|Citation: Kassem, M. N., & Bartha, R. (2003). Quantitative proton short‐echo‐time LASER spectroscopy of normal human white matter and hippocampus at 4 Tesla incorporating macromolecule subtraction. Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine, 49(5), 918-927.|
Accurate quantification of in vivo short-echo-time (TE) 1H spectra must account for contributions from both mobile metabolites and less mobile macromolecules, which can fluctuate in disease. The purpose of this study was to develop an approach for the acquisition and processing of macromolecule information to optimize metabolite quantification accuracy and precision. Human parietal white matter (8-cm3 voxel) and posterior hippocampus (1.7-cm3 voxel) metabolite levels were quantified, following manomolecule subtraction, from short-echo-time spectra (TE = 46 ms) acquired at 4.0 Tesla with localization by adiabatic selective refocusing (LASER). Nineteen metabolites were fit using a time domain Levenberg-Marquardt minimization that incorporated prior knowledge of metabolite lineshapes. The macromolecule contribution to the spectrum was reduced by 87% (P < 0.05) when the acquisition of single averages of the full spectrum and macromolecule spectrum were interleaved to reduce subtraction errors due to motion. Subtracting the Hankel Lanczos singular value decomposition (HLSVD) fit of the macromolecule spectrum, which contained no random noise, did not alter quantified metabolite levels but did not increase metabolite quantification precision. Several metabolites had higher concentrations in the posterior hippocampus compared to parietal white matter, which emphasizes the need to carefully control for partial volume contamination in hippocampal spectroscopy studies. Magn Reson Med 49:918 –927, 2003. © 2003 Wiley-Liss, Inc.
Key words: white matter; hippocampus; quantification; macro- molecule; 1H spectroscopy
Proton (1H) magnetic resonance spectroscopy (MRS) has been used in clinical studies to detect in vivo metabolic changes in many pathologic conditions (1). The majority of these studies have used long echo times (TE > 100 ms) to detect changes in N-acetylaspartate (NAA), creatine (Cr), and choline (Cho), which are sensitive but nonspecific indicators of disease. The popularity of this approach is mainly due to the simplicity of the long-TE spectrum, which contains easily resolvable peaks for the metabolites listed above and does not contain overlapping signals from less mobile macromolecules (2). In contrast, short-TE (<50 ms) acquisitions, which minimize J-coupling modulations, can be used to detect additional metabolites, including glutamate (Glu), glutamine (Gln), and myo-inositol (Myo). However, broad macromolecule signals also con- tribute significant area to the short-TE spectrum (2).
Behar et al. (2) characterized the underlying macromolecule spectrum in short-TE (17 ms) localized 1H spectra of normal human brain. Resonances of broad macromolecules were tentatively assigned to methyl and methylene resonances of amino acids in proteins (3). Although prominent macromolecular and lipid resonances appear in the 0.5–2.0 ppm region of the spectrum, these broad signals also extend beneath the 2– 4 ppm region, overlapping with many mobile metabolite resonances (i.e., Glu, Gln, and Cr). These macromolecule signals, in combination with a low signal-to-noise ratio (SNR) and low metabolite resolution, increase the complexity of metabolite quantification (2,4) in short-TE 1H spectroscopy, and have limited its use in clinical diagnosis. Quantification accuracy and precision can be increased by fitting linear combinations of known metabolite lineshapes to the in vivo spectrum (5). With this approach, accurate metabolite levels can be measured when all contributions to the in vivo spectrum (metabolite and macromolecule) are correctly accounted for during fitting (6). Although the use of prior knowledge of metabolite lineshapes derived from in vitro samples has been successfully demonstrated by several groups (5–11), many recent short-TE 1H spectroscopy studies have not included patient-specific prior knowledge about the macromolecule component. To account for the macromolecule contribution to the spectrum, Provencher (5) proposed the use of a regularization algorithm that simultaneously fits metabolite lineshapes and a baseline spline function; however, the spline function does not exactly reproduce the macromolecule lineshape (12), potentially leading to inaccuracy in metabolite level measurements.
Alternatively, the macromolecule contribution to each in vivo spectrum can be separately measured and subtracted. Macromolecule resonances have a significantly shorter T1 (~250 ms (2)) compared to the mobile metabolite resonances (~1500 ms (13)) and thus can be acquired separately by inversion-nulling the metabolite signal (14,15). The use of multiple inversion pulses can produce adequate metabolite nulling for metabolites with a range of T1 values (2,14,15). Once it is acquired, the isolated mac- romolecule contribution can be subtracted from the original spectrum. The disadvantages of this approach include the addition of noise in the resulting spectrum, sensitivity to subject motion in unanesthetized human subjects, and longer total scan times. To overcome the third limitation, Tkac et al. (16) proposed the use of the same average macromolecule spectrum in the prior knowledge basis set. However, macromolecule resonances have been shown to fluctuate during disease conditions, making this approach inappropriate for patient studies (17). Another approach for eliminating the macromolecule signal is to acquire a metabolite-only spectrum with the macromolecule component nulled. However, this approach causes a significant reduction in the SNR of the acquired metabolite spectrum due to signal saturation. In addition, the macromolecule signal may contain important physiological information and should be considered an important component of an MRS exam.
Other sources outside the designated voxel may also produce unwanted artifacts in the spectrum. Seeger et al. (15) examined the source of lipid contamination in single- volume stimulated-echo acquisition mode (STEAM) spectroscopy with respect to the influence of remote out-of- slice excitation, and demonstrated that the spoiling of unwanted coherences alone was not sufficient to eliminate lipid contamination, even if the volume of interest (VOI) was placed 2 cm or more away from fat-containing regions. The additional application of outer volume presaturation in conjunction with single-voxel spectroscopy decreases contamination; however, the use of outer volume saturation techniques can also decrease the SNR of metabolites within the VOI due to magnetization transfer effects (18). In this study, spectra were acquired with localization by adiabatic selective refocusing (LASER), which has been shown to have excellent slice-selection profiles due to the use of multiple adiabatic full passage (AFP) pulses for the excitation of each orthogonal plane (19). Therefore, outer volume saturation was not required. The LASER sequence is also capable of reducing J-coupling modulation when the time between AFP pulses (Tcp) is kept short, despite the use of long TEs (20), and unlike STEAM it does not suffer a factor of 2 decrease in SNR compared to point-resolved spectroscopy (PRESS).
The purpose of this study was to design an optimized approach to quantify metabolite levels from in vivo short-TE 1H spectra integrating the subtraction of the macromolecule contribution from each spectrum. Although the subtraction of the macromolecule spectrum is conceptually straightforward, two methods (differing in sensitivity to subject motion) for acquiring the macromolecule spectrum were compared. In the first method a macromolecule spectrum was acquired after the full (metabolite + macromolecule) spectrum. In the second method, single averages of the macromolecule and full spectrum were alternated. The effect of reducing random noise in the final spectrum by first fitting the macromolecule spectrum using the Hankel Lanczos singular value decomposition (HLSVD) (21,22) and then subtracting the fit result, which contains no random noise, was also investigated. The HLSVD routine requires no user interaction and has been used successfully to fit other portions of the spectrum, such as the residual water (10,11,21,23). Although the HLSVD fit result contains no random noise, it does contain uncertainty associated with fit parameter estimates. The HLSVD routine models the spectrum using a series of damped exponential model functions, which is the most basic form of prior knowledge. Since the incorporation of prior knowledge in spectral fitting generally increases quantification precision (22), we hypothesized that HLSVD fitting of the macromolecule spectrum prior to subtraction would reduce the uncertainty associated with the macromolecule baseline estimate, and lead to increased metabolite quantification precision. Human parietal white matter and posterior hippocampus were studied because of their involvement in schizophrenia and epilepsy, which are of clinical interest in our laboratory. Measurements were made at 4.0 Tesla, which provided sufficient SNR to position a small (1.7 cm3) voxel within the hippocampus, thereby minimizing partial volume contamination. Spectral dispersion also increases with field strength, enhancing the resolution of overlapping metabolites such as Glu and Gln compared to that obtained at lower fields (24). Simulations were used to test the macromolecule subtraction process for a range of SNRs.
Twenty-one neurologically normal volunteers (10 males and 11 females, 35.3 ± 11.2 years old (mean ± SD)) participated in this study. All of the subjects gave informed consent, which was approved by the University of Western Ontario Health Sciences Research Ethics Board. Data were acquired using a 16-element quadrature hybrid birdcage RF coil on a 4.0 Tesla Varian Unity Inova whole-body MR imager equipped with a Siemens Sonata gradient coil. Each study began with a manual global shim using linear and Z2 shims followed by the acquisition of sagittal localizer images. These localizer images were used to plan two sets of volumetric images parallel to the hippocampus using a T1-weighted, inversion-prepared (Ti 500 ms), segmented turbo fast low-angle shot (FLASH) sequence to produce high gray matter/white matter contrast. The first set consisted of 64 slices (5 mm thick, TR/TE 10/5 ms) covering the entire brain, and the second set consisted of 16 slices (1 mm thick, TR/TE 11/6 ms) covering the hippocampus.
Spectroscopic data were acquired from a 2 x 2 x 2 cm3 voxel in the right parietal white matter (Fig. 1a) and a 1.3 x 1.3 x 1 cm3 voxel in the right posterior hippocampus of each subject (Fig. 1b), using the LASER localization sequence (19) (500 µs dwell time, 2 KHz receiver bandwidth, 1024 complex points, Tcp 6 ms) preceded by VAPOR (variable pulse power and optimized relaxation delays) water suppression (16). Lateralization to the right side was chosen arbitrarily. Magnetic field inhomogeneity was manually optimized within each voxel using only the linear shims. Three sets of data were acquired in each location: one spectrum contained the metabolite and macromolecule signals (full spectrum), one spectrum contained only the macromolecule signals (metabolite-nulled macromolecule spectrum), and one spectrum contained the unsuppressed water signal. Metabolite nulling was achieved with a two-pulse inversion technique (25,26), using AFP (HS2-R10 (19)) pulses rather than conventional 180° pulses. The time intervals between the first and second AFP inversion pulses (T1) and the second AFP inversion and the start of the LASER sequence (T2), which achieved maximum metabolite suppression, were determined empirically to be 2.20 s and 0.69 s, respectively.
Two data-averaging schemes were tested to determine the effect of subject motion on the macromolecule subtrac- tion efficiency. In the first method, complete full and mac- romolecule spectra were acquired in succession (TR/TE 3200/46 ms, 128 averages, 7-min acquisition/spectrum). In the second method, single acquisitions of the full and macromolecule spectra were alternated during a total acquisition time of 14 min. To maintain the same total acquisition time as the serial technique, the TR for the full and macromolecule acquisitions were modified to 2.15 s and 4.25 s, respectively. The unsuppressed water spectrum (used to remove lineshape distortions and as an internal standard for concentration referencing) was acquired last.
Data Processing and Macromolecule Subtraction
Spectra were postprocessed by combined QUALITY de-convolution and eddy current correction (QUECC, 27) with the junction between techniques set at 100 MS (10,27). Following lineshape correction, any remaining unsuppressed water was removed from the spectrum using HLSVD (21,22), which required no prior knowledge. Resonances between 4.1 and 5.1 ppm (water ~ 4.7 ppm), as determined by the HLSVD algorithm, were subtracted from the data (10,11,21,23).
Since the saturation of the macromolecule signal was different in the full spectrum compared to the macromolecule spectrum due to the different TRs used during their acquisitions, the macromolecule spectrum was scaled prior to subtraction. The scale factor was theoretically determined based on the steady-state solution to the Bloch equations (14) Mz M0 [(1-2exp (–T2/T1)) + 2exp (–(T1+ T2)/T1)]. The T1 of the macromolecule signal at 0.93 ppm was estimated in vivo by Behar et al. (2) to be 250 ± 36 ms at 2.1 Tesla. An ~10% increase in T1 is expected at 4.0 Tesla based on the scaling of metabolite T1’s between 1.5 T and 4.0 T. Therefore, setting T1 275 ms, and using T12200 ms and T2 686 ms, the expression above evaluates to Mz83 Mo. Therefore, the macromolecule spectrum was scaled by 1/0.83 ~ 1.2 prior to subtraction. This theoretically determined value was verified experi- mentally by varying the applied scale factor between 0.8 – 1.6 and measuring the standard deviation (SD) in the region of the spectrum between 0.75–1.8 ppm. This region was chosen because it contains exclusively macromolecule and lipid resonances in the normal brain spectrum. After the macromolecule signal is completely removed, the SD following subtraction in the specified region should equal the SD of the noise in the spectrum.
Two methods of macromolecule subtraction were tested. The first method involved the direct subtraction of the macromolecule spectrum (which added random noise to the resultant spectrum). The second method utilized the noninteractive HLSVD fitting routine to fit the macromol- ecule spectrum and then subtract the resultant fit. The fit spectrum contained no random noise, but was influenced by uncertainty in the fit parameter estimates as a result of the noise. The number of singular values and points used for HLSVD fitting were systematically varied to achieve the lowest SDs between 0.75–1.8 ppm.
Resultant metabolite spectra were fit in the time domain using a Levenberg-Marquardt minimization routine (28) incorporating prior knowledge from 19 metabolite line- shapes (12). Phase and delay time were included as parameters in the fitting (10,11), thereby eliminating user interaction, which can bias the results. The analysis soft- ware (Fitman (10,11)) was incorporated into a graphical user interface written in our laboratory in the IDL programming language (Version 5.4 Research Systems Inc., Boulder, CO).
The acquisition of prior knowledge has been described previously in detail (11). Prior knowledge for all metabolites was acquired using the same LASER pulse sequence used to acquire all in vivo data. Briefly, high-resolution in vitro spectra (7,8) were acquired from solutions (pH ad- justed to 7.04) of N-acetyl aspartate (NAA), glutamate (Glu), glutamine (Gln), -y-aminobutyric acid (GABA), as- partate (Asp), N–acetylaspartyl-glutamate (NAAG), taurine (Tau), glucose (Glc), phosphorylethanolamine (PEth), eth- anolamine (Eth), creatine (Cr), myo-inositol (Myo), glutathione (Glth), lactate (Lac), alanine (Ala), scyllo-inositol (Syl), glycerophosphorylcholine (GPC), and phosphoryl- choline (PC) (10,11). The use of GPC and PC rather than Cho resulted in a significant improvement of the in vivo fit in the region between 3.6 –3.7 ppm. The imposition of prior knowledge assumes that the in vitro metabolite resonances have the same chemical shifts, relative Lorentzian dampings, relative amplitudes, and phase modulations (due to J-coupling) as the corresponding in vivo metabolite resonances. Each of these characteristics was parameterized as described previously to constrain resonances dur- ing in vivo fitting (10,11). The scyllo-inositol (Syl) and glycine (Gly) singlets were modeled based on information in the literature (12).
Metabolite levels were normalized to the level of unsuppressed water within each voxel and then corrected for T1 and T2 relaxation effects using values from the literature. The average ratio of gray matter/white matter/CSF was estimated at ~ 90%/5%/5% in the hippocampus and 5%/ 90%/5% in the white matter voxels. The fractional water content was taken as 81% and 71% in the hippocampus and white matter, respectively (8). Table 1 summarizes the relaxation times used for the conversion calculations. T1 and T2 values for water, NAA, and Cr were taken from previous work at 4.0 Tesla (20,29,30) using the same LASER localization sequence with a similar interpulse time (Tcp ~ 6 ms). This sequence has been shown to reduce signal loss due to the diffusion of spins between regions of tissue with different magnetic susceptibility by limiting the time for diffusion between AFP pulses, resulting in T2 values at 4 Tesla approaching the T2 values typically observed at 1.5 Tesla measured with conventional Hann- echo sequences. Therefore, T2 values for PC/GPC, which have not been reported at 4.0 Tesla using the LASER sequence, were taken from the literature regarding 1.5 Tesla (31). The T1 and T2 values for the nondominant metabolites in the spectrum were approximated by taking the mean of the NAA, Cr, and PC/GPC relaxation times. Although studies have shown clear differences in water relaxation times in gray and white matter, there is a large overlapping variation in reported metabolite relaxation times in different tissue types (32); therefore, the same values were used for both tissues.
Table 1. T1 and T2 Relaxation Times Used for Conversion of Metabolite Levels to mM/liter VOI Units
Gray matter (hippocampus)
ªRelaxation times approximated as the average of NAA, CR, and PC/GPC.
Simulated data were generated to determine the effect of SNR on the macromolecule subtraction process and metabolite quantification precision. A single white matter full spectrum and macromolecule spectrum (both apodized by a 1 Hz exponential function) from one volunteer were used as the gold standard. Ten separate in vivo noise spectra were independently acquired from another volunteer by setting the amplitude of the RF pulses in the LASER sequence to zero. A set of simulated spectra was generated at one SNR by adding each of the 10 noise spectra to the same white matter spectrum. Ten similar series of simulated spectra were created with SNRs ranging from 8:1 to 60:1 by scaling the noise prior to adding it to the white matter spectrum. SNR was calculated using the intensity of the NAACH3 resonance divided by the SD of the noise on the spectral baseline (in the frequency domain) after Fourier transforming the entire data set. Each series was processed identically to the method described above to subtract the macromolecule component and fit the resultant spectrum using the prior knowledge template. Metabolite levels, SDs, and Crame´r-Rao lower bounds (33) on quantification precision were determined for each series. The Crame´r- Rao lower bounds (33) provide a useful estimate of the limit of quantification reliability.
After the macromolecule spectrum was subtracted, the SD of the baseline (0.75–1.8 ppm) was compared to the SD of the baseline in the original full spectrum using analysis of variance (ANOVA). This comparison was made for both serial and interleaved data acquisition schemes. ANOVA was also used to compare quantified metabolite levels after direct macromolecule subtraction and after subtracting the HLSVD fit of the macromolecule spectrum. Finally, metab- olite levels were compared between white matter and the hippocampus (HLSVD fit subtraction method) using ANOVA. In all cases, P-values < 0.05 were considered statistically significant.
Figure 1 shows the location of the right parietal white matter (Fig. 1a) and the right posterior hippocampus (Fig. 1b) voxels studied in one volunteer. Both full and metab- olite-suppressed macromolecule spectra were successfully acquired in 21 subjects in the right parietal white matter and in 19 subjects in the right posterior hippocampus. The average SNR and full width at half maximum (FWHM) of the unsuppressed water in the white matter data were 44 ± 11 and 9.4 ± 1.7 Hz, respectively; in the posterior hippocampus they were 9 ± 3 and 10.2 ± 1.5 Hz, respectively. The shape of the macromolecule spectrum in the region between 0.75–1.8 ppm did not match the shape of the full spectrum in the serially acquired data (11 subjects). This difference in shape precluded the accurate subtraction of the macromolecule component from the full spectrum. A comparison of the SD in the region of the spectrum be- tween 0.75–1.8 ppm showed no differences before or after direct macromolecule subtraction (SDbefore subtraction 2.8 x 10–3 ± 1.0 x 10–3 vs. SDafter subtraction 2.7 x 10–3 ± 0.5 x 10–3, P 0.88).
FIG. 1. The position of (a) a 2 x 2 x 2 cm3 voxel in right parietal white matter, and (b) a 1.3 x 1.3 x 1 cm3 voxel in the right posterior hippocampus in the same subject.
In contrast, Fig. 2 demonstrates the similarity in the 0.75–1.8 ppm region between the full and macromolecule spectrum for the interleaved acquisition method (10 subjects). Individual spectra are identified in the figure caption. The macromolecule component was successfully subtracted in all subjects when the interleaved acquisition method was used. After the subtraction, the SD between 0.75–1.8 ppm approached the SD of the noise on the base- line (1.6 x 10–3) measured in the region of the spectrum between 7.6 –9.9 ppm (SDbefore subtraction 2.5 x 10–3 ± 0.7 x 10–3 vs. SDafter subtraction 1.7 x 10–3 ± 0.1 x 10–4, P 0.05).
FIG. 2. The consistency of the macromolecule subtraction is demonstrated. (a) A right parietal white matter full spectrum and (b) a macromolecule spectrum from an interleaved acquisition are shown following QUECC lineshape correction. c: The corresponding macromolecule spectrum after fitting with HLSVD. d: The result of the macromolecule subtraction is shown (spectrum a minus spectrum c) for one volunteer following residual water removal. e and f: Spectra for two other volunteers demonstrate the consistency of the macromolecule subtraction. g: A right posterior hippocampus spectrum from one volunteer. The dotted lines represent the spectral region used to calculate the baseline SD in Fig. 3.
Figure 3 shows the average SD of the region between 0.75–1.8 ppm in the original full spectra (scale factor 0) and after macromolecule subtraction using the scale factors 0.8, 1, 1.2, 1.4, and 1.6 for the white matter data. Figure 3a shows the results when the macromolecule spectrum was subtracted directly. Figure 3b shows the results when the HLSVD fit of the macromolecule spectrum was subtracted. The horizontal gray dashed line represents the SD of the noise calculated in another part of the spectrum (7.6 –9.9 ppm) after subtraction. The results indicate that optimum subtraction of the macromolecule component (lowest SD) is achieved with a scale factor of 1–1.2, in agreement with the theoretical value of 1.2 determined with the steady-state solution to the Bloch equations. The use of the HLSVD fit rather than the original spectrum to subtract the macromolecule signals from the full spectrum resulted in an 87% decrease in the SD of the spectrum between 0.75–1.8 ppm, and a 27% increase in the SNR (considering random noise only) in the resultant spectrum. Optimal HLSVD fits were achieved by using the first 128 – 256 points of the time domain signal and choosing the maximum number of singular values detected by the algorithm.
FIG. 3. The average baseline SD in the white matter spectra (N 10) from the region between 0.75–1.80 ppm following the scaling and subtraction of the macromolecule signal. Error bars represent 1 SD. Results are shown for (a) the direct subtraction of the macromolecule spectra and (b) the subtraction of the HLSVD fit macro- molecule spectra. The average SD in the spectra before subtraction is depicted at scale factor 0. The horizontal dashed lines represent the SD of the noise on the baseline (7.6 –9.9 ppm) after subtraction. Asterisks indicate the scale factors that produced a significant decrease in the SD of the 0.75–1.8 ppm region of the spectrum (P < 0.05) after macromolecule subtraction.
Figure 4a shows the white matter spectrum from one volunteer following lineshape correction, residual water removal, and macromolecule subtraction (corresponds to the spectrum in Fig. 2d), with the linear combination fit result superimposed on it. The residual is shown beneath (Fig. 4b). Individual metabolite components contributing to the in vivo spectrum are also displayed in Fig. 4c–p.
FIG. 4. The result of linear combination fitting (dark line) is shown superimposed on the white matter spectrum (light line) of one vol- unteer in part a above (b) the residual. Below are the individual metabolite components, which were visible in this spectrum: (c) NAA, (d) Glu, (e) Gln, (f) Glth, (g) Asp, (h) NAAG, (i) Tau, (j) Glc, (k) Cr, (l) PC, (m) GPC, (n) Myo, (o) Gly, and (p) Syl.
The relationships between SNR and quantified levels of NAA, Cr, Glu, Myo, and Gln in simulated data following direct macromolecule subtraction and subtraction of the HLSVD fit macromolecule spectrum are shown in Fig. 5a and d. These metabolites represent the dominant resonances in the spectrum. The corresponding relationships between metabolite measurement uncertainties (% SD) and SNR are shown in Fig. 5b and e, and the corresponding change in % Crame´r-Rao SD as a function of SNR is shown in Fig. 5c. Generally, SD and Crame´r-Rao uncertainty increased as SNR decreased, while estimated metabolite levels remained relatively constant. Comparing the SDs between methods, similar results were observed above SNR of 20:1. However, below 20:1, the SDs were notably lower when the macromolecule spectrum was fit prior to subtraction. Calculation of the Crame´r-Rao lower bounds on measurement precision involves an estimate of the SD of the noise in the signal. Since noise in the fit subtracted spectrum does not incorporate the uncertainty in parameter estimates introduced by the HLSVD fitting, the Crame´r-Rao lower bounds would be underestimated. Therefore, the Crame´r-Rao bounds calculated from the data obtained by direct subtraction should be considered the minimum variance for both techniques.
FIG. 5. SNR simulation results are displayed following (a– c) direct macromolecule subtraction and (d– f) subtraction of the HLSVD fit of the macromolecule spectrum. a and d: Average quantified metabolite levels are shown for five prominent metabolites as a function of SNR, which was calculated on the full spectrum before subtraction. b and e: The corresponding % SDs. c: The corresponding % Crame´ r-Rao SDs. Vertical dotted lines indicate the average SNR of the in vivo hippocampus (9:1) and white matter (44:1) full spectra prior to subtraction.
Tables 2 and 3 summarize the quantified metabolite levels, SDs, and Crame´r-Rao SD from the right parietal white matter and right posterior hippocampus, respectively. There were no significant differences in metabolite levels between subtraction techniques. Of note, no in- crease in quantification precision was associated with the HLSVD fit subtraction method in the white matter despite the 27% increase in the apparent SNR for these data. Except for NAA, metabolite levels were generally higher (identified by superscript b in Table 3) in the hippocampus compared to white matter. Cr, Myo, and Glc remained significantly elevated following a Bonferroni correction for multiple comparisons.
Table 2. Metabolite Levels in Parietal White Matter (N 10)
|levelª ± SD||level ± SD||SD|
|NAA||9.37 ± 0.95||9.32 ± 0.96||0.17|
|Glu||5.49 ± 0.54||5.47 ± 0.56||0.09|
|Gln||1.84 ± 0.53||1.92 ± 0.35||0.11|
|Asp||3.16 ± 0.59||3.38 ± 0.45||0.20|
|NAAG||0.42 ± 0.46||0.60 ± 0.70||0.19|
|Tau||0.61 ± 0.54||0.65 ± 0.53||0.10|
|Glc||3.32 ± 0.66||3.37 ± 0.71||0.09|
|Eth||0.14 ± 0.27||0.05 ± 0.12||0.23|
|Cr||7.36 ± 0.79||7.42 ± 0.80||0.14|
|Myo||5.64 ± 0.72||5.66 ± 0.68||0.09|
|Glth||0.98 ± 0.31||1.10 ± 0.29||0.07|
|Lac||0.15 ± 0.23||0.15 ± 0.20||0.18|
|Ala||0.33 ± 0.35||0.33 ± 0.34||0.15|
|Syl||0.28 ± 0.14||0.31 ± 0.13||0.09|
|Gly||0.20 ± 0.28||0.31 ± 0.25||0.46|
|GPC||1.22 ± 0.28||1.17 ± 0.25||0.08|
|PC||0.98 ± 0.11||1.07 ± 0.11||0.24|
ªLevels reported in mM/liter VOI-incorporating T1 and T2 relaxation times taken from the literature for NAA, Cr, and GPC/PC, and approximated for all other metabolites (Table 1). SD, standard deviation; n.d., not detected.
Table 3. Metabolite Levels in Posterior Hippocampus (N 8)
|levelª ± SD||level (mM) ± SD||SD|
|NAA||9.20 ± 1.96||9.18 ± 2.13||0.72|
|Glu||8.03 ± 2.11||7.67 ± 2.39||0.57|
|Gln||3.92 ± 1.38||4.08 ± 2.14b||0.60|
|GABA||0.98 ± 1.54||1.24 ± 2.31||0.97|
|Asp||5.46 ± 4.46||3.97 ± 3.39||1.10|
|NAAG||0.59 ± 1.31||0.51 ± 1.03||1.10|
|Tau||2.68 ± 1.66||2.07 ± 1.69b||0.50|
|Glc||5.21 ± 2.44||5.37 ± 1.06b||0.52|
|Eth||0.55 ± 0.92||0.20 ± 0.56||1.46|
|Cr||11.64 ± 2.23||11.71 ± 2.55b||0.78|
|Myo||9.78 ± 1.34||10.28 ± 1.32||0.54|
|Glth||1.04 ± 0.74||0.97 ± 0.76||0.36|
|Lac||1.05 ± 1.42||1.18 ± 1.32b||1.04|
|Ala||0.45 ± 0.81||0.68 ± 0.86||0.82|
|Syl||0.73 ± 0.60||0.60 ± 0.50||0.56|
|Gly||1.84 ± 2.08||2.06 ± 1.89b||2.61|
|GPC||1.86 ± 1.25||1.97 ± 1.36||0.46|
|PC||1.71 ± 0.75||1.66 ± 0.82||1.23|
ªLevels reported in mM/liter VOI-incorporating T1 and T2 relaxation times taken from the literature for NAA, Cr, and GPC/PC, and approximated for all other metabolites (Table 1). bSignificant increase in metabolite level (P < 0.05) compared to parietal white matter. SD, standard deviation; n.d., not detected.
In this study, a separately acquired metabolite-nulled macromolecule spectrum was used to remove the macromolecule contribution to the LASER-localized short-TE 1H spectrum on a case by case basis. The remaining information in the spectrum was attributed exclusively to metabolite resonances, and quantified using a linear combination fitting technique. Metabolite levels were quantified from parietal white matter and posterior hippocampus volumes. Quantified levels of NAA, Cr, and Cho were consistent with previous long-TE studies, which are not complicated by macromolecule resonances (32). Levels of most other metabolites, including Glu and Gln, were within the normal range (6,32,34,35), which demonstrates the importance of incorporating all spectral signals into the quantification process.
For volume selection, the LASER localization sequence was applied (19), which uses only AFP pulses to excite three orthogonal planes. These adiabatic (HS2-R10) pulses have excellent slice-selection profiles (19) compared to STEAM and PRESS. Therefore contamination from lipid signals outside the region of interest, which can fluctuate due to subject motion, is reduced. Additional outer volume suppression was not required to reduce lipid signal contamination in the spectrum, which avoided metabolite signal reduction within the voxel due to magnetization transfer effects. The only disadvantage of the LASER sequence is its high power requirements. In this study, a quadrature hybrid birdcage volume coil was required to study deep brain regions. The maximum power used was ~2.5 W/Kg, which is within the FDA guidelines of <3.2 W/Kg.
The large difference in metabolite and macromolecule T1’s (~ 1500 ms vs. ~ 250 –300 ms, respectively) was exploited to selectively acquire the macromolecule signal by T1-nulling the metabolite signal. In our first attempt to apply this technique in a group of human subjects, a full spectrum was acquired followed by a macromolecule spectrum (the serial method). This approach yielded macromolecule spectra that did not resemble the macromolecule signals in the full spectrum in the region < 2.0 ppm. Therefore the macromolecule signals could not be properly subtracted from the full spectrum. The most likely reason for this discrepancy was subject motion between the acquisitions of the full and macromolecule data. To reduce motion artifacts an interleaved acquisition approach was designed. This method alternated the acquisitions of the full and macromolecule spectra, thereby minimizing spectral differences between acquisitions due to motion despite the long (~14 min) total acquisition time. The interleaved approach resulted in a macromolecule spectrum that closely resembled the macromolecule signal in the full spectrum, which allowed the removal of the macromolecule component of the spectrum and left a flat baseline in the region < 2.0 ppm. The theoretically calculated scale factor of 1.2, applied to the macromolecule spectrum to account for saturation differences between the full and macromolecule spectra, was experimentally con- firmed (Fig. 3). The success of the interleaved data acquisition compared to the serial data acquisition illustrates the sensitivity of this subtraction technique to patient motion. Since interleaved acquisitions of full and macromolecule spectra are acquired in closer temporal proximity, differences in subject position are minimized between acquisitions.
Although the subtraction of the macromolecule component of the spectrum is essential for optimum metabolite quantification accuracy and precision, the subtraction of two spectra also increases the random noise in the resultant spectrum. Therefore the direct subtraction of the macromolecule spectrum from the full spectrum imposes a SNR penalty, which can reduce metabolite quantification precision. Fitting the macromolecule spectrum and then using the fit result for subtraction reduces the random noise fluctuations in the resultant spectrum, and thus increases the apparent SNR (a 27% increase in the white matter data was observed). However, the uncertainty arising from the noise is incorporated in the uncertainty-of-fit parameter estimates. Therefore, following subtraction of the fit spectrum, which does not appear noisy in the usual sense, the normal calculation of SNR does not correctly account for spectral uncertainty. Imposing prior knowl- edge during spectral fitting generally leads to increased quantification precision (22). Since the HLSVD routine imposes an exponentially damped sinusoid model on the macromolecule spectrum, the fit representation of this spectrum was expected to contain less uncertainty than the original data.
The simulations summarized in Fig. 5 demonstrate that quantified metabolite levels remained constant for SNR > 20:1 (Fig. 5a and d) and began to deteriorate notably at SNR < 10:1 for both the direct subtraction and fit subtraction approaches. The uncertainty (% SD) associated with the metabolite measurements also increased as SNR decreased, as expected. The percent SDs associated with metabolite measurements were similar between the direct subtraction and fit subtraction techniques at SNR > 20:1, consistent with the in vivo results from white matter. In contrast, at SNR < 20:1, the fit subtraction technique simulations demonstrated lower % SDs compared to the direct subtraction technique, which was not observed in the in vivo hippocampus data. This discrepancy can be attributed to a reduction in sensitivity due to the additional biological variation present in the in vivo data.
Metabolite levels were quantified in mM/L VOI and corrected for the effects of T1 and T2 relaxation. Relaxation times were taken from the literature for NAA, Cr, and GPC/PC, and approximated for all other metabolites (Table 1) since these values have not been measured. In the parietal lobe white matter, the metabolite levels reported were generally consistent with levels previously reported in both the MR literature (in vivo and extract studies) and other physiological measurements (6,32,34,35). Measured levels of NAA (9.4 mM/L VOI), Cr (7.4 mM/L VOI), and Myo (5.6 mM/L VOI) were consistent with previous in vivo spectroscopy measurements and high-resolution extract studies (32,34). The level of total Cho (GPC + PC, 2.2 mM/L VOI) was slightly higher than previously reported (32), with a GPC : PC ratio of approximately 1:1. The level of Glu (5.5 mM/L VOI) was slightly lower than previous in vivo measurements (34,36), but within the lower range for concentrations reported from in vitro studies of biopsy materials (37). The lower levels of Glu measured in this study (compared to previous studies) may also be due to the inclusion of Glth in the prior knowledge template, which overlaps with Glu in the spectrum. Failure to separate the contributions from these metabolites in the past, as well as the underlying macromolecule components, may have lead to an overestimation of Glu. The level of Gln was approximately one-third that of Glu. GABA was not detected in white matter, consistent with [3H]GABA binding studies in human brain (38). PEth was also not detected in the parietal white matter. The level of glucose measured in the white matter (3.4 ± 0.7 mM/L VOI) was notably higher than expected based on previous 1H and 13C in vivo spectroscopy measurements. For example, Gruetter et al. (39) measured brain glucose levels of ~1 mM/L VOI in human visual cortex at euglycemia using the 5.23 ppm glucose resonance in the 1H spectrum. In this study, the glucose resonances between 3.4 –3.9 ppm were used to estimate glucose concentration (Fig. 4). The over- estimation of glucose in the present study compared to previous in vivo measurements may be explained by a systematic error in macromolecule removal. One possibility is that the T1 of the macromolecule signal in the 3.4 – 3.9 ppm region is significantly longer than the T1 of the macromolecule in the region of the spectrum < 2 ppm. A careful characterization of macromolecule T1 is required to make this determination.
Metabolite levels were also quantified in the posterior hippocampus. The small voxel size (1.3 x 1.3 x 1 cm3) used in this study ensured that data were acquired with minimal partial volume contamination. The level of NAA observed in the hippocampus gray matter was virtually identical to that observed in the white matter (9.2 mM/L VOI). This level is consistent with other previous in vivo 1H spectroscopy studies (32). Most other metabolites were found to have a much higher concentration compared to parietal white matter, which can be verified visually by comparing spectra from white matter and the hippocampus. For example, Fig. 2d–f represent white matter spectra, while Fig. 2g represents a hippocampus spectrum. Since the NAA levels in both are similar, a comparison of the relative intensities of the Cr, GPC/PC, Myo, and Glu peaks suggests that the concentration of these metabolites is higher in the hippocampus than in white matter, consistent with a previous report (35).
The small voxel size used in the right posterior hippocampus to avoid partial volume contamination resulted in a SNR for hippocampus data < 10:1. The simulations shown in Fig. 5 demonstrate that some variability in metabolite levels can be attributed to the low SNR of these data. However, the higher metabolite levels observed in the posterior hippocampus compared to white matter (i.e., Glu: 40% increase; Gln: 113% increase; Myo: 82% in- crease; and Cr: 58% increase) cannot be solely attributed to this increased uncertainty. In fact, Cr, Myo, and Glc remained significantly elevated after a Bonferroni correction for multiple statistical comparisons was applied. The higher Cr levels in this region and the similarity of NAA levels with white matter suggest that partial volume correction must be incorporated into studies that report NAA ratios to Cr or total Cho when voxels include notable contributions from both gray and white matter.
The two-pulse metabolite-nulling sequence used in this study resulted in excellent metabolite suppression despite small differences in metabolite T1’s. However, some quantification bias occurs with this approach because all metabolite signals cannot be identically suppressed. The amount of residual signal was estimated based on the steady-state solution of the Bloch equations and the sequence timing. Metabolites with T1 1500 ms had ~3% residual signal in the macromolecule spectrum, while metabolites with T1 1200 ms or 2000 ms had absolute residual signal intensities of ~5%, resulting in a maxi-mum error of ~6% after macromolecule subtraction. Conversely, macromolecule signals with T1 > 275 ms (estimated for the 0.93-ppm macromolecule) will be partially nulled. For example, a macromolecule signal with a 400-ms T1 would be underestimated by ~23% following macromolecule subtraction. Underestimating the macromolecule component would lead to an overestimation of overlapping metabolite concentrations in the final analysis.
Metabolite levels were quantified in parietal white matter and the posterior hippocampus using LASER-localized proton short-TE 1H spectroscopy, incorporating macromolecule subtraction. Optimum macromolecule subtraction (87% reduction) was achieved when single averages of full and macromolecule spectra were alternately acquired. The use of HLSVD to fit the macromolecule spectrum prior to subtraction increased the apparent SNR of the resultant spectrum by 27% in white matter and did not lead to any differences in quantified metabolite levels, but did not increase quantification precision. The separate acquisition and subtraction of macromolecule signals from the short-TE 1H spectrum resulted in accurate and consistent metabolite measurements, and is suitable for individual patient studies in which macromolecule signals may fluctuate. Quantified metabolite levels in the posterior hippocampus were significantly higher than those in parietal white matter for most metabolites (except NAA), which emphasizes the need to control for partial volume contamination when studying this region.
The authors thank Dr. Michael Garwood for providing the original LASER pulse sequence, and Dr. Ronald deBeer for providing the HLSVD fitting routine. The authors also thank Jeff Herbynchuk and John-Paul Lobos for software development, and Lisa Jong for data analysis.
- Ross B, Michaelis T. Clinical applications of magnetic resonance spec- Magn Reson Q 1994;10:191–247.
- Behar KL, Rothman DL, Spencer DD, Petroff OA. Analysis of macro- molecule resonances in 1H NMR spectra of human brain. Magn Reson Med 1994;32:294 –302.
- Behar KL, Ogino Assignment of resonance in the 1H spectrum of rat brain by two-dimensional shift correlated and J-resolved NMR spec- troscopy. Magn Reson Med 1991;17:285–303.
- Graham GD, Hwang JH, Rothman DL, Prichard JW. Spectroscopic as- sessment of alterations in macromolecule and small-molecule metabo- lites in human brain after stroke. Stroke 2001;32:2797–2802.
- Provencher SW. Estimation of metabolite concentrations from localized in vivo proton NMR Magn Reson Med 1993;30:672– 679.
- Hofmann L, Slotboom J, Jung B, Maloca P, Boesch C, Kreis R. Quanti- tative 1H-magnetic resonance spectroscopy of human brain: influence of composition and parameterization of the basis set in linear combi- nation model-fitting. Magn Reson Med 2002;48:440 – 453.
- de Graaf AA, Bovee WM. Improved quantification of in vivo 1H NMR spectra by optimization of signal acquisition and processing and by incorporation of prior knowledge into the spectral Magn Reson Med 1990;15:305–319.
- Stanley JA, Drost DJ, Williamson PC, Thompson RT. The use of a priori knowledge to quantify short echo in vivo 1H MR spectra. Magn Reson Med 1995;34:17–24.
- Mierisova S, van den Boogaart A, Tkac I, Van Hecke P, Vanhamme L, Liptaj T. New approach for quantitation of short echo time in vivo 1H MR spectra of brain using NMR Biomed 1998;11:32–39.
- Bartha R, Drost DJ, Menon RS, Williamson PC. Comparison of the quantification precision of human short echo time (1)H spectroscopy at 1.5 and 4.0 Tesla. Magn Reson Med 2000;44:185–192.
- Bartha R, Drost DJ, Williamson Factors affecting the quantification of short echo in vivo 1H MR spectra: prior knowledge, peak elimina- tion, and filtering. NMR Biomed 1999;12:205–216.
- Pfeuffer J, Tkac I, Provencher SW, Gruetter R. Toward an in vivo neurochemical profile: quantification of 18 metabolites in short-echo- time (1)H NMR spectra of the rat brain. J Magn Reson 1999;141:104 –
- Posse S, Cuenod CA, Risinger R, Le Bihan D, Balaban RS. Anomalous transverse relaxation in 1H spectroscopy in human brain at 4 Tesla. Magn Reson Med 1995;33:246 –252.
- Knight-Scott J. Application of multiple inversion recovery for suppres- sion of macromolecule resonances in short echo time (1)H NMR spec- troscopy of human J Magn Reson 1999;140:228 –234.
- Seeger U, Mader I, Nagele T, Grodd W, Lutz O, Klose U. Reliable detection of macromolecules in single-volume 1H NMR spectra of the human brain. Magn Reson Med 2001;45:948 –954.
- Tkac I, Starcuk Z, Choi IY, Gruetter R. In vivo 1H NMR spectroscopy of rat brain at 1 ms echo Magn Reson Med 1999;41:649 – 656.
- Hwang JH, Graham GD, Behar KL, Alger JR, Prichard JW, Rothman Short echo time proton magnetic resonance spectroscopic imaging of macromolecule and metabolite signal intensities in the human brain. Magn Reson Med 1996;35:633– 639.
- Luo Y, de Graaf RA, DelaBarre L, Tannus A, Garwood M. BISTRO: an outer-volume suppression method that tolerates RF field inhomogene- ity. Magn Reson Med 2001;45:1095–1102.
- Garwood M, DelaBarre L. The return of the frequency sweep: designing adiabatic pulses for contemporary NMR. J Magn Reson 2001;153:155–
- Michaeli S, Garwood M, Zhu XH, DelaBarre L, Andersen P, Adriany G, Merkle H, Ugurbil K, Chen W. Proton T2 relaxation study of water, N-acetylaspartate, and creatine in human brain using Hahn and Carr- Purcell spin echoes at 4T and Magn Reson Med 2002;47:629 – 633.
- van den Boogaart A, Ala-Korpela M, Jokisaari J, Griffiths JR. Time and frequency domain analysis of NMR data compared: an application to 1D 1H spectra of lipoproteins. Magn Reson Med 1994;31:347–358.
- de Beer R, Ormondt D. Analysis of NMR data using time domain fitting NMR Basic Princ Progr 1992;26:201–248.
- Cabanes E, Confort-Gouny S, Le Fur Y, Simond G, Cozzone PJ. Opti- mization of residual water signal removal by HLSVD on simulated short echo time proton MR spectra of the human brain. J Magn Reson 2001;150:116 –125.
- Tkac I, Andersen P, Adriany G, Merkle H, Ugurbil K, Gruetter R. In vivo 1H NMR spectroscopy of the human brain at 7 T. Magn Reson Med 2001;46:451–
- Dixon WT, Sardashti M, Castillo M, Stomp GP. Multiple inversion recovery reduces static tissue signal in angiograms. Magn Reson Med 1991;18:257–268.
- Starcuk Z, Tkac I, Starcuk Zj, Horky J. Measurement of macromolecular or metabolite nulled in vivo proton NMR spectra of the rat brain. In: Proceedings of the 7th Annual Meeting of SMR, Philadelphia, 1999. p
- Bartha R, Drost DJ, Menon RS, Williamson Spectroscopic lineshape correction by QUECC: combined QUALITY deconvolution and eddy current correction. Magn Reson Med 2000;44:641– 645.
- Marquardt DW. An algorithm for least-squares estimations of nonlinear parameters. J Soc Indust Appl Math 1963;11:431–
- Bartha R, Michaeli S, Merkle H, Adriany G, Andersen P, Chen W, Ugurbil K, Garwood M. In vivo 1H2O T2+ measurement in the human occipital lobe at 4T and 7T by Carr-Purcell MRI: detection of micro- scopic susceptibility Magn Reson Med 2002;47:742–750.
- Mason GF, Chu WJ, Hetherington HP. A general approach to error estimation and optimized experiment design, applied to multislice imaging of T1 in human brain at 1 T. J Magn Reson 1997;126:18 –29.
- Choi CG, Frahm J. Localized proton MRS of the human hippocampus: metabolite concentrations and relaxation times. Magn Reson Med 1999; 41:204 –207.
- Kreis R. Quantitative localized 1H MR spectroscopy for clinical use. J Prog NMRS 1997;31:155–195.
- Cavassila S, Deval S, Huegen C, van Ormondt D, Graveron-Demilly Cramer-Rao bounds: an evaluation tool for quantitation. NMR Biomed 2001;14:278 –283.
- Michaelis T, Merboldt KD, Bruhn H, Hanicke W, Frahm J. Absolute concentrations of metabolites in the adult human brain in vivo: quan- tification of localized proton MR spectra. Radiology 1993;187:219 –227.
- Hajek M, Burian M, Dezortova M. Application of LCModel for quality control and quantitative in vivo 1H MR spectroscopy by short echo time STEAM sequence. MAGMA 2000;10:6 –17.
- Pouwels PJ, Frahm J. Regional metabolite concentrations in human brain as determined by quantitative localized proton MRS. Magn Reson Med 1998;39:53–
- Perry TL, Hansen S, Berry K, Mok C, Lesk D. Free amino acids and related compounds in biopsies of human J Neurochem 1971;18: 521–528.
- Iloyd KG, Shemen L, Hornykiewicz O. Distribution of high affinity sodium-independent [3H]gamma-aminobutyric acid [3H]GABA bind- ing in the human brain: alterations in Parkinson’s disease. Brain Res 1977;127:269 –278.
- Gruetter R, Ugurbil K, Seaquist ER. Steady-state cerebral glucose concentrations and transport in the human brain. J Neurochem 1998;70: 397–408.