In vivo measurement of brain metabolites using two-dimensional double-quantum MR spectroscopy—exploration of GABA levels in a ketogenic diet



A localized proton 2D double-quantum (DQ) spin-echo spectroscopy technique was implemented on 1.5 T clinical MRI scanners for the detection of γ-aminobutyrate (GABA) in the brain. The 2D approach facilitates separation of peaks overlapping with GABA in 1D DQ-filtered (DQF) spectra. This technique was applied to four normal adult volunteers and four children with intractable epilepsy. The coefficient of variation of the level of GABA and overlapping macromolecules at F2 = 3.0 ppm and F1 = 4.8 ppm was 0.08 in normal subjects. Three patients received 2D MRS scans before and after initiation of the ketogenic diet (KD): one patient showed a trend of decreasing GABA throughout the study, and two patients showed low initial GABA levels that increased over time. In addition to major metabolites and GABA, low-level metabolites (valine, leucine, and glutathione) were also identified in the 2D spectra. Magn Reson Med 49:615–619, 2003. © 2003 Wiley-Liss, Inc.

Magnetic resonance spectroscopy (MRS) is the only tool capable of noninvasively measuring metabolite concentrations in the brain, and it has proven to be a valuable technique for evaluating neurologic diseases. Proton MRS is particularly informative concerning the nuclei available for signal detection (1). However, there are currently many limitations in clinical proton MRS examinations for both 1.5T and 3T MRI scanners. Signals from most metabolites are found in a narrow aliphatic spectral window between a chemical shift of about 1 and 4 parts per million (ppm), and numerous metabolites present in low concentrations are overshadowed by the resonance of a few metabolites present in high concentrations. Consequently, much potential information remains unrecovered.

Because most metabolites have coupled spins, two-dimensional (2D) correlation spectroscopy (2) can be used to provide a means of spreading the spectral peaks in a second frequency dimension, greatly increasing the information content of the data and facilitating peak quantification and assignment. The 2D approach is widely used for in vitro studies, and can be adapted for in vivo human studies (3). There are many technical benefits to be derived from the use of 2D MRS. These methods will allow us to detect metabolites that are difficult to detect even with editing techniques. For example, some metabolites may have peaks with small chemical shift separations, and, consequently, selectively exciting one spin group without disturbing the others for editing purposes is difficult. However, 2D MRS can be employed to detect these metabolites without using selective pulses in the frequency domain. Furthermore, the use of a 2D correlation spectroscopic technique will enhance our ability to interpret and assign unexpected peaks found in patient examinations that are difficult to assign in 1D spectra.

This report describes the application of 2D double-quantum (DQ) measurements of GABA levels in the brain of normal subjects and of patients undergoing ketogenic diet (KD) treatment for intractable epilepsy. There are an estimated 70,000–128,700 new cases of epilepsy per year in the United States (4). For most epilepsy patients (70–80%), seizures are well controlled by conventional or investigational antiepileptic drugs (AEDs). However, the remaining 20–30% have intractable epilepsy, and the chance that they will be seizure-free after further AED manipulation is disappointingly small (5–10%). The KD was developed by the Mayo Clinic in 1921 (5), and an efficacy of 50–75% has been reported in this subgroup of patients (6). The mechanism behind the KD is poorly understood. It mimics the fasting state by maintaining metabolic ketosis. Basic scientific research suggests that GABA, the major inhibitory neurotransmitter, is present in low levels in the brain of patients with intractable epilepsy. Changes in synaptic GABA concentrations have also been implicated as a possible mechanism of the KD (7). The objectives of the present study were twofold: 1) to explore the feasibility of using a 2D technique in routine clinical examinations, and 2) to determine whether the 2D DQ-MRS technique can be used to study the effect of KD on the brain GABA levels in epilepsy patients.


All studies were done on two identical Magnetom Vision 1.5T clinical MRI scanners (Siemens Medical Systems, Iselin, NJ) using an adult head coil. The pulse sequences were developed using the system pulse sequence generation software “PARGEN.” A DQ-filtered (DQF), single-voxel, point-resolved spectroscopy (PRESS) technique optimized for GABA detection (8) was modified to obtain a 2D DQ-MRS pulse sequence by data acquisition with a series of t1 values (Fig. 1). Outer-volume saturation pulses were applied to decrease the signal contamination by lipids from the scalp. Spatially-selective RF pulses were generated on a SUN Ultrasparc station using Interactive Data Language (IDL; Research Systems, Inc., Boulder, CO). The phase of the spatial RF pulses was calculated according to the voxel center position. The pulses were a sinc function with a length of 5.12 msec and a bandwidth corresponding to the localization B0 gradient (0.1 gauss/cm for the 90° pulse, and 0.067 gauss/cm for the 180° pulses) multiplied by the voxel dimension and by a cosine window function from –π/2 to π/2. The phase of pulse 6 was set such that the value at the center of this spatial RF pulse was π. The phase value at the center of pulse 5, ϕ1, was adjusted for the strongest GABA signal from a phantom in a 1D DQF measurement; this replaced the phase calibration procedure described in Ref. 8. The data were acquired sequentially using four pulse sequences that interleaved t1 to obtain 128 t1 values of 0.5–64 msec, because each pulse sequence can accommodate 32 t1 signals. For each echo, 1024 points were recorded in t2 with a dwell time of 1 msec. The resulting raw data files were transferred to a SUN workstation for analysis using in-house-written software in IDL. The 2D time domain data were zero-filled to 256 points in the t1 direction and Fourier-transformed to the frequency domain. The IDL program displays the contour plot of the modulus 2D spectra and, from the same data, a stack plot of 1D single-quantum absorption spectra for a range of DQ frequencies. The phase adjustment of the 2D spectra consisted of adjusting a constant phase, a linear phase in F1, and a linear phase in F2. In the 2D spectra, the peak position in F1 is the sum of the chemical shifts of two coupled spins, and F2 is the single-quantum dimension (2). For example, in lactate, which has CH3 protons at 1.31 ppm and CH protons at 4.13 ppm, the 2D DQ spectrum has a doublet centered at [F1, F2] = [5.44, 1.31] ppm and a quadruplet at [F1, F2] = [5.44, 4.13] ppm. This was verified in our solution phantom studies.

Figure 1.

2D spin-echo DQ pulse sequence. The pulse sequence can be divided into seven parts: i) CHESS water suppression. ii) Outer volume fat suppression. iii) Multiple-quantum coherence generation, where pulses 5 and 6 are also used for spatial localization in x and y. iv) Removal of single-quantum signals and refocusing of the DQ coherence. v) DQ coherence evolution time period t1. vi) Conversion of the DQ coherence to single-quantum signal and spatial localization in the z direction, combined with spin refocusing with twice the gradient pulse area of that in part iv. Pulses 9 and 10 form one composite coherent transfer pulse, which selectively excites spins at 1.9 ppm without exciting those at 3.0 ppm. Pulse 11 is a refocusing pulse that also achieves localization in the z direction. vii) Localized single-quantum evolution and detection. Parameters are: τ = 34 msec, ϵ = 5.5 msec, δ = 7.1 msec. Phases ϕ1 and ϕ2 are adjustable to take into account the fact that the GABA signal is not on resonance.

The technical objective of this study was to measure GABA C4 protons at 3.0 ppm that are coupled to GABA C3 protons at 1.9 ppm. However, there is also a macromolecule signal at 3.0 ppm that is coupled to protons at 1.7 ppm (9). While the GABA peak is centered at [F1, F2] = [4.9, 3.0] ppm, the macromolecule signal is present at [F1, F2] = [4.7, 3.0] ppm in the 2D DQ spectra. The two components are not resolved and they merge into a multiplet at [F1, F2] = [4.8, 3.0] ppm, the volume of which was quantified. The metabolite peak area under the absorption spectra is calculated for curves in the stack plot and summed to obtain a peak volume. The GABA signal was measured in arbitrary units and calibrated using the water signal of a phantom, taking into account coil loading (10):

equation image(1)

where S is the GABA level in arbitrary units, S0 is the peak volume of the [4.8, 3.0] ppm multiplet measured from the 2D absorption spectrum of the subject, Sw,phantom is the water signal amplitude from a phantom measured with a single quantum single-voxel spectroscopy pulse sequence, and Vref is the voltage at the output of the RF power amplifier needed for a 0.5-msec rectangular 180° pulse.

The study in humans was approved by the Institutional Review Board of the Children's Hospital of Philadelphia. Four normal volunteers (three males and one female, 19–20 years old) and four patients (two males and two females, 2–7 years old) undergoing KD treatment (11) were studied by use of the 2D MRS technique. The patients were sedated for the studies following a standard protocol for MRI examinations (age 2–5 years, pentobarbital 2 mg/kg and maximum 6 mg/kg; age 6–11 years, morphine 0.1 mg/kg followed by pentobarbital 2 mg/kg, maximum 6 mg/kg). For patients undergoing KD therapy, three serial studies were conducted: before the initiation of the KD, and at 2 weeks and 3 months into KD. The region of interest (ROI) was a 7 × 7 × 7 cm3 voxel enclosing both the parietal lobe and the central brain. Other parameters were TR = 1.6 sec, and number of acquisitions = 6. Data acquisition required 20 min to complete.


Tests on a GABA solution confirmed that the corresponding 1D DQF sequence (setting t1 = 0 msec in the 2D sequence) is capable of detecting GABA C4 protons with an efficiency of 36%, which is close to the theoretical prediction of 37.5% for 2τ = 1/2J (12). A 2D GABA solution study is presented in Fig. 2 as a stacked plot of absorption spectra with different F1. In the F1 dimension, the GABA signal is broadened by transverse relaxation, B0 field inhomogeneity, and a conversion of in-phase DQ to anti-phase DQ coherence caused by the J-coupling to passive spins (2) (C2 protons). For in vivo human studies, the water line width was typically 13 Hz after shimming. Figure 3a shows a 1D DQF spectrum from the same ROI used for the 2D study in a normal control subject. Figure 3b shows the contour plot of the 2D DQ modulus spectrum from this normal subject. The GABA signal, with a contribution from macromolecules (13), appears as twin peaks and is labeled in the figure. Figure 3c is a stacked plot for the GABA signal from the same data in Fig. 3b. Corrections for coil loading and scanner sensitivity were made using Eq. [1], and the resulting GABA level, including the macromolecule contribution, is listed in Table 1 in arbitrary units. The coefficient of variation of the combined GABA and macromolecule signal intensity was 8% for the four control subjects, which is comparable to the 0.1/1.6 reported by Rothman et al. (13). Patient 1 showed a trend of decreased GABA levels 3 months into the therapy. Patients 2–4 had low GABA levels before they started the KD treatment. Patients 2 and 3 showed an increase of GABA levels over time under the KD. Patient 4 was examined once at the beginning of the study.

Figure 2.

Spectrum of GABA solution with a concentration of 100 mM: a stacked plot of absorption spectra with different F1. The number on each curve is the DQ shift F1 in ppm. A low level of lactate was also added to the phantom as a frequency reference.

Figure 3.

Proton MR spectrum from the brain of a control subject. a: 1D DQF spectrum from a 7 × 7 × 7 cm3 voxel enclosing the parietal lobe and the central brain. b: Contour plot of modulus 2D DQ spectrum for the same ROI as in part a. The labels are defined as: Glx = glutamine and glutamate; Val = valine; Leu + MM = leucine and macromolecule; NAA = N-acetyl-aspartate; mI = myoinositol; GSH = glutathione; Lac = lactate; GABA + MM = γ-aminobutyrate and macromolecule. c: Stacked plot of absorption spectra from the same data in b. The number on each curve is the DQ shift F1 in ppm. Arrows 1 and 2 indicate the expected center positions for GABA and the macromolecule, respectively.

Table 1. GABA Level in Controls and Patients Undergoing KD (Arbitrary Unit)
Subject1st study (0 week)2nd study (2 weeks)3rd study (3 months)
Control 11.58 × 104  
Control 21.33 × 104  
Control 31.44 × 104  
Control 41.57 × 104  
Patient 11.64 × 1041.41 × 104
Patient 27.1 × 1038.7 × 1031.36 × 104
Patient 31.22 × 1041.32 × 1041.64 × 104
Patient 45.2 × 103  

Although the pulse sequence is optimized for GABA measurement, a variety of low-level metabolites can be detected, along with strong signals from NAA, glutamine, glutamate, and myoinositol (Fig. 3b). Leucine and valine, whose branched chain methyl signals overlap at 0.9 ppm, can be separated in the 2D spectrum. The glutathione (GSH) signal is present at [F1, F2] = [7.5, 2.9] ppm. The lactate methyl signal is also present at an appreciable level. Not all signals have been assigned to date, and further investigation is needed.


According to Rothman et al. (13), the 3.0 ppm signal detected with the subtraction editing technique contains an approximately 40% contribution from macromolecules in the normal brain. The macromolecules are also expected to contribute to the 3.0 ppm multiplet in the 1D DQF spectrum, although the exact amount is not known. McLean et al. (14) estimated that the macromolecule contribution to the 2.9 ppm component is about 10%, and that the contribution to the 3.1 ppm component is substantially higher. It is recognized that other metabolites or peptides also contribute to the 3.0 ppm multiplet (14). The 2D technique allows separation of the various peaks according to the DQ frequency. From Fig. 3b, it is apparent that signal components with a DQ frequency of approximately 7–8 ppm contribute to the 3.0 ppm twin-peak in the 1D DQF spectrum, including GSH at [F1, F2] = [7.5, 2.9] ppm. GSH is present at a significant level of 0.8–3.1 mM in a normal human brain (15). For Fig. 3b, the peak height ratio of GSH to GABA is 1.4, and the peak height ratio for the [F1, F2] = [7.2, 3.1] ppm peak to GABA is 2.5. These strong peaks interfere with GABA in the 1D spectrum acquired with DQF. It was initially hoped that the 2D DQ technique would allow us to separate GABA from the macromolecules. However, according to our data obtained at 1.5 T, the macromolecules and GABA were still not resolved from each other and formed one multiplet at [F1, F2] = [4.8, 3.0] ppm. Higher B0 field strengths may be needed to resolve the two components. In our studies, we used a large voxel size to overcome the difficulty of a low signal-to-noise ratio, which increased the degree of difficulty in shimming. Because the effects of the KD on GABA are expected to be diffuse and affect the whole brain, a large sampling volume will not prevent us from obtaining valuable information from the MRS study. Currently, more patients receiving KD treatment are being studied with this technique.

Studying GABA levels in patients undergoing KD treatment may shed light on the mechanism behind this important therapy. Moreover, the use of the 2D DQ approach will allow measurements of other metabolites that may help improve our understanding of the development of epilepsy and seizures. Furthermore, it is anticipated that the 2D DQ technique will provide improved separation of glutamine and glutamate compared with 1D techniques. In this exploratory study, the normal subjects and the patients were of different ages. The data from the normal adults mainly serve to demonstrate the variability of the measurements. The normal GABA level for the age range of our patients is probably different from that found in adults, due to developmental changes and differences in gray and white matter partial volumes. This is an important issue to be addressed in future investigations. The effects of sedation on the GABA level in patients are not known. Because each patient was under sedation in the current MRS studies, we anticipate similar effects in each follow-up study, and therefore MRS can be used to follow changes in GABA levels resulting from KD therapy.

Many reports concerning the implementation of 2D MRS in clinical scanners or similar magnets for animal studies have been published over the past 15 years (16–31). Early localized multiple-quantum techniques were not able to detect signals with optimal sensitivity. Recently, however, multiple-quantum spectroscopy was combined with PRESS localization without loss of detection sensitivity (32). This significant technical advance led to practical applications in human studies. There are many different ways to perform 2D MRS measurements, including J-resolved spectroscopy, correlated spectroscopy (COSY), spin echo correlation spectroscopy (SECSY), 2D zero-quantum spectroscopy, and 2D DQ spectroscopy. The use of 2D zero-quantum spectroscopy is appealing because there is no line-broadening in the zero-quantum dimension due to B0 field inhomogeneity. However, even a little system instability and patient motion can cause the uncoupled peaks to spread into the zero-quantum dimension and interfere with the detection of spins with weaker signals. This type of difficulty is also shared by conventional COSY-type experiments, in which uncoupled diagonal peaks have a strong presence. On the other hand, in 2D DQ-MRS uncoupled peaks are removed, which enables metabolites to be measured without interference from strong signals of uncoupled spins. Therefore, the 2D DQ-MRS is a more robust technique for patient studies. However, the removal of uncoupled spins from the spectra has the drawback that the 3.0 ppm creatine peak, a convenient internal reference, is no longer available from the same scan.


We are grateful to Dr. John A. Detre for his support and valuable discussions. We thank Brenden Tavelli, Qi Cao, and Maxwells Scherzer for assistance. We acknowledge Rolf Schulte for discussions on GSH detection, and Dr. Andreas Trabesinger for discussions on technical issues regarding DQF spectroscopy.