|Home | About | Journals | Submit | Contact Us | Français|
To develop and optimize a 1H MRS method for measuring brain glutathione (GSH) levels.
Phantom experiments and density operator simulations were performed to determine the optimal TE for measuring GSH at 3 T using J-difference spectral editing. In vivo data collected from eleven normal volunteers (forty-three measurements) and five stroke patients (ten measurements) were processed using a new spectral alignment method (adaptive spectral registration).
In phantom experiments and density operator simulations where relaxation effects were ignored, close to maximum GSH signal (2.95 ppm) was obtained at TE ≈ 131 ms with minimum N-acetyle-aspartate (NAA) signal interference. Using adaptive spectral registration, GSH levels in healthy volunteers were found to be 1.20 ± 0.14 mM (mean ± SD). GSH levels in stroke patients were found to be 1.19 ± 0.24 mM in lesion and 1.25 ± 0.19 mM in contralateral normal tissue. In comparison, the standard deviations were significantly larger when only the NAA singlet (2.01 ppm) was used as a navigator for spectral alignment.
Spectral editing using J-differences is a reliable method for measuring GSH levels in volunteers and stroke patients.
Glutathione (GSH) is an antioxidant and detoxifier in the human body that plays an important role in defense against oxidative stress. Acute ischemic stroke is associated with significant oxidative stress and subsequently precipitates changes in the glutathione system. Clinical studies have shown that subjects with risk factors for stroke have relatively low levels of GSH, and patients with acute ischemic stroke develop elevated blood GSH levels during the first hours to days post ictus (1,2). However, previous studies of the GSH system in stroke have focused on its changes in the blood or urine. In vivo measurement of GSH in stroke patients by magnetic resonance spectroscopy (MRS) can provide spatially localized GSH levels in lesion and normal brain tissue, which could be valuable in predicting clinical outcome after stroke and evaluating the efficacy of drug treatments.
Several in vivo 1H spectroscopy techniques for GSH measurement have been reported. One of the techniques extracts GSH concentration from LCModel analysis of short echo time spectra (3,4). This technique requires high quality spectra to eliminate ambiguity in separating GSH from other overlapping resonances such as creatine (Cr), γ-aminobutyric acid (GABA), and macromolecules. Editing techniques have been widely used to detect targeted metabolite by eliminating overlapping resonances. Two editing techniques have been adopted for measuring GSH in vivo, which are double quantum coherence filtering (DQF) and J-difference spectroscopy (5–13). Both techniques measure the concentration of the cysteinyl β-protons resonating around 2.95 ppm by exploiting their J-coupling to the cysteinyl α-proton resonating at 4.56 ppm.
J-difference spectroscopy is conceptually simpler than DQF and preserves singlet resonances in both the “on” and “off” spectra, whereas DQF needs a second echo to obtain the singlet resonances (14). The singlet resonances can serve as internal concentration references. When J-difference editing is combined with PRESS localization (15), the pulse sequence can be readily implemented on clinical scanners by modifying the existing PRESS sequence. A variety of TE values (68–136 ms) have been used for in vivo GSH detection using J-difference editing. A TE of 68 ms (7) has been used on a 4 T scanner and TE values of 80 ms (8) and 94 ms (12) have been used on 3 T scanners. We have proposed to measure GSH at 3 T with TE = 131 ms (10). Meanwhile, GSH detection at 7 T with TE = 136 ms has also been reported (13).
The goal of this work was to test the feasibility of using J-difference editing to reliably measure GSH levels in stroke patients in a typical clinical setting. A major challenge for this work and for subtraction-based J-difference editing spectroscopy in general comes from subtraction errors. Subtraction between the “on” and “off” acquisitions often results in subtraction errors due to subject motion, instrumental instabilities, and different residual water signals between the “on” and “off” acquisitions. Since the frequency of cysteinyl α-proton (4.56 ppm) is very close to water frequency (4.68 ppm), the baseline differences between the “on” and “off” acquisitions can be very large. One way to reduce subtraction errors is to bring down the baseline errors by using a longer TE. If the GSH signal does not become much smaller or even become larger at a longer TE, the overall quality of the GSH difference spectrum would be improved. In addition to using a longer TE, an improved post-processing technique could further reduce subtraction errors. It has been shown that frequency and phase corrections are important in reducing subtraction errors (16). Currently, frequency and phase corrections for processing J-difference spectra are based on the alignment of a singlet peak such as the N-acetyle-aspartate (NAA) peak at 2.01 ppm. The singlet peak signal from a single-acquisition (7) or a Lorentzian curve (16) is used as the reference for the spectral alignment. The quality of the alignment could be compromised if the baseline differences are too large or the singlet peak has a poor signal-to-noise ratio (SNR). In stroke patients, for example, the NAA singlet peak in lesions can have poor SNR due to reduced NAA concentration levels. A new spectral alignment method (adaptive spectral registration) was developed to perform frequency, phase, and linear baseline corrections in order to make the post-processing method less susceptible to errors. In this article, TE optimization and spectral alignment based on adaptive spectral registration are described. Comparison with a conventional singlet-based frequency and phase correction method is also discussed using data collected from both normal volunteers and stroke patients.
A PRESS-based J-difference editing pulse sequence (7) with rearranged crusher gradients (10) was implemented on a 3 T Philips Achieva whole-body scanner (Philips Medical Systems, Best, The Netherlands). A standard 30 cm diameter T/R volume head coil was used to perform the experiments. For the purpose of understanding the dependence of GSH peak amplitude versus TE, a 20 mM GSH phantom was scanned nine times with different TE values incrementing from 70 ms to 150 ms. A 5 × 3 × 3 cm3 volume of interest (VOI) in the center of the phantom was localized. The selective inversion pulses were Gaussian pulses with 15.5 ms duration and 80 Hz bandwidth at half-height. A TR of 2 s was used and 64 acquisitions were performed in each scan. The GSH difference spectra were apodized with Lorentzian linebroadening of 9 Hz. The GSH phantom experiments were also simulated using the GAMMA C++ libraries (17). TE was varied from 55 ms to 170 ms with 5 ms increment. Relaxation effects were ignored in the simulations and the difference spectra were apodized with Lorentzian linebroadening of 9 Hz.
For evaluating the spectral interference from the co-edited aspartate moiety of NAA signals, NAA difference spectra were simulated with TE = 126, 131, and 136 ms using the same pulse sequence parameters as for the in vivo experiments. Relaxation effects were ignored in the simulations and the difference spectra were apodized with Lorentzian linebroadening of 7 Hz.
Eleven normal volunteers and five stroke patients were studied, all of whom gave informed consent in accordance with procedures approved by the institutional review board. The same pulse sequence was used for scanning both normal volunteers and stroke patients, which had a TR of 2 s, a TE of 131 ms, spectral width of 2000 Hz, 2048 data points for each acquisition, and 256 acquisitions for each measurement. The total scan time for each measurement was 8.5 min. The selective inversion pulses were Gaussian pulses with 28.2 ms duration and 55 Hz bandwidth at half-height. A constant VOI size of 5 × 3 × 3 cm3 was used for all in vivo experiments. RF frequency was maintained within 2 Hz using the navigator based frequency lock option provided by the manufacturer. Prior to each GSH scan, shim correction up to second order and chemical shift-selective (CHESS) water suppression were optimized for the selected VOI (18).
A total of forty-three GSH measurements were performed on the eleven normal volunteers (five male, six female, age = 30 ± 11 years). The VOI for each measurement was located in the left or right parietal lobe. For eight of the normal volunteers, two measurements were performed on a VOI in the left parietal lobe and two measurements were performed on another VOI in the right parietal lobe. For one of the normal volunteers, three measurements were performed on one VOI and two measurements were performed on the contralateral side. For two of the normal volunteers, only one VOI was measured three times.
A total of ten measurements were performed on the five ischemic stroke patients (three male, two female, average age = 57 ± 15 years). Each patient received two GSH measurements with one measurement localized in the lesion and the other on the contralateral side (Fig. 1). Two of the patients were scanned one day after stroke onset and their lesions were smaller than 1 cm3 in the diffusion weighted images (DWI). The rest three patients were scanned thirty, five, and six days after stroke onset, respectively. Each of the three patients had a large DWI lesion in the middle cerebral artery (MCA) territory that occupied most of the 5 × 3 × 3 cm3 VOI.
The adaptive spectral registration routine is programmed in IDL (ITT Visual Information Solutions, Boulder, Colorado) and is fully automatic. Time domain raw data for all 256 acquisitions in each measurement are saved for the purpose of spectral alignment. The 2048 data points in each acquisition are zero padded to 8192 points, multiplied by a 4 Hz exponential decay function, and then Fourier transformed into the frequency domain to generate a spectrum for each acquisition. Among all acquisitions, half are “on” acquisitions with the selective inversion pulses placed at the cysteinyl α-proton resonating at 4.56 ppm and the other half are “off” acquisitions with the selective inversion pulses shifted to a higher frequency (+200Hz). Due to the editing effects of the selective inversion pulses, spectra from the “on” and “off” acquisitions are intrinsically different. Aligning the “on” spectra and “off” spectra separately would allow us to fit the spectra over a broader range of data instead of one singlet peak.
Because the procedures for aligning the “on” spectra and “off” spectra are similar, only the procedure for aligning the “on” spectra is detailed here. An important factor affecting the quality of spectral alignment is the quality of the reference spectrum. In order to generate a reference spectrum of good quality, a pair-wise fitting and selection procedure is developed. For the case of 256 acquisitions, the 128 “on” spectra are sequentially grouped into 64 pairs. The two spectra in each pair are aligned with each other by frequency-shifting one of them. The complex-valued spectral data over the range of 1.9 to 3.3 ppm are used in this fitting process. The root mean square (RMS) error between each pair of spectra over the range of 1.9 to 3.3 ppm is recorded. After this frequency alignment, each pair of spectra is averaged into one spectrum, resulting in a total of 64 averaged spectra. The RMS errors of the 64 spectra are sorted into ascending order. Half of the spectra (32 spectra) with highest RMS errors are discarded from this process. The remaining spectra (32 spectra) are grouped into 16 pairs, then aligned and averaged by the same process. This process repeats until there is only one spectrum left, which is selected as the “on” reference spectrum. This reference spectrum is an average of 16 most consistent “on” spectra. Next, the initial 128 individual “on” spectra are aligned with the “on” reference spectrum one by one. Complex-valued data over the same 1.9 to 3.3 ppm range are used in the fitting process. The difference between an individual spectrum and the reference spectrum which is minimized in the fitting process can be expressed as:
where n counts data points in each spectrum; diff(n) is the difference term being minimized; spect(n-Δn) represents the nth data point of an individual “on” spectrum shifted by Δn points; i is the imaginary unit; ϕ0 is the zero-order phase; b0 and b1 represent complex-valued zero and first order baselines, respectively; ref(n) represents the nth data point of the reference spectrum. Four variables Δn, ϕ0, b0, and b1 need to be determined in this non-linear optimization process, for which a Levenberg-Marquardt optimization subroutine is used. The variable Δn is the frequency shift between the individual and reference spectrum in unit of data points. When Δn is not an integer, cubic spline interpolation is used to shift the individual spectrum to fit with the reference spectrum. After aligning each “on” spectrum with the reference spectrum, the RMS errors are sorted into ascending order. Five spectra with largest RMS errors are discarded. The average of the remaining “on” spectra is taken as the averaged “on” spectrum.
The averaged “off” spectrum is generated similarly, and subsequently aligned with the averaged “on” spectrum by frequency, phase, and baseline adjustments based on Eq. . Two segments of data are used in this fitting process, one surrounding the NAA singlet peak (1.81 to 2.21 ppm), and the other in-between the Cr and choline (Cho) peaks (3.08 to 3.17 ppm). The two segments of data are given different weighting factors in the fitting process, with 1 for the first segment and 3 for the second segment. After this spectral alignment, the difference between the averaged “on” and “off” spectra is taken as the difference spectrum. Meanwhile, the average of the averaged “on” and “off” spectra is taken as the averaged spectrum.
At this stage, both of the difference spectrum and averaged spectrum have complex values. Because both spectra have the same phase, zero-order phase for both of them can be obtained by fitting NAA, Cr, and Cho basis spectra to the averaged spectrum. These basis spectra were obtained by scanning 20 mM NAA, Cr, and Cho phantoms separately. The final real-valued difference spectrum is generated by removing the zero-order phase from the complex-valued difference spectrum. In the above fitting process, Cr signal is also obtained.
The GSH signal is determined by fitting GSH and NAA basis spectra (difference spectra) to the real-valued difference spectrum. The GSH basis spectrum was obtained by scanning a 20 mM GSH phantom. The GSH/Cr ratio is then computed as the ratio between the GSH signal and the Cr signal obtained in the previous step.
GSH levels in normal volunteers were measured from the GSH/Cr ratio assuming Cr level was 8 mM based on typical published values (19,20). Standard deviations (SD) for the GSH measurements were also computed. For stroke patients, Cr levels are generally decreased in lesions. Hence, the Cr level in the contralateral normal tissue was used as reference to estimate GSH levels in both lesion and contralateral normal tissue. Furthermore, within subject standard deviation Sw (21) for each VOI was estimated for GSH measurements on normal volunteers, which was made possible by measuring each VOI at least twice. The square of Sw is given by the sum of squares about subject mean divided by the degrees of freedom,
where N is the total number of VOIs; Mi is the number of GSH measurements for the ith VOI; Gij is the GSH concentration value for the jth measurement in the ith VOI; i is the averaged GSH concentration for the ith VOI. Within subject standard deviation Sw excludes the contribution of GSH variations due to different VOI locations and natural differences among different subjects.
For comparison purposes, a singlet-based conventional frequency and phase correction method was also used to process the in vivo data. The NAA singlet peak in the 16th individual spectrum was arbitrarily chosen as the reference for spectral alignment. Each individual spectrum was fitted with the reference spectrum using complex-valued data over the range of 1.81 to 2.21 ppm. Frequency shift and zero order phase of each individual spectrum were determined by the Levenberg-Marquardt optimization subroutine. Five “on” spectra and five “off” spectra with the largest RMS errors were thrown away to make it a fair comparison with the proposed spectral alignment method. The RMS errors were computed over the data range of 1.81 to 2.21 ppm. After spectral alignment, the averaged spectrum was computed as the average of all spectra, and the difference spectrum was computed as the difference between the averaged “on” and “off” spectra. Procedures for obtaining the real-valued difference spectrum and the GSH concentration level were the same as those in the proposed method, which have been described above.
As an estimate of data quality, instabilities in frequency shift, zero order phase, zero order baseline, and first order baseline were computed during the spectral alignment process for each GSH measurement. When an individual “on” or “off” spectrum was aligned with the “on” or “off” reference spectrum, values of four variables were recorded, i.e. frequency shift, zero order phase, zero order baseline, and first order baseline. The standard deviations of the four variables for all spectra in the GSH measurement were defined as the instabilities of the corresponding variables.
Plots of difference spectra for the GSH peak from selected phantom experiments are shown in Fig. 2(a). The maximum GSH peak amplitude is reached at TE = 130 ms. At shorter TE values, the GSH peak amplitudes become smaller. At TE = 70 ms, for example, the GSH peak amplitude is about 70% of that at TE = 130 ms.
Plots of difference spectra for the GSH peak from density operator simulations using GAMMA are shown in Fig. 2(b). The GSH peak shapes are similar to those from phantom experiments but the peak amplitude grows larger when TE becomes longer because relaxation effects were ignored in the density operator simulations. A curve of GSH peak amplitude versus TE obtained from density operator simulations is plotted in Fig. 2(c). GSH peak amplitude reaches maximum value at TE = 145 ms. GSH peak amplitude at TE = 70 ms is only about 50% of the maximum peak amplitude. GSH peak amplitude at TE = 131 ms is over 97% of the maximum peak amplitude.
Density operator simulations of the NAA difference spectra are plotted in Fig. 3. At TE = 131 ms, the NAA spectrum is flat at 2.95 ppm where the GSH peak locates. When the TE value is decreased to 126 ms or increased to 136 ms, the dispersive component of the NAA signal interferes with GSH detection. Therefore TE = 131 ms is a good choice for minimizing NAA signal interference.
Difference spectra for the forty-three measurements on eleven normal volunteers were reconstructed with both the conventional spectral alignment method using the NAA singlet as the navigator and the proposed spectral alignment method. Spectra from the first measurement on each normal volunteer are plotted in Fig. 4. In Fig. 4(a), difference spectra were reconstructed using the conventional method. Substantial subtraction errors at NAA (2.01 ppm), Cr (3.03 ppm), and Cho (3.22 ppm) frequencies are present in several spectra. The GSH peak at 2.95 ppm has relatively large variations in shape. Since GSH at 2.95 ppm overlaps with Cr at 3.03 ppm, relatively large error in GSH quantification due to subtraction errors is expected. GSH concentration determined from the forty-three measurements was 1.13 ± 0.21 mM (mean ± SD) relative to 8 mM Cr. The standard deviation was 19% of the mean. The within subject standard deviation Sw was 0.11 mM (9.7% of the mean). In Fig. 4(b), difference spectra were reconstructed with the proposed method. Subtraction errors are greatly reduced and the GSH peak at 2.95 ppm has more consistent and better defined shapes. GSH concentration was found to be 1.20 ± 0.14 mM. The standard deviation was 12% of the mean. The within subject standard deviation Sw was found to be 0.095 mM (7.9% of the mean). By using the proposed method, the standard deviation dropped from 19% to 12% of the mean. Sw dropped from 9.7% to 7.9%.
Data from the ten GSH measurements on the five stroke patients were processed with both the conventional method and the proposed method. Difference spectra reconstructed with the conventional method (Fig. 5(a)) had very large subtraction errors due to patient motion. GSH levels were found to be 1.16 ± 0.31 mM in lesion and 1.06 ± 0.65 mM in contralateral normal tissue, where the standard deviations were 27% and 61% of the means, respectively. By using the proposed method, subtraction errors were effectively minimized (Fig. 5(b)). GSH levels were found to be 1.19 ± 0.24 mM in lesion and 1.25 ± 0.19 mM in contralateral normal tissue, where the standard deviations were 20% and 15% of the means, respectively.
For the forty-three GSH measurements in normal volunteers, Instabilities of frequency shift, zero order phase, zero order baseline, and first order baseline were found to be 0.98 ± 0.49 Hz, 0.13 ± 0.036 rad, 0.75 ± 1.1, and 0.75 ± 0.067 ppm−1, respectively. The values for the zero order and first order baselines were in absolute units and were proportional to the MR signal amplitude. For the ten GSH measurements in stroke patients, instabilities of frequency shift, zero order phase, zero order baseline, and first order baseline were found to be 1.4 ± 0.47 Hz, 0.19 ± 0.043 rad, 1.2 ± 0.48, and 0.96 ± 0.14 ppm−1, respectively. It can be seen that instabilities of the four variables were all higher in stroke patients, which is expected because subject motion in stroke patients is more severe and occurs more frequently than in normal volunteers.
Density operator simulations show that GSH peak amplitude at TE = 131 ms is approximately twice that at TE = 70 ms, in which the T2 relaxation effects are ignored. For the same amount of signal contaminants from water and macromolecules, the contrast between the GSH signal and background contaminants is improved approximately 100% at TE = 131 ms than at TE = 70 ms when the T2 relaxation effects are ignored. Because GSH is a small molecule metabolite, it is safe to assume that the T2 of GSH in the brain is much longer than those of water and macromolecules. Therefore, the contrast between the GSH signal and background signal contaminants from water and macromolecules will have an even greater improvement (over 100%) for in vivo experiments when the 131 ms TE is chosen over the 70 ms TE. In addition, a longer TE allows the use of longer editing pulses for better frequency selectivity. For a given TE value, variations in pulse sequence timing and RF pulse shapes will change the GSH peak amplitude to a small extent because GSH is a strong coupling system. Our simulations show that these small changes do not change the trend of the GSH peak amplitude versus TE curve shown in Fig. 2(c). Furthermore, as shown in Figs. 2(a), 2(b), and and3,3, dispersive phase abnormalities from both GSH and NAA are absent at TE = 131 ms. This absence of spectral interference allows more accurate spectral quantification of GSH. GSH quantification will be more affected at a long TE than at a short TE if the T2 value of GSH in lesion is different from that in normal tissue. This could be disadvantageous for measuring the true GSH concentration levels but also could be beneficial for detecting a difference between GSH signals in lesion and normal brain tissue.
In an “on” acquisition, the editing pulses are applied at the cysteinyl α-proton at 4.56 ppm which is very close to water frequency. As a result, the residual water signal is greatly affected by the editing pulses. In an “off” acquisition, the water signal is mostly unaffected because the frequency of the editing pulses is shifted away from water frequency. Therefore, there are relatively large baseline differences between the “on” and “off” GSH spectra compared to J-difference spectroscopy experiments measuring GABA. In stroke patients, patient motion during scans is hard to avoid which will cause further baseline differences due to variations in the effectiveness of water suppression. The proposed method has an integrated baseline correction procedure which corrects baseline differences between two spectra in the frequency range of interest rather than the baseline per se. This integrated baseline correction is very effective yet simple and time efficient.
Because spectra from the “on” and “off” acquisitions are intrinsically different, conventional frequency and phase corrections for processing J-difference spectra are based on the alignment of a singlet peak. By processing the “on” and “off” spectra separately, we are able to perform spectral alignment over a broader range of data (22). This makes the proposed method more reliable than the conventional method, especially when the reference peak (e.g. NAA) has poor SNR or shape due to disease, artifacts, or random noise. In addition, the “on” and “off” reference spectra in the proposed method are generated from the data by a pair-wise fitting and selection process, which makes the reference spectra to have better qualities and higher SNR’s than a reference signal created from a single-acquisition spectrum.
By comparing difference spectra reconstructed with the conventional and proposed spectral alignment methods, we have shown that the proposed spectral alignment method is more effective in reducing subtraction errors. In an earlier reported work (7), GSH was measured with TE = 68 ms and spectra were aligned with the conventional method in which the NAA singlet was used as the navigator. Difference spectra from twelve normal volunteers have been presented. Large subtraction errors at around 2.0 ppm are present in seven out of the twelve spectra. The percentage of spectra that have large subtraction errors is similar to that of Fig. 4(a) in this paper where the spectra were aligned with the conventional method. Fig. 4(b) shows substantially reduced subtraction errors over Fig. 4(a) and the twelve in vivo spectra in the earlier work. The GSH concentration value of 1.20 ± 0.14 mM (mean ± SD, n = 43) for the normal volunteers are in good agreement with the published GSH levels (4,7) in normal brain tissue indicating that no bias was introduced by our data processing method. For the five stroke patients, no clinical conclusion can be reached at this stage because they were scanned a different number of days after stroke onset and only three patients had large lesions. Even though patient motion was visually observed for four of the five stroke patients during GSH measurements, all ten difference spectra have reasonably good quality, which demonstrates the robustness of the proposed spectral alignment method. In our experiments, the 5 × 3 × 3 cm3 VOI allows single average comparison. For more typical VOI sizes in brain studies such as 10 cm3 or less, this spectral alignement method may not work due to insufficient SNR.
In summary, GSH measurement by J-difference spectroscopy was performed on a commercial 3 T scanner with a 30 cm birdcage T/R head coil. A relatively long TE of 131 ms was found to be an optimal choice for reducing subtraction errors while maintaining a high GSH yield. The 131 ms TE also minimizes spectral interference from signals originated from the aspartate moiety of NAA. A fully automatic spectral alignment method was developed to reduce subtraction errors by performing frequency, phase, and linear baseline corrections on each individual spectrum. Compared to the conventional spectral alignment method, the proposed method was much more effective in reducing subtraction errors. GSH levels were determined to be 1.20 ± 0.14 mM (mean ± SD, n = 43) in eleven normal volunteers. In five stroke patients, GSH concentrations were found to be 1.19 ± 0.24 mM in lesion and 1.25 ± 0.19 mM in contralateral normal tissue. This work demonstrated the feasibility of reliably measuring GSH levels in stroke patients in the clinical setting, which paved the way for clinically relevant studies on larger stroke populations.
We gratefully acknowledge Mr. Christopher Johnson for making metabolite phantoms, Mr. Kevin Scott for assistance in scanner operations, Dr. Christabel Lee for assistance in scanning normal volunteers, and staff of the NIH stroke team for assistance in recruiting and caring for the stroke patients. This work was supported by the intramural programs of the NIH, NINDS, NIMH, and NIH Clinical Center.
Grant Support: This work was supported by the intramural programs of the NIH, NINDS, NIMH, and NIH Clinical Center.