Unedited in vivo detection and quantification
of g-aminobutyric acid in the occipital cortex
using short-TE MRS at 3 T
Jamie Near
a,b,c
*, Jesper Andersson
c
, Eduard Maron
d
, Ralf Mekle
e,f
,
Rolf Gruetter
f,g
, Philip Cowen
b
and Peter Jezzard
c
Short-TE MRS has been proposed recently as a method for the in vivo detection and quantification of g-aminobutyric
acid (GABA) in the human brain at 3 T. In this study, we investigated the accuracy and reproducibility of short-TE
MRS measurements of GABA at 3 T using both simulations and experiments. LCModel analysis was performed on
a large number of simulated spectra with known metabolite input concentrations. Simulated spectra were generated
using a range of spectral linewidths and signal-to-noise ratios to investigate the effect of varying experimental
conditions, and analyses were performed using two different baseline models to investigate the effect of an
inaccurate baseline model on GABA quantification. The results of these analyses indicated that, under experimental
conditions corresponding to those typically observed in the occipital cortex, GABA concentration estimates are
reproducible (mean reproducibility error, <20%), even when an incorrect baseline model is used. However, simula-
tions indicate that the accuracy of GABA concentration estimates depends strongly on the experimental conditions
(linewidth and signal-to-noise ratio). In addition to simulations, in vivo GABA measurements were performed using
both spectral editing and short-TE MRS in the occipital cortex of 14 healthy volunteers. Short-TE MRS measurements
of GABA exhibited a significant positive correlation with edited GABA measurements (R = 0.58, p < 0.05), suggesting
that short-TE measurements of GABA correspond well with measurements made using spectral editing techniques.
Finally, within-session reproducibility was assessed in the same 14 subjects using four consecutive short-TE GABA
measurements in the occipital cortex. Across all subjects, the average coefficient of variation of these four GABA
measurements was 8.7 4.9%. This study demonstrates that, under some experimental conditions, short-TE MRS
can be employed for the reproducible detection of GABA at 3 T, but that the technique should be used with caution,
as the results are dependent on the experimental conditions. Copyright © 2013 John Wiley & Sons, Ltd.
Keywords: g-aminobutyric acid (GABA); spin-echo full-intensity acquired localised (SPECIAL); MRS; short-TE; spectral editing
INTRODUCTION
g-Aminobutyric acid (GABA) is the primary inhibitory neurotrans-
mitter in the mammalian brain and plays an important role in the
regulation of neuronal activity (1). Altered tissue GABA levels
have been observed in various pathologies, including both
epilepsy (2) and major depression (3,4), and individual variations
in tissue GABA levels of healthy individuals have been shown to
correlate with both functional MRI activity (5,6) and behaviour
(7,8). In vivo MRS detection of GABA is challenging because of
its relatively low concentration and the large overlapping
resonances from other metabolites and macromolecules (MMs).
As a result, GABA detection is most commonly performed using
spectral editing techniques, which enable the selective observation
of GABA by separation of the C4-GABA multiplet from the back-
ground of overlapping resonances (9–12). Of the available spectral
editing methods, probably the most commonly used is the
Mescher–Garwood point-resolved spectroscopy (MEGA-PRESS)
* Correspondence to: J. Near, FMRIB Centre, Nuffield Department of Clinical
Neurosciences, John Radcliffe Hospital, Headington, Oxford OX3 9DU, UK.
E-mail: jnear@fmrib.ox.ac.uk
a J. Near
Douglas Mental Health University Institute and Department of Psychiatry, Mc-
Gill University, Montreal, QC, Canada
b J. Near, P. Cowen
Department of Psychiatry, University of Oxford, Oxford, UK
c J. Near, J. Andersson, P. Jezzard
FMRIB Centre, Nuffield Department of Clinical Neurosciences, University of
Oxford, Oxford, UK
d E. Maron
Faculty of Medicine, Imperial College, London, UK
e R. Mekle
Laboratory for Functional and Metabolic Imaging (LIFMET), Ecole
Polytechnique Federale de Lausanne, Lausanne, Switzerland
f R. Mekle, R. Gruetter
Department of Radiology, University of Lausanne, Lausanne, Switzerland
g R. Gruetter
Centre d’Imagerie Biomedicale, Ecole Polytechnique Federale de Lausanne,
Lausanne, Switzerland
Abbreviations used: CRLB, Cramer–Rao lower bound; GABA, g-aminobutyric
acid; LW, linewidth; MEGA-PRESS, Mescher–Garwood point-resolved spectros-
copy; MM, macromolecule; NAA, N-acetylaspartate; SNR, signal-to-noise ratio;
SPECIAL, spin-echo full-intensity acquired localised; VAPOR, variable power
radiofrequency pulses with optimised relaxation delays.
Research article
Received: 20 January 2012, Revised: 15 March 2013, Accepted: 18 March 2013, Published online in Wiley Online Library: 22 May 2013
(wileyonlinelibrary.com) DOI: 10.1002/nbm.2960
NMR Biomed. 2013; 26: 1353–1362 Copyright © 2013 John Wiley & Sons, Ltd.
1353
technique (10), which combines MEGA editing (10) with PRESS
localisation (13) to achieve reliable quantitative measurements of
GABA concentrations within a localised region of tissue. Studies
have shown that the MEGA-PRESS technique provides reproduc-
ible GABA measurements, with intra-subject reproducibility values
in the range 7–12% (14,15). However, this acquisition method also
has a number of associated drawbacks. First, when optimised for
the observation of GABA, the MEGA-PRESS technique does not
enable the optimal detection of many other metabolites simulta-
neously. Thus, it provides a very limited amount of useful
metabolic information. Second, because the MEGA editing scheme
makes use of narrow-band frequency-selective pulses, it is very sen-
sitive to B
0
field drift, and small drifts of only a few hertz can affect
significantly the quantification accuracy (12,16). Third, the tech-
nique is inefficient; the observed GABA signal in the MEGA-PRESS
experiment typically consists of less than 40% of the available
signal from just one of GABA’s three methylene groups (16). The
remainder is lost in the editing process.
Short-TE MRS provides a possible alternative approach for the
detection of GABA, and may provide several advantages over the
more standard spectral editing techniques. Specifically, short-TE
MRS enables the detection of a large number of metabolites
(including GABA) simultaneously, thus increasing the available
amount of metabolic information. Furthermore, in contrast with
spectral editing techniques, short-TE MRS is relatively insensitive
to B
0
field drift and is highly efficient as it minimises signal decay
from T
2
relaxation and scalar coupling phase evolution. The de-
tection and quantification of GABA using short-TE MRS in combi-
nation with LCModel analysis has been demonstrated previously,
but has been mainly restricted to rodent studies and/or very
high field strengths (≥7 T) (17–20). Recently, however, Mekle
et al. (21) demonstrated short-TE (6 ms) MRS detection of GABA
at 3 T in the occipital cortices of six human subjects, with an av-
erage Cramer–Rao lower bound (CRLB) uncertainty of 8%. This
suggests that reliable GABA detection is feasible at clinical field
strengths, without the need for spectral editing. Following this
promising initial study, further investigation is required to con-
firm the initial results and to validate the use of short-TE MRS
for the detection of GABA at 3 T.
Therefore, the purpose of this study was to assess the accuracy
and reproducibility of GABA detection using the short-TE
approach under normal experimental conditions, and to investi-
gate how a change in the experimental conditions, such as
linewidth (LW) and signal-to-noise ratio (SNR), influence the ac-
curacy and reproducibility of this technique. These investigations
were performed using a methodology similar to that described
by Hancu (22,23), namely a 3-T, short-TE spectrum was simulated
by combining metabolite basis spectra in approximate physio-
logical concentrations followed by the addition of noise. The
simulated spectrum was then analysed using LCModel, and the
resulting metabolite concentration estimates were compared
with the known concentrations in the simulated input spectrum.
The above procedure was repeated many times to enable the es-
timation of the overall accuracy and reproducibility of the GABA
measurements. In addition to the simulated GABA measure-
ments described above, two in vivo experiments were performed
to further evaluate the reproducibility of short-TE MRS detection
of GABA. First, in vivo GABA measurements were performed
using both spectral editing and short-TE MRS in the occipital
cortex of 14 healthy volunteers, and the GABA concentration
estimates from short-TE MRS were compared with the gold
standard edited GABA measurement. Finally, within-session
reproducibility was assessed in the same 14 subjects using four
consecutive short-TE GABA measurements in the occipital
cortex.
METHODS
Simulated spectra
A complete set of metabolite basis spectra consisting of 22 indi-
vidual metabolites (Table 1) was simulated using an in-house
MATLAB-based implementation (MathWorks, Natick, MA, USA)
of the density matrix formalism (12). All metabolite chemical
shifts and coupling constants were taken from Govindaraju
et al. (24), with the exception of GABA, which was defined using
the modified spin system parameters provided by Kaiser et al.
(25). Metabolites were simulated under the influence of an ideal
spin-echo sequence to approximate the short-TE spin-echo full-
intensity acquired localised (SPECIAL) technique (21) (2048
points; spectral width, 2000 Hz; TE = 8.5 ms) with a field strength
of 3 T. Lipid and MM signals were simulated using the same
Gaussian basis functions that are simulated by default within
the LCModel software (Table 2). A residual water basis spectrum,
modelled as a two-proton singlet at 4.7 ppm, was also simulated.
All basis spectra were line broadened according to the desired
LW of the final spectrum, and an additional line-broadening
factor was applied to lipid and MM basis spectra, as specified in
Table 2 and in the LCModel user manual. Following line broaden-
ing, a simulated spectrum was generated by combining all basis
spectra in approximate in vivo concentrations. The average metab-
olite input concentrations were estimated from the literature (24)
and are specified in Table 1, whereas the average lipid and MM in-
put concentration values (specified in Table 2) were estimated
from 12 LCModel outputs of previously obtained in vivo datasets
Table 1. List of simulated metabolites and average
concentrations
Metabolite Average concentration (m
M)
Alanine 0.5
Ascorbate 0.8
Aspartate 1.5
Creatine 5.25
Phosphocreatine 4.75
g-Aminobutyric acid 1.2
Glutamine 3.2
Glutamate 9.5
Glutathione 1.5
Myo-inositol 6
Lactate 0.4
N-Acetylaspartate 12
Scyllo-inositol 0.45
Taurine 1.5
Glucose 1.0
N-Acetylaspartylglutamate 1.5
Phosphoethanolamine 1.3
Glycerophosphocholine 1.0
Phosphocholine 0.6
Glycine 0.7
Serine 0.4
b-Hydroxybutyrate 0.1
J. NEAR ET AL.
wileyonlinelibrary.com/journal/nbm Copyright © 2013 John Wiley & Sons, Ltd. NMR Biomed. 2013; 26: 1353–1362
1354
acquired in the occipital cortex of normal subjects using the rele-
vant pulse sequence and timing parameters. As with the metabo-
lites, the concentrations of each of the six MMs and lipid signals
were allowed to vary independently of each other. Normally dis-
tributed random noise was then added to the simulated spectrum
to achieve the desired SNR, which was defined as the maximum
metabolite peak height divided by the standard deviation of the
added noise. For a given SNR and LW, the above procedure was
repeated 500 times with different noise seeds. In each of the
500 repetitions , to account for n ormal subject- to-s ubject varia-
tion, the input concentration of each basis spectrum was
allowed to vary randomly w ith a standard deviation of 15% of
its mean value. The only exception to this was the residual water
signal, which was assigned an average value of 0 m
M and a stan-
dard deviation of 20 m
M.
Up to this point in the simulation process, 500 simulated
spectra will have been generated in which the experimental con-
ditions (LW and SNR) are identical, and the relative metabolite
concentrations will vary as they might in a normal human popu-
lation. Then, the same procedure was repeated for different sets
of experimental conditions. In total, every combination of 11 dif-
ferent LW values (ranging from 2 to 12 Hz in integer steps) and
18 different SNR values (ranging from 50 to 900, in steps of 50)
was tested, for a total of 198 experimental conditions and
99 000 simulated spectra. Each spectrum was then processed
twice in LCModel; once without a baseline fitting component
(achieved by setting ‘NOBASE = T’) and once using the default
LCModel baseline setting, in which LCModel attempts to find
the smoothest possible baseline that is still consistent with the
data. For analysis of the simulated data in this study, the
baseline-free analysis is the more appropriate option, as all of
the peaks in the simulated spectra (with the exception of the tail
of the residual water peak) should be accounted for by the basis
set. However, a second analysis was performed using the default
baseline setting to examine the effect of an incorrect baseline
model on the accuracy and reproducibility of GABA concentra-
tion estimates. Both LCModel analyses were performed using
the same basis set as was employed to generate the simulated
datasets and the LCModel estimates of the metabolite concen-
trations, and CRLB uncertainties were recorded for each analysis.
Critically, the metabolite input concentrations are precisely
known for each simulated spectrum; therefore, it is possible to
assess the accuracy and reproducibility of the metabolite con-
centration estimates obtained from LCModel.
Assessment of measurement bias and reproducibility
For each set of experimental conditions, the measurement bias
and reproducibility were assessed. Measurement bias was
assessed by calculating the mean estimation error (%E
E
), which
was defined as the average percentage difference between the
estimated and actual GABA concentration values:
%E
E
¼
1
N
X
N
i
C
estimated
i
C
actual
i
C
actual
i
100 ¼
1
N
X
N
i
E
i
C
actual
i
100 [1]
where C
estimated
is the concentration estimate from LCModel,
C
actual
is the known input concentration, updated to correctly
account for subject-to-subject variation in the data, E is the
difference between the estimated and actual metabolite concen-
trations and N is the total number of simulated spectra per ex-
perimental condition (500). Reproducibility was assessed by
calculating the reproducibility error (%E
R
). This was performed
by first plotting the measured GABA concentrations versus the
actual GABA input concentrations and performing a linear
least-squares fi t to the data. %E
R
was then given by the standard
deviation of the values of the fitted residuals, divided by the av-
erage of the estimated concentration values:
%E
R
¼
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1
N
X
N
i
C
estimated
i
C
fit
i
ðÞ
2
s
1
N
X
N
i
C
estimated
i
100 ¼
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1
N
X
N
i
e
2
i
s
1
N
X
N
i
C
estimated
i
100
[2]
where C
fit
is the concentration calculated from the linear data fit
described above and e
i
are the fitted residuals. By this definition,
%E
R
is expressed as a percentage and provides an approximate
measure of the coefficient of variation. It should be noted that,
because the reproducibility error %E
R
is expressed relative to
the average measured value C
estimated
, its value is dependent
on the measurement bias. Therefore, it is informative to define
a second measure of the reproducibility error (%E
R2
), which is in-
dependent of measurement bias. This is achieved by, instead, ex-
pressing the reproducibility error relative to the average actual
concentration value C
actual
. Although this is no longer a true
measurement of the coefficient of variation, it provides a means
of assessing the reproducibility independently of any potential
measurement bias.
Table 2. Details of simulated lipid (Lip) and macromolecule (MM) basis spectra
Lipid/macromolecule Frequency (ppm) Line-broadening factor (Hz) Amplitude (# protons) Concentration (m
M)
MM09 0.91 21.0 3 9.0
Lip20 2.04 24.6 1.33 1.0
2.25 18.5 0.67
1.95 24.6 0.87
MM20 2.08 22.2 1.33 16.0
2.25 24.6 0.33
1.95 18.5 0.33
3.00 24.6 0.40
MM12 1.21 24.6 2.0 3.75
MM14 1.43 24.6 2.0 8.0
MM17 1.67 21.0 2.0 4.0
UNEDITED g-AMINOBUTYRIC ACID MRS AT 3 T
NMR Biomed. 2013; 26: 1353–1362 Copyright © 2013 John Wiley & Sons, Ltd. wileyonlinelibrary.com/journal/nbm
1355
In vivo experiments
All volunteers (n = 14; age, 23.3 5.4 years; eight women, six
men) provided informed, written consent and were scanned on
a 3-T Siemens TIM Trio scanner (Erlangen, Germany) with a body
coil transmitter and a 32-channel receive head array. Both short-
TE and edited GABA spectra were acquired in the same scan ses-
sion from the same localised region measuring 2.5 2.5 2.5
cm
3
in the occipital cortex. Shimming was performed using
the vendor-provided automated shim tool. Short-TE MR spectra
were acquired using the SPECIAL sequence (4096 points; spectral
width, 4000 Hz; TR/TE = 3000/8.5 ms; 192 averages) and edited
GABA spectra were acquired using the MEGA-SPECIAL sequence
(2048 points; spectral width, 2400 Hz; TR/TE = 3000/68 ms; 192
averages). The MEGA-SPECIAL sequence was implemented with
20-ms editing pulses and an MM-unsuppressed editing
scheme (12). To limit the amount of frequency drift in any single
scan, the MEGA-SPECIAL acquisition was broken into three
blocks of 64 averages, and a system frequency adjustment was
performed prior to the start of each acquisition block. For
both short-TE and edited acquisitions, outer volume suppres-
sion was applied prior to each scan to saturate spins on all
six sides of the region of interest, and VAPOR (variable power
radiofrequency pulses with optimised relaxation delays) water
suppressionwasused(19).Finally,eightaveragesofwater-
unsuppressed data were also acquired with the same outer
volume suppression scheme as above.
Post-processing and analysis
The same post-processing chain was applied to both the edited
and short-TE datasets. First, 32-channel data were recombined in
a weighted fashion, with coil weights and phases determined
using the magnitude and phase, respectively, of the first time
domain point of the water-unsuppressed data. Following coil re-
combination, the subspectra resulting from SPECIAL pre-
inversion on/off scans were subtracted from each other,
resulting in properly localised scans. Following subtraction of
the subspectra, a strict procedure to remove motion-corrupted
scans was employed. To identify motion-corrupted scans, a met-
ric was developed to measure the ‘unlikeness’ of each scan to
the average of all scans. Specifically, an ‘unlikeness metric’ was
calculated for each individual scan by subtracting the scan from
the average of all scans (to obtain a difference spectrum) and
then taking the root-mean-square of all of the spectral points
in the difference spectrum. Scans whose ‘unlikeness metrics’ fell
more than 2.6 standard deviations above the average were
deemed to have been corrupted by motion, and were removed.
The 2.6 standard deviation threshold was determined from expe-
rience to be fairly successful at removing outlier scans without
removing uncorrupted averages. Following the removal of
motion-corrupted scans, but prior to signal averaging, a fre-
quency and phase drift correction was performed. This was
achieved by least-squares fitting of each scan to the first scan
in the series, using frequency and phase as adjustment parame-
ters. To reduce computational load, this procedure was
performed in the time domain, using only the first 40 ms of data.
For the MEGA-SPECIAL edited data, both the removal of motion-
corrupted scans and the frequency and phase alignment were
performed separately for edit-on and edit-off spectra, with the
constraint that the number of edit-on and edit-off scans removed
must be equal. Following frequency and phase alignment of the
scans, signal averagin g was performed, resulting in a fully
processed short-TE spectrum, and fully processed edit-on and
edit-off MEGA-SPECIAL data. The edit-on and edit-off scans in
the MEGA-SPECIAL data w ere then manually frequency and
phase aligned to minimise the residual choline difference sig-
nal, and the edit-on and edit-off scans were then subtracted,
resulting in three fully processed difference-edited spectra
(one for each of the three acquisition blocks). Finally, the three
fully processed MEGA-SPECIAL data blocks were combined
using the same automated time domain frequency and phase
alignment algorithm as used for drift correction, resulting in a
single fully processed difference edited spectrum.
All experimentally acquired short-TE SPECIAL MRS data
were analysed in LCModel using the default baseline setting
and the same basis set as used for the analysis of the simu-
lated data as described above. Edited MEGA-SPECIAL MRS
data were analysed by peak fitting using jMRUI software
(26), according to the method described previously (12). Me-
tabolite concentration estimates were c orrected for T
2
relaxa-
tion during TE by assuming T
2
values of 88 and 116 ms for
GABA and creatine, respectively.
Within-session reproducibility was assessed by splitting the
previously acquired short-TE SPECIAL data into four equal and
consecutive blocks, each containing 48 averages. Each of these
four blocks was then pre-processed identically, as described
above, with the sole restriction that the same number of aver-
ages was discarded from each of the four blocks for removal of
motion corruption. Each of the four processed datasets was then
analysed using LCModel employing the default baseline setting,
and the GABA concentration estimates were recorded. The coef-
ficient of variation of the GABA estimates across the four scan
blocks was then calculated for each subject.
For the measurement of SNR in experimental data, the ‘signal’
was defined as the maximum intensity of the real part of the me-
tabolite signal between 0.2 and 4.2 ppm, which always
corresponded with the N-acetylaspartate (NAA) peak. The ‘noise’
was calculated by performing a second-order polynomial de-
trend of each of the spectral regions between [–2.5, –1.5], [–1.5,
–0.5], [10, 11] and [11, 12] ppm, and then taking the average of
the standard deviations of the real parts of the signals in these
regions.
RESULTS
The contour plots in Figure 1 illustrate the measurement bias and
reproducibility errors of GABA estimates as a function of LW
and SNR for the simulat ed data when the baseli ne component
is omitted from LCModel fitting. The measurement bias (%E
E
),
reproducibility error (%E
R
), average CRLB and absolute reproduc-
ibility error (%E
R2
) are shown in Figures 1a–d, respectively.
Figure 2 illustrates the same parameters as in Figure 1 when
using the default baseline setting in LCModel instead of the
baseline-free option. The measurement bias (%E
E
), reproducibility
error (%E
R
), average CRLB and absolute reproducibility error (%E
R2
)
are shown in Figures 2a–d, respectively.
Figure 3a shows a representative in vivo spectrum acquired
using the short-TE SPECIAL sequence, together with an
anatomical image showing the location o f the volume of
interest in the occipital cortex. Across the 14 subjects
scanned, the average LW was 5.5 0.8 Hz and the average
SNR was 665 93. The average CRLB for GABA, as measured
J. NEAR ET AL.
wileyonlinelibrary.com/journal/nbm Copyright © 2013 John Wiley & Sons, Ltd. NMR Biomed. 2013; 26: 1353–1362
1356
using LCModel , was 1 1.7 2.8%. Figure 3b shows an example
of an edited spectrum acquired using the MEGA-SPECIAL se-
quenceinthesamesubjectandvoxel.
In Figure 4, the GABA concentrations measured with short-TE
SPECIAL MRS are plotted against the edited GABA measure-
ments made using MEGA-SPECIAL. A significant positive linear
Reproducibility Error (%ER)
Linewidth (Hz)
SNR
Average CRLB
Linewidth (Hz)
SNR
Reproducibility Error (%ER2)
Linewidth (Hz)
SNR
a. b.
c. d.
−50
−40
−40
30
−30
−30
−20
−20
−20
−10
−10
−10
0
0
0
0
0
10
10
10
1
0
20
20
20
30
30
30
40
40
40
50
50
50
Mean Estimation Error (%EE)
Linewidth (Hz)
SNR
2 4 6 8 10
12
100
200
400
600
800
5
5
10
10
10
10
15
15
15
15
20
20
20
20
20
20
20
20
25
25
25
2
5
25
40
40
40
4
0
60
60
6
0
80
2 4 6 8 10 12
100
200
400
600
800
10
10
10
15
15
15
15
20
20
20
20
20
20
25
2
5
25
25
25
4
0
40
40
60
6
0
80
2 4 6 8 10 12
100
200
400
600
800
5
5
5
10
10
10
10
10
15
15
15
15
1
5
15
20
20
20
20
20
20
20
25
25
25
25
25
2 4 6 8 10 12
100
200
400
600
800
Figure 2. Contour plots showing mean estimation error (%E
E
) (a), reproducibility error (%E
R
) (b), average Cramer–Rao lower bound (CRLB) (c) and
absolute reproducibility error (%E
R2
) (d) of g-aminobutyric acid (GABA) for simulated data as a function of the linewidth (LW) and signal-to-noise ratio
(SNR). LCModel analysis was performed using the default baseline model.
a. b.
c. d.
50
50
50
5
0
50
50
50
5
0
100
100
100
100
150
1
50
150
150
200
25
0
Mean Estimation Error (%)
Linewidth (Hz)
SNR
2 4 6 8 10 12
100
200
400
600
800
4
4
6
6
6
6
8
8
8
10
10
10
12
12
12
14
14
14
1
4
16
16
16
18
Reproducibility Error (%)
Linewidth (Hz)
2 4 6 8 10 12
100
200
400
600
800
6
6
8
8
8
8
8
8
8
10
10
10
10
1
0
10
10
10
12
12
1
2
12
12
12
14
16
18
2
0
Average CRLB (%)
Linewidth (Hz)
SNR
2 4 6 8 10 12
100
200
400
600
800
5
5
10
10
10
10
1
0
15
15
15
20
20
20
20
25
25
25
25
25
30
30
30
30
35
35
35
4
0
45
50
Reproducibility Error (Absolute)
Linewidth (Hz)
SNR
2 4 6 8 10 12
100
200
400
600
800
20
SNR
Figure 1. Contour plots showing mean estimation error (%E
E
) (a), reproducibility error (%E
R
) (b), average Cramer–Rao lower bound (CRLB) (c) and
absolute reproducibility error (%E
R2
) (d) of g-aminobutyric acid (GABA) for simulated data as a function of linewidth (LW) and signal-to-noise ratio
(SNR). LCModel analysis was performed with no baseline component by setting ‘NOBASE = T’.
UNEDITED g-AMINOBUTYRIC ACID MRS AT 3 T
NMR Biomed. 2013; 26: 1353–1362 Copyright © 2013 John Wiley & Sons, Ltd. wileyonlinelibrary.com/journal/nbm
1357
