High Speed
‘H
Spectroscopic Imaging in Human Brain
by
Echo Planar Spatial-Spectral Encoding
Stefan Posse, Gioacchino Tedeschi, Robert Risinger, Robert
Ogg,
Denis
Le
Bihan
We introduce a fast and robust spatial-spectral encoding
method, which enables acquisition of high resolution short
echo time (13 ms) proton spectroscopic images from human
brain with acquisition times as short as
64
s
when using
surface coils. The encoding scheme, which was implemented
on a clinical 1.5 Tesla whole body scanner, is a modification of
an echo-planar spectroscopic imaging method originally pro-
posed by Mansfield
Magn.
Reson.
Med.
1,370-386
(1984),
and
utilizes a series of read-out gradients to simultaneously en-
code spatial and spectral information. Superficial lipid signals
are suppressed by a novel double outer volume suppression
along the contours of the brain. The spectral resolution and
the signal-to-noise per unit time and unit volume from reso-
nances such as N-acetyl aspartate, choline, creatine, and
inositol are comparable with those obtained with conventional
methods. The short encoding time of this technique enhances
the flexibility of
in
vivo
spectroscopic imaging by reducing
motion artifacts and allowing acquisition of multiple data sets
with different parameter settings.
Key words: proton spectroscopy; spectroscopic imaging;
echo planar encoding; human brain.
INTRODUCTION
Proton spectroscopic imaging
(HSI)
is increasingly being
used to study human brain metabolism
(1-11).
Brain
tumors
(1,
6,
8,
lo),
multiple sclerosis
(5,
lo),
brain
infarction
(7),
epilepsy
(Z),
and acquired immunodefi-
ciency syndrome
(4)
manifest locally altered metabolite
levels. At the same time clinical applications of HSI have
been limited by the procedure’s length which tests pa-
tient tolerance and limits the integration of
HSI
with
diagnostic imaging. Recently, a method based on multi-
ple individually phase encoded spin echoes has been
proposed to reduce scan times
(12).
However, short echo
time measurements are not feasible with this technique,
because it uses echoes at different echo times. It also has
a limited spectral resolution and suffers from T,-weight-
ing in k-space. Methods using chemical shift encoded
fast scans have similar limitations (13-16).
Echo-planar spectroscopic imaging (EPSI), a much
faster method proposed by Mansfield and further devel-
MRM
33:34-40
(1995)
From the Diagnostic Radiology Department, The Warren Grant Magnuson
Clinical Center (S.P., D.L.B.). the Neuroimaging Branch, National Institute of
Neurological Disorders and Stroke (G.T.) and the Section on Clinical Phar-
macology, National Institute of Mental Health
(R.R.),
National Institutes of
Health, Bethesda, Maryland and St Jude Children’s Research Hospital,
Memphis, Tennessee
(R.O.).
Address correspondence to Stefan Posse, Ph.D, Institute of Medicine,
Research Center Julich GmbH,
P.O.
Box 1913, D-52425 Julich, Germany.
Received July 14, 1994; revised October 11, 1994; accepted October
12,
1994.
1994
SMR
Young Investigators’ Rabi Award Finalist.
0740-3194/95 $3.00
Copyright
0
1995 by Williams
&
Wilkins
All
rights of reproduction in any form reserved.
oped by others, avoids these limitations by using a series
of periodically inverted gradients to generate a train of
echoes that contain both spatial and spectral information
(17-23). However, initial implementations suffered from
limitations in gradient performance and from spectral
artifacts due to the convolution of spectral and spatial
information. Recently, we have improved the EPSI-based
acquisition technique by enhancing the orthogonaliza-
tion (separation) of the spectral and spatial information
within the hardware constraints of a clinical whole-body
scanner and demonstrated short echo time three-dimen-
sional spatial encoding in human brain in clinically fea-
sible acquisition times
(24).
Here, we investigate the feasibility
of
short echo time
(13 ms) two-dimensional EPSI in human brain with ac-
quisition times as short as
64
s
by using surface coils to
improve the signal-to-noise ratio (SNR). The spatial sup-
pression of strong lipid signals in the proximity of the
surface coil is improved by a modified pulse sequence
which permits localized tuning of the spatial suppres-
sion RF pulses. The
SNR
of N-acetyl aspartate (NAA) is
evaluated for different acquisition times, voxel sizes, and
locations with respect to the surface coil. We also eval-
uate gradient performance and robustness of the local-
ization scheme.
MATERIALS AND METHODS
For
echo-planar spatial-spectral encoding, a trapezoidal
read out gradient was inverted periodically to encode
k-space in a zig-zag trajectory (Fig.
1).
The solid line
trajectory corresponds to the original encoding scheme
proposed by Mansfield
(17)
where the length of each
trapezoid
(7)
corresponds to the inverse of the spectral
width in the reconstructed spectra. Reversal of the read-
out gradient in the presence of local magnetic field inho-
mogeneities and asymmetries in gradient switching in-
troduce periodicities in k-space that lead to aliasing
artifacts. Aliasing can be removed by separating (editing)
the echoes obtained with positive and negative gradients
at the expense of reducing the spectral width to
1/~~)
(17,
19,
20,
22).
Evolution in time convolves spectral and
spatial information which leads to chemical shift
arti-
facts. To simultaneously eliminate aliasing and to reduce
chemical shift artifacts while retaining the desired spec-
tral width (I/T), we employ spatial-spectral oversampling
as represented by the dotted-line trajectories in Fig.
1.
For example, when using an oversampling ratio
(n)
of
2,
the read-out gradient strength doubles, thus reducing
chemical shift artifacts twofold. After eventodd echo ed-
iting (Fig.
2)
and time reversal of echoes obtained with
negative gradients, the data sets are reconstructed sepa-
rately and the results are added together to maintain the
SNR.
34
High Speed Echo Planar Spectroscopic Imaging
RFro--l-:
35
-9-:
n-1
-
n=2
--
-w
n-4
___)
t
T
0
FIG.
1.
KY-t
space trajectories of
EPSl
acquisition methods
with
different spatial-spectral oversampling ratios
(n).
The time interval
T
determines the desired spectral
width
1h
in
the reconstructed
spectra. The
k,
domain, which
is
orthogonal to the
ky-t
plane,
is
not shown.
rkl
nd
&
7
7
7
t t
t
Gy
m
1234 1112212231324142
1114
24 34 44
RF
FIG.
2.
Echo
editing
scheme for an individual EPSl data trace
obtained
with
different oversampling ratios
(n).
With
two-fold over-
sampling
(n
=
2)
two echo trains from alternate echoes are ob-
tained. With four-fold oversampling
(n
=
4)
four echo trains are
obtained. Two-dimensional fourier transformation of these echo
trains provides a series of spatially localized spectra.
The method was implemented on a clinical 1.5 Tesla
whole body scanner
(GE
Medical Systems, Milwaukee,
WI)
with actively shielded gradients of
10
mT/m
strength. Spatial-spectral encoding with twofold over-
sampling was performed along the y-axis. At 1.5 Tesla a
spectral width of
488
Hz covers almost all signals that are
observable
in
vim
(between
0
and
7.64
ppm). The length
of each read-out gradient lobe was
1024
ps.
For spatial
resolutions between 5 and
10
mm, the read-out gradient
amplitudes were between
4.6
and
2.3
mT/m, and the
gradient ramp times were limited by hardware between
160
and
80
ps.
Due
to
the convolution of spatial and
spectral information, the water signal could be posi-
tioned on resonance and any signal outside of a fre-
quency range of
+/-244
Hz would alias back into the
spectrum. For each data trace
16384
complex data points
(=512 spectral points
X
32
spatial points) were sampled
continuously with a spectral width of
32
kHz to yield a
digital frequency resolution of 1.9 Hz.
No
gradient tuning
was required. The x-dimension was localized with con-
ventional
32
step phase encoding.
Volume prelocalization
of
an axial slice was obtained
by a three-pulse sequence to generate a stimulated echo
(Fig.
3) (24).
Surface lipid were suppressed by a series of
eight slice selective suppression pulses with gradient
dephasing (SSl), each of which suppressed a different
slice positioned along the brain contours and orthogonal
to the stimulated echo selected axial slice. The position
and orientation
of
each suppression slice could be ad-
justed individually. The series of suppression pulses was
repeated during the TM period
(SS2)
to improve lipid
suppression. For localized tuning of the suppression RF
pulses, the slice selective stimulated echo pulse se-
quence could be converted into a volume selective pulse
sequence to receive signal only from a selected suppres-
sion slice. This was accomplished by changing the ori-
entation of the slice selection gradient of the last RF
pulse (Fig.
3).
The residual signal from a selected sup-
pression slice could be observed in spectroscopy mode to
minimize residual lipid signals
or
in imaging mode for
spatially resolved tuning.
3
times
8
times
8
times
I
STEAM
I
Soatid
encodina
I
ws2
I
ss2
I
I
ws1
I
SSl
I
FIG.
3.
Water suppressed
EPSl
pulse sequence
with
volume pre-
localization.
An
axial slice
is
selected
by
three
RF
pulses (gray
symbols) to form a stimulated echo. Spatial suppression
(SS1,
SS2)
is
applied orthogonal to the axial slice to suppress superficial
lipid
signals. Eight spatial suppression pulses
with
subsequent
gradient dephasing are applied
during
each presaturation period
(SS1,
SS2)
in
different spatial orientations and locations along the
brain contours. Three CHESS water suppression
pulses
are ap-
plied
in
front
of
the prelocalization scheme
(WS1)
and one CHESS
pulse
is
applied
during
TM
(WS2).
Spatial localization
is
achieved
by
Echo-Planar spectral-spatial encoding along the y-axis and
phase encoding along the x-axis. For localized suppression slice
tuning,
the orientation of the slice selection gradient of the last
RF
pulse
is
changed (see Methods).
36
Posse
et
al.
Three chemical shift selective water suppression
pulses
(CHESS)
with
50
Hz bandwidth were applied
before the localization scheme. The flip angles and tim-
ing sequence of the water suppression pulses were nu-
merically optimized
(25).
A
fourth water suppression
pulse was applied during the TM period to further en-
hance water suppression. With this slice selective pulse
sequence, the gradient dephasing requirements during
the
TE/2
periods were strongly reduced with respect to
volume selective stimulated echo methods. Thus, the
motion sensitivity was similar to that of previously pub-
lished methods
(lo),
despite the longer TM period. Sig-
nal losses due to J-modulation and due to zero-quantum
modulations during TM were minimized by using short
echo times
(26).
Protocol
Measurements were performed on five normal volunteers
and on some patients with neuropsychiatric disorders.
Informed consent was obtained from all subjects prior to
the measurements according to institutionally reviewed
protocols. In these studies, different anatomical regions,
including occipital brain, frontal brain and cerebellum,
were investigated using a circular receive-only surface
coil
(12.7-cm
diameter) with body coil excitation. FDA
guidelines for
RF
power deposition were observed and
hearing protection was provided. Fast spin echo scans
(TE
85
ms,
TR:
2s,
8
echoes)
or
gradient recalled echo
scans
(TE: 5
ms, TR:
33
ms, flip angle:
30”)
were acquired
to locate the volume of interest. In some cases, we used
automatic higher order shimming based on phase sensi-
tive gradient echo imaging, as described by Webb and
Macowski
(27).
Water-suppressed and nonwater-sup-
pressed spectroscopic imaging data were acquired at
TR:
2000
ms,
TE
13
ms, and TM:
120
ms from
1-
or
2-cm
thick axial slices. Data acquisition and echo-planar gra-
dient encoding started
1
ms prior to the top of the stim-
ulated echo to minimize first order phase errors in the
spectra. The minimum acquisition time for a
32
X
32
spatial matrix was
64
s.
When necessary, averaging was
performed to improve the SNR.
Data
Processing
Data processing was performed off-line using the SA/GE
software
(GE
Medical Systems, Milwaukee, WI) on a
SUN
Sparc I1 workstation. Data representing “even” and “odd”
echoes were rearranged as shown in Fig.
2
to yield sep-
arate data sets that had conventional data format and that
could be processed separately. Spectral filtering con-
sisted
of
3
or
4
Hz exponential line-broadening. Fermi
filtering in the spatial domains (radius:
16
points, width:
4
points) reduced Gibb’s ringing. Residual water signals
were removed by low frequency filtering in the time
domain: A
131
point binomial filter was applied to the
spatially localized time domain data, and the result was
subtracted from the original time-domain data. Absorp-
tion mode spectra from the two data sets were added after
automatic and manual zero order phase correction to
maintain signal-to-noise. Magnitude spectroscopic im-
ages were created by spectral integration over a spectral
width of
12
Hz.
RESULTS
In
vitro
studies performed on a daily quality assurance
phantom
(GE
Medical Systems, Milwaukee, WI) that con-
tained water and glycerol confirmed that the spatial lo-
calization obtained with EPSI in
64
s
was similar to that
of
conventional HSI acquired in
32
min (Fig.
4).
A slight
loss in spatial resolution due to continuous data sam-
pling during the gradient ramps
(10-20%
depending on
the gradient ramp time) was only apparent before spatial
filtering was performed. Spectral resolution and spectral
artifacts with either method were similar (Fig.
5),
sug-
gesting only minor eddy current effects due to the oscil-
lating read-out gradient. The SNR of EPSI obtained in
1
min
in
vitro
was approximately
5-6
times less than that
in conventional
HSI
acquired in
32
min, a result consis-
tent with the shorter acquisition time. We investigated
possible localization errors caused by interactions with
local susceptibility related gradients by changing shim
gradients. We detected no significant spectroscopic im-
age distortions for the range of shim currents typically
used
in
vivo,
despite significant line broadening in local-
ized spectra. The amplitude of the dephasing gradient
lobe could be changed by up to
10%
without noticeable
changes in localization. Gradient tuning and stability
were verified by prelocalizing a
3
x
3
x
1
cc volume in
a homogeneous phantom containing distilled water. The
volume was shimed to a line width of
2
Hz.
In
the
k,,-t
plane, the symmetry of the magnitude and the phase of a
selected echo train demonstrate excellent gradient per-
formance (Fig.
6).
In
vivo
measurements were acquired in
1-16
min and
with voxel sizes ranging from
0.4-2
cc (Figs.
7
and
8).
a
b
C
d
FIG.
4.
Comparison between conventional
HSI
and
EPSl
on a daily
quality assurance phantom obtained
with
a quadrature head coil.
(a) Gradient echo localizer
(TE
5
ms,
TR:
100
ms,
(Y:
45”,
FOV:
24
cm).
(b)
Conventional spectroscopic image obtained
with
phase
encoding
in
32
min.
This
image was obtained
by
integration over
the entire spectral range. (c,
d)
Spectroscopic images obtained
with
EPSl
in
1
min.
(c) Integrated image computed from even echo
data.
(d)
Integrated image computed from odd echo data
(Tf:
13
ms,
TR:
1
s,
TM:
120
ms,
Matrix:
32
x
32,
Number of averages:
2).
High Speed Echo Planar Spectroscopic Imaging
37
IPP~I
a
FIG.
5.
Selected individual spectra from the data
sets
in
Fig.
4.
(a) Conventional
HSI.
The two resonances represent water (on resonance)
and glycerol (off resonance). The EPSl spectrum
(b)
was obtained
by
adding absorption mode spectra from the even- and odd-echo data
sets
in
Fig.
4.
The
total spectral range of
7.64
pprn
is
shown and demonstrates the absence of aliasing artifacts.
There
is
little
evidence
for additional eddy current effects due to the
EPSl
read-out gradients
(see
distortions to the right of the water line).
FIG.
6.
Single raw data trace from
an EPSl data
set
after echo
edit-
ing. The two echo trains corre-
spond to the even-odd echo data
format shown
in
Fig.
2.
(a) Echo
magnitudes.
(b)
Echo phases
(prelocalized volume:
3
x
3
x
1
cc, spatial resolution:
7.5
mrn).
a
b
The spectral pattern was similar to that found in our
previous experiments with short echo time single voxel
spectroscopy and phase encoded spectroscopic imaging
(lo).
We detected singlet resonances from choline, crea-
tine, and NAA, as well as multiplet resonances from
compounds such as inositol, glutamine, and glutamate
(Fig.
8).
The SNR in spectra obtained at different depths
was measured in selected data sets covering a wide range
of acquisition times and voxel sizes. The normalized
SNR
at a given distance from the coil per square root of
unit acquisition time and per unit volume (NSNR(x)) can
be used to compare the
SNR
obtained with different
parameter settings. For example, at a depth of
2-3
cm
from the surface of the head, immediately adjacent to the
region of outer volume suppression, the NSNR of the
NAA resonance was between
0.5
S-'.~C~-~
and
0.75
S-'.~C~-~,
a result consistent with our previous experi-
ence using surface coil single voxel spectroscopy. At a
depth of approximately
6
cm from the surface of the
head, the SNR of NAA decreased by approximately
50%,
which
is
consistent with the expected sensitivity profile
of the surface coil. The useful depth range measured from
the edge of the lipid suppressed region was typically
between
3
and
4
cm.
Lipid suppression pulses were tuned individually by
observing the residual lipid signals from individual outer
volume suppression slices (see Methods). The slice se-
lection gradients for outer volume suppression were
maximized within the constraints of the required slice
width and the maximum
RF
power, to reduce chemical
shift artifacts. Tuning parameters obtained in initial mea-
surements were used for later measurements without
further modifications. Residual lipid signals in the re-
sulting EPSI data sets were less than
or
equal to the level
of the metabolite resonances, except in some cases where
errors in the positioning of the outer volume suppression
were made.
DISCUSSION
EPSI provides an
SNR
per unit time and unit volume that
is comparable with that obtained with conventional tech-
niques at the same bandwidth per data point. Local gra-
dients interfere with the echo formation in k-space and
lead to aliasing artifacts which degrade the SNR unless
eventodd echo editing is used. Proper phasing of each
edited and processed data set is required to maintain the
SNR in absorption mode spectra after adding the two
38
a
Posse et
al.
MRI
4
s
(256x128)
b
NAA
64
s
(32x32)
. . .
. . . . .
.
FIG.
7.
EPSl
data set obtained on a
normal volunteer in occipital brain
in 1 min with a voxel size
of
1.1 cc
(7.5
x
7.5
x
20
mm3).
(a) Gradient
recalled echo localizer scan. (b)
NAA
map. (c) Zoomed spectral ar-
ray from a subset of spectra in the
proximtty of the surface coil dis-
playing a spectral range from 1.5
to
3.8
ppm. (d) Individual spectrum
displaying the entire spectral
range.
d
i
lr~,~l,~,~~,,,~
III,III,,,I/I
,,,I
,,,,
v
8
6
4
2
data sets. Because automatic phasing may fail due to
residual water signals, we are currently exploring the
possibility of using a nonwater-suppressed data set as a
spectral reference for phasing. The extra acquisition time
for such a reference data set is short due to the speed
of
this encoding
scheme.
The spatial resolution and spectral bandwidth are lim-
ited by gradient rise times and eddy current performance.