.
There was no significant difference in the CRLBs be-
tween GM and WM except for Glx. The CRLBs of Glx were
higher in WM than in GM (Table 2), in accordance with the
regional concentration difference. The root-mean-square
(RMS) errors between the GRAPPA-reconstructed and the
fully sampled data are summarized in Table 2. RMS errors
were similar between GM and WM for all metabolites
except for Glx, which showed higher errors in WM than in
GM (Table 2). In general, both RMS errors and CRLBs
increased with the acceleration rate. The whole-slice-av-
eraged RMS errors were 18.65%, 17.9%, and 27.84% for
NAA, tCr, and Cho at R ? 4. For mI and Glx those at R ?
2 were 18.65% and 23.98%, respectively.
In Vivo Experiments Using the 8-Channel Array Coils
SENSE and GRAPPA reconstructions showed similar spa-
tial uniformity (Fig. 5). The whole-slice-averaged CRLBs
FIG. 3. Two representative spec-
tra from the fully sampled data
and the 2-, 3-, 4-, and 5-fold
GRAPPA accelerated data (top to
bottom). Spectra were selected
from the voxels in the white mat-
ter at the right hemisphere (left)
and in the gray matter at the left
lateral area (right), as indicated in
Fig. 1. For each spectrum we
show the spectral ranges of the
metabolites (1–4 ppm) and the
noise (6–8 ppm).
32-Channel GRAPPA PEPSI993
(?SD) at R ? 2 were 9.7 (?4.11)% for NAA, 8.4 (?2.62)%
for tCr, 9.4 (?2.42)% for Cho in GRAPPA, and 8.2 (?3.3)%
for NAA, 7.5 (?3.2)% for tCr, 7.9 (?4.5)% for Cho in
SENSE. GRAPPA seemed to have higher averaged CRLBs
than SENSE but overall the CRLBs were still in an accept-
able range (?20%). The concentration difference between
SENSE and GRAPPA reconstructions was not statistically
significant for any of the metabolites (P ? 0.1) (Fig. 6).
Metabolite maps from SENSE and GRAPPA reconstruc-
tions were quantitatively similar. Averaged RMS errors
over ROIs defined in GM, WM, and across the whole slice
are summarized in Fig. 6. No significant difference in RMS
errors was found between the data reconstructed by
SENSE and GRAPPA. RMS errors were higher in WM than
in GM at 3-fold acceleration except for Cho using GRAPPA
reconstruction, which was consistent with the increased
reconstruction artifacts and the CRLBs found in WM ver-
sus those found in GM (Fig. 5).
FIG. 4. Metabolite concentration maps (a)
and CRLB maps (b) for NAA, tCr, and Cho in
vivo from the fully sampled PEPSI data and
the 2-, 3-, 4-, and 5-fold GRAPPA acceler-
ated data. The concentrations were quanti-
fied by LCModel. Values shown at the upper
left corner of each CRLB map denote the
number of voxels with CRLBs over 20%.
994 Tsai et al.
The results of this study demonstrate the feasibility of
combining the GRAPPA parallel MRI technique with
PEPSI to further reduce the data acquisition time required
for metabolite mapping in the human brain using a 32-
channel coil array at 3T. The improved temporal resolu-
tion achieved with this approach relies on the intrinsic
SNR gains afforded by the smaller RF coil elements in the
array and their closer proximity to the head, as well as by
the coil geometry, which yields more spatially distinct
information from the different channels. In our previous
study using an 8-channel coil array at 3T we found that
only 2-fold acceleration with 32 sec of measurement time
was feasible for maintaining acceptable spectral quality
(21). Our present results show that the mean CRLBs were
less than 11% for NAA, tCr, and Cho at 4-fold acceleration
at 3T with the use of a 32-channel array, such that the scan
time for a single signal-averaged short TE PEPSI experi-
ment with the 32 ? 32 image matrix, 24 cm FOV, 1.9 cm3
voxel size can be reduced from 64 sec to 16 sec. Such a
reduction in the data acquisition time may enable novel
MRSI applications. For example, highly accelerated MRSI
can be used in hyperpolarized
transiently high SNR (31). In addition to improving tem-
poral resolution, partial parallel imaging combined with
MRSI is particularly effective for reducing the long scan-
ning time for 3D spatial encoding. As demonstrated here,
1D acceleration by GRAPPA can be utilized to accelerate
slow spatial phase encoding in a 2D PEPSI experiment.
Further accelerations can be achieved with 2D GRAPPA
acceleration in 3D PEPSI experiments in two orthogonal
spatial phase-encoding directions. In fact, the spherically
symmetric layout of this 32-channel head RF array coil
design encourages the use of 2D acceleration.
Here, the achievable GRAPPA acceleration rate was
evaluated qualitatively by the spatial variability (homoge-
neity) in the metabolite maps, and quantitatively by the
CRLB and RMS errors in the metabolite concentrations
estimated by LCModel. In general, the CRLB from LC-
Model indicates spectral quality that combines informa-
tion about SNR, spectral line width, and spectral line
shape (28). With our short TE imaging protocol using
1.9 cm3nominal spatial resolution (2.1 Hz in spectral
resolution and TR/TE of 2000/15 ms), the mean CRLBs are
13C experiments for its
Average and Standard Deviations of LCModel-Quantified Metabolite Concentrations (mM)
14.46 ? 1.79
8.71 ? 0.98
3.19 ? 0.45
6.08 ? 0.83
11.01 ? 3.20
15.36 ? 2.62
11.97 ? 1.67
3.19 ? 0.56
6.11 ? 1.14
18.40 ? 3.19
6.58 ? 1.83
5.23 ? 0.72
5.16 ? 0.75
7.47 ? 1.18
12.58 ? 5.20
4.37 ? 1.40
4.51 ? 0.85
5.20 ? 0.80
7.69 ? 2.00
8.00 ? 1.78
15.30 ? 1.77
9.16 ? 1.02
3.60 ? 0.63
6.57 ? 1.51
12.47 ? 4.01
15.93 ? 3.27
11.93 ? 2.28
3.10 ? 0.81
6.71 ? 1.47
17.38 ? 4.76
7.05 ? 1.90
6.72 ? 0.73
6.33 ? 0.81
10.98 ? 2.61
14.74 ? 4.12
5.77 ? 2.34
5.86 ? 1.26
7.20 ? 1.80
10.71 ? 3.48
12.86 ? 6.50
15.46 ? 2.17
10.25 ? 2.17
3.74 ? 0.72
6.53 ? 2.29
12.91 ? 4.48
15.93 ? 3.12
11.94 ? 2.35
3.33 ? 0.80
6.42 ? 2.11
17.49 ? 5.11
7.98 ? 2.15
7.51 ? 1.30
7.35 ? 1.15
15.21 ? 6.36
17.88 ? 8.16
6.14 ? 1.78
6.71 ? 1.18
8.20 ? 2.36
14.71 ? 10.74
15.89 ? 6.15
15.36 ? 2.73
9.43 ? 2.38
3.85 ? 1.33
6.14 ? 2.63
13.49 ? 5.69
16.63 ? 4.49
11.76 ? 3.48
3.59 ? 1.43
7.42 ? 3.22
19.06 ? 8.97
10.65 ? 3.27
9.77 ? 1.97
9.67 ? 1.81
24.33 ? 18.50
23.62 ? 11.00
8.11 ? 5.14
8.83 ? 4.44
9.97 ? 3.75
16.29 ? 15.76
22.03 ? 16.56
15.61 ? 5.77
9.82 ? 4.15
5.25 ? 2.90
5.98 ? 3.56
13.60 ? 10.48
14.38 ? 6.96
12.13 ? 5.48
3.79 ? 1.78
8.03 ? 4.05
14.15 ? 9.36
17.23 ? 12.01
16.79 ? 10.48
13.93 ? 8.40
40.34 ? 37.10
68.13 ? 74.26
17.43 ? 28.78
13.24 ? 7.80
17.62 ? 19.84
25.91 ? 21.60
51.42 ? 69.23
RMS Error (%)R?1R?2R?3R?4R?5
Fitting errors (CRLB in %) and root-mean-square (RMS) errors in metabolite concentrations (%) of NAA, total creatine (tCR), choline (Cho),
myo-inositol (mI), and glutamate ? glutamine (Glx) in vivo for the fully sampled data and for 2-, 3-, 4-, and 5-fold GRAPPA accelerations.
Values were calculated from the ROIs defined in the white matter (WM) and the gray matter (GM) with 60 and 58 voxels, respectively.
32-Channel GRAPPA PEPSI995
less than 20% and RMS errors are less than 30% for NAA,
tCr, and Cho at 4-fold acceleration and for mI and Glx at
2-fold acceleration. However, the maximal acceleration
rate should be adjusted to obtain the SNR required for a
particular MRSI experiment. It should be noted that the
CRLB decreases as SNR increases but the CRLB increase is
not expected to be linearly proportional to the SNR reduc-
We did not find significant spatial dependence in SNR
reductions in the phantom experiments and in the in vivo
experiments. For Glx there are significantly higher RMS
errors and CRLBs in WM than in GM. Since the fitting of
Glx is known to be very sensitive to the noise level, this
result can be explained by the lower concentrations of Glx
in WM. This finding is also consistent with the higher
intrasubject variations of Glx concentrations in WM
(28%). For the other metabolites quantified here the intra-
subject variations are less than 20% for all other metabo-
lites, which is consistent with values reported in our pre-
vious study (30).
Two metabolites, tCr and Glx, show strong GM/WM
contrast, which is consistent with the findings from previ-
ous reports (30,33). The GM/WM ratios in our study are in
the reasonable ranges, although a little lower than the
reported values (1.6 ? 0.2 for tCr and 1.8 ? 0.4 for Glx) due
to lack of partial volume correction. The concentration
values derived here are higher than those reported in pre-
vious studies (34,35) due to the lack of partial volume and
FIG. 5. Metabolite concentration maps (a)
and CRLB maps (b) for NAA, tCr, and Cho
from data acquired with an 8-channel phase
array coil on another subject. GRAPPA and
SENSE reconstructions were performed for
2- and 3-fold accelerated data.
FIG. 6. RMS errors of NAA, tCr, and Cho concen-
tration values from the data acquired with an
8-channel phase array coil. The RMS values are
the averages over ROIs defined in WM (66 voxels),
in GM (71 voxels), and across the whole slice of the
metabolite maps shown in Fig. 5. The numbers in
brackets are P-values for the statistical signifi-
cance of differences in concentrations between
SENSE and GRAPPA.
996 Tsai et al.
relaxation corrections but in agreement with those mea-
sured in our previous works using the 8-channel coil (21).
The SDs of the averaged concentrations across subjects are
less than 15% for all metabolites. The intersubject coeffi-
cients of variations are thus consistent with previous stud-
ies (30,36,37) (Table 3). Recently, it has been demonstrated
that the PEPSI technique combined with relaxation and
partial volume corrections provide metabolic concentra-
tions that are consistent with previous studies (30).
According to our results by comparing the metabolite
concentrations, CRLBs, and RMS errors, the difference
between GRAPPA and SENSE reconstructions is not sta-
tistically significant. The autocalibrating technique is
known to help improve the reconstruction of spatial har-
monics-based parallel MRI methods such as GRAPPA. In
this study we implemented the standard GRAPPA algo-
rithm, where the reconstruction kernel was estimated from
the fully sampled NWS data. Because NWS data are ac-
quired for automatic phasing, frequency drift, and eddy
current corrections, using the fully sampled NWS PEPSI
data as ACS lines is an option that requires no more
scanning time than standard PEPSI. Alternatively, a faster
reference dataset can be collected using the PEPSI se-
quence with short TR to complete data acquisition in
several seconds. Park et al. (38) recently compared image
quality using different reconstruction kernels in GRAPPA
reconstruction. They noted that the separation of high-
frequency and low-frequency k-space ACS lines to esti-
mate the reconstruction kernels improves reconstruction
quality. Because the low-frequency k-space signals tend to
more seriously perturb the reconstruction process, using
the same kernel for all k-space reconstructions is not op-
timal. This concept, which implies that GRAPPA PEPSI
can be further improved by separating the reconstruction
kernels used in the high- and low-frequency regions, is
currently being investigated. Our recent results also indi-
cate that GRAPPA calibration is sensitive to the noise
amplitude in the ACS lines, which will influence the re-
construction quality of the accelerated PEPSI data (39). On
the other hand, an incorporation of prior spectral informa-
tion may improve GRAPPA reconstruction (40). Because
some information, for instance, the spectral pattern, is
known before acquisition, it may be possible to further
accelerate MRSI by incorporating such prior information
for reconstruction (41).
In this study we demonstrated that the SNR improve-
ments associated with the use of a head array with a larger
number of small receiver coil elements in a close-fitting
helmet design benefit PEPSI spectral quality and acquisi-
tion acceleration. The SNR gains observed can also be
attributed to the smaller size of the coil elements and their
close proximity to the head (25). This SNR improvement is
exchanged for a reduction in the acquisition time that can
be used for applications for anatomical and functional
imaging sequences. Acceleration also reduces motion sen-
sitivity, an improvement desirable for clinical studies. The
higher acceleration rate is achievable in part because of
this soccer-ball coil geometry, which provides more spa-
tially disparate information from different channels in the
array (Table 1). However, simply increasing the number of
array channels may not enable further reduction in scan-
ning time. Electromagnetic studies have shown that accel-
eration beyond 4-fold in one direction at 3T produces
dramatic decays in SNR due to fast increases in the geom-
etry factors (42). Nevertheless, the property of increased
SNR combined with reduced geometry factors of this 32-
channel coil array is still attractive to reduce the lengthy
acquisition time of 3D PEPSI using 2D acceleration, which
is under development (43). On the other hand, it has been
proven that PEPSI provides linear gains in sensitivity with
field strengths increased from 1.5T to 7T (44) and enables
the mapping of multiplet resonances at high field strength
(30). We expect that a short echo time PEPSI technique
with large-scale coil array at higher field strength could
further improve the performance of parallel PEPSI. How-
ever, the technique challenges for short echo time PEPSI at
high fields are 1) limitations in the gradient rise times will
limit the spectral width at high fields (45), and 2) eddy
currents from fast switching of the gradient coils lead to
line-shape distortions, particularly at short TE.
In conclusion, combining a 32-channel coil array with
GRAPPA reconstruction can improve the temporal resolu-
tion of MRSI experiments to reduce encoding time.
This work was supported by National Institutes of Health
Grants R01 HD040712, R01 NS037462, P41 RR14075, R21
EB007298, National Science Council—Taiwan (NSC 96-
2320-B-002-085), National Health Research Institute—Tai-
wan (E29C97N), and the Mental Illness and Neuroscience
Discovery Institute (MIND).
1. Maudsley AA, Matson GB, Hugg JW, Weiner MW. Reduced phase
encoding in spectroscopic imaging. Magn Reson Med 1994;31:645–
2. Duyn JH, Moonen CT. Fast proton spectroscopic imaging of human
brain using multiple spin-echoes. Magn Reson Med 1993;30:409–414.
3. Dreher W, Leibfritz D. Fast proton spectroscopic imaging with high
signal-to-noise ratio: spectroscopic RARE. Magn Reson Med 2002;47:
4. Ebel A, Dreher W, Leibfritz D. A fast variant of (1)H spectroscopic
U-FLARE imaging using adjusted chemical shift phase encoding. J
Magn Reson 2000;142:241–253.
Comparison of Intersubject Variations (%) of the Metabolite Concentrations between Different Studies
NAA tCr ChomIGlx
Schubert et al. (36)
Hurd et al. (37)
Posse et al. (30)
N denotes the number of subjects enrolled in the study.
32-Channel GRAPPA PEPSI 997
5. Guimaraes AR, Baker JR, Jenkins BG, Lee PL, Weisskoff RM, Rosen BR, Download full-text
Gonzalez RG. Echoplanar chemical shift imaging. Magn Reson Med
6. Adalsteinsson E, Irarrazabal P, Topp S, Meyer C, Macovski A, Spielman
DM. Volumetric spectroscopic imaging with spiral-based k-space tra-
jectories. Magn Reson Med 1998;39:889–898.
7. Mansfield P. Spatial mapping of the chemical shift in NMR. Magn
Reson Med 1984;1:370–386.
8. Posse S, DeCarli C, Le Bihan D. Three-dimensional echo-planar MR
spectroscopic imaging at short echo times in the human brain. Radiol-
9. Webb P, Spielman D, Macovski A. A fast spectroscopic imaging method
using a blipped phase encode gradient. Magn Reson Med 1989;12:306–
10. Posse S, Dager SR, Richards TL, Yuan C, Ogg R, Artru AA, Muller-
Gartner HW, Hayes C. In vivo measurement of regional brain metabolic
response to hyperventilation using magnetic resonance: proton echo
planar spectroscopic imaging (PEPSI). Magn Reson Med 1997;37:858–
11. Posse S, Tedeschi G, Risinger R, Ogg R, Le Bihan D. High speed 1H
spectroscopic imaging in human brain by echo planar spatial-spectral
encoding. Magn Reson Med 1995;33:34–40.
12. Dager SR, Friedman SD, Heide A, Layton ME, Richards T, Artru A,
Strauss W, Hayes C, Posse S. Two-dimensional proton echo-planar
spectroscopic imaging of brain metabolic changes during lactate-in-
duced panic. Arch Gen Psychiatry 1999;56:70–77.
13. Dager SR, Friedman SD, Parow A, Demopulos C, Stoll AL, Lyoo IK,
Dunner DL, Renshaw PF. Brain metabolic alterations in medication-
free patients with bipolar disorder. Arch Gen Psychiatry 2004;61:450–
14. Sodickson DK, Manning WJ. Simultaneous acquisition of spatial har-
monics (SMASH): fast imaging with radiofrequency coil arrays. Magn
Reson Med 1997;38:591–603.
15. Pruessmann KP, Weiger M, Scheidegger MB, Boesiger P. SENSE: sen-
sitivity encoding for fast MRI. Magn Reson Med 1999;42:952–962.
16. Griswold MA, Jakob PM, Heidemann RM, Nittka M, Jellus V, Wang J,
Kiefer B, Haase A. Generalized autocalibrating partially parallel acqui-
sitions (GRAPPA). Magn Reson Med 2002;47:1202–1210.
17. Sanchez-Gonzalez J, Tsao J, Dydak U, Desco M, Boesiger P, Pruessmann
KP. Minimum-norm reconstruction for sensitivity-encoded magnetic
resonance spectroscopic imaging. Magn Reson Med 2006;55:287–295.
18. Dydak U, Weiger M, Pruessmann KP, Meier D, Boesiger P. Sensitivity-
encoded spectroscopic imaging. Magn Reson Med 2001;46:713–722.
19. Zhao X, Prost RW, Li Z, Li SJ. Reduction of artifacts by optimization of
the sensitivity map in sensitivity-encoded spectroscopic imaging.
Magn Reson Med 2005;53:30–34.
20. Dydak U, Pruessmann KP, Weiger M, Tsao J, Meier D, Boesiger P.
Parallel spectroscopic imaging with spin-echo trains. Magn Reson Med
21. Lin FH, Tsai SY, Otazo R, Caprihan A, Wald LL, Belliveau JW, Posse S.
Sensitivity-encoded (SENSE) proton echo-planar spectroscopic imag-
ing (PEPSI) in the human brain. Magn Reson Med 2007;57:249–257.
22. Serafini S, Steury K, Richards T, Corina D, Abbott R, Dager SR,
Berninger V. Comparison of fMRI and PEPSI during language process-
ing in children. Magn Reson Med 2001;45:217–225.
23. de Zwart JA, Ledden PJ, van Gelderen P, Bodurka J, Chu R, Duyn JH.
Signal-to-noise ratio and parallel imaging performance of a 16-channel
receive-only brain coil array at 3.0 Tesla. Magn Reson Med 2004;51:
24. Hardy CJ, Cline HE, Giaquinto RO, Niendorf T, Grant AK, Sodickson
DK. 32-element receiver-coil array for cardiac imaging. Magn Reson
25. Wiggins GC, Triantafyllou C, Potthast A, Reykowski A, Nittka M, Wald
LL. 32-Channel 3 Tesla receive-only phased-array head coil with soc-
cer-ball element geometry. Magn Reson Med 2006;56:216–223.
26. Blaimer M, Breuer F, Mueller M, Heidemann RM, Griswold MA, Jakob
PM. SMASH, SENSE, PILS, GRAPPA: how to choose the optimal
method. Top Magn Reson Imaging 2004;15:223–236.
27. Zhu X, Ebel A, Ji JX, Schuff N. Spectral phase-corrected GRAPPA
reconstruction of three-dimensional echo-planar spectroscopic imag-
ing (3D-EPSI). Magn Reson Med 2007;57:815–820.
28. Provencher SW. Estimation of metabolite concentrations from local-
ized in vivo proton NMR spectra. Magn Reson Med 1993;30:672–679.
29. Srinivasan R, Sailasuta N, Hurd R, Nelson S, Pelletier D. Evidence of
elevated glutamate in multiple sclerosis using magnetic resonance
spectroscopy at 3 T. Brain 2005;128(Pt 5):1016–1025.
30. Posse S, Otazo R, Caprihan A, Bustillo J, Chen H, Henry PG, Marjanska
M, Gasparovic C, Zuo C, Magnotta V, Mueller B, Mullins P, Renshaw P,
Ugurbil K, Lim KO, Alger JR. Proton echo-planar spectroscopic imaging
of J-coupled resonances in human brain at 3 and 4 Tesla. Magn Reson
31. Golman K, Ardenkjaer-Larsen JH, Petersson JS, Mansson S, Leunbach I.
Molecular imaging with endogenous substances. Proc Natl Acad Sci
U S A 2003;100:10435–10439.
32. Kreis R. Issues of spectral quality in clinical 1H-magnetic resonance
spectroscopy and a gallery of artifacts. NMR Biomed 2004;17:361–381.
33. Srinivasan R, Cunningham C, Chen A, Vigneron D, Hurd R, Nelson S,
Pelletier D. TE-averaged two-dimensional proton spectroscopic imag-
ing of glutamate at 3 T. Neuroimage 2006;30:1171–1178.
34. Gasparovic C, Song T, Devier D, Bockholt HJ, Caprihan A, Mullins PG,
Posse S, Jung RE, Morrison LA. Use of tissue water as a concentration
reference for proton spectroscopic imaging. Magn Reson Med 2006;55:
35. Soher BJ, Hurd RE, Sailasuta N, Barker PB. Quantitation of automated
single-voxel proton MRS using cerebral water as an internal reference.
Magn Reson Med 1996;36:335–339.
36. Schubert F, Gallinat J, Seifert F, Rinneberg H. Glutamate concentrations
in human brain using single voxel proton magnetic resonance spectros-
copy at 3 Tesla. Neuroimage 2004;21:1762–1771.
37. Hurd R, Sailasuta N, Srinivasan R, Vigneron DB, Pelletier D, Nelson SJ.
Measurement of brain glutamate using TE-averaged PRESS at 3T. Magn
Reson Med 2004;51:435–440.
38. Park J, Zhang Q, Jellus V, Simonetti O, Li D. Artifact and noise sup-
pression in GRAPPA imaging using improved k-space coil calibration
and variable density sampling. Magn Reson Med 2005;53:186–193.
39. Rueckert M, Otazo R, Posse S. GRAPPA reconstruction of sensitivity
encoded 2D and 3D proton echo planar spectroscopic imaging (PEPSI)
with SNR adaptive recalibrating. In: Proc 14th Annual Meeting
ISMRM, Seattle; 2006. p 296.
40. Lin FH. Prior-regularized GRAPPA reconstruction. In: Proc 14th An-
nual Meeting ISMRM, Seattle; 2006. p 3656.
41. Lin FH, Kwong KK, Belliveau JW, Wald LL. Parallel imaging recon-
struction using automatic regularization. Magn Reson Med 2004;51:
42. Wiesinger F, Boesiger P, Pruessmann KP. Electrodynamics and ulti-
mate SNR in parallel MR imaging. Magn Reson Med 2004;52:376–390.
43. Otazo R, Tsai SY, Lin FH, Posse S. Accelerated short-TE 3D proton
echo-planar spectroscopic imaging using 2D-SENSE with a 32-channel
array coil. Magn Reson Med 2007;58:1107–1116.
44. Otazo R, Mueller B, Ugurbil K, Wald L, Posse S. Signal-to-noise ratio
and spectral linewidth improvements between 1.5 and 7 Tesla in
proton echo-planar spectroscopic imaging. Magn Reson Med 2006;56:
45. Ebel A, Maudsley AA, Weiner MW, Schuff N. Achieving sufficient
spectral bandwidth for volumetric 1H echo-planar spectroscopic imag-
ing at 4 Tesla. Magn Reson Med 2005;54:697–701.
998 Tsai et al.