|Home | About | Journals | Submit | Contact Us | Français|
The utility of multi-voxel two-dimensional chemical shift imaging in the clinical environment will ultimately be determined by the imaging time and the metabolite peaks that can be detected. Different k-space sampling schemes can be characterized by their minimum required imaging time. The use of spiral-based readout gradients effectively reduces this minimum scan time due to simultaneous data acquisition in three k-space dimensions (kx, ky, and kf2). A 3T spiral-based multi-voxel two-dimensional spectroscopic imaging sequence using the PRESS excitation scheme was implemented. Good performance was demonstrated by acquiring preliminary in vivo data for applications including brain glutamate imaging, metabolite T2 quantification, and high spatial resolution prostate spectroscopic imaging. All protocols were designed to acquire the data within a 17 minute scan time at a field strength of 3T.
In vivo proton magnetic resonance spectroscopy (MRS) provides valuable information regarding biochemical processes. The spectroscopic information content can be enriched in several different ways. Short echo time spectroscopy (echo time ≤ 35 ms) has been extensively used to enable the robust quantification of metabolites with both long and short T2 relaxation times. Magnetic resonance spectroscopic imaging (MRSI) can provide additional topological knowledge. In addition, two-dimensional (2D) spectroscopy can be used to better detect J-coupled metabolites [1-3]. While short echo MRS can be implemented without sacrificing scan time, multidimensional techniques typically require more scan time to gather the desired information. This additional data acquisition can be formulated as extended k-space coverage. In the case of MRSI, the coverage is represented as kx, ky, and kz, while 2D spectroscopy requires sampling in both the kf1 (= t1: indirect evolution time, Fourier domain of f1) and kf2 (=t2: direct detection time, Fourier domain of f2) dimensions.
Approaches that combine MRSI with 2D spectroscopy have recently been introduced [4-8]. These multi-voxel 2D spectroscopic imaging techniques can gather spatial distributions of metabolites with complex coupled spectra. Despite the prospect of many new applications, the necessary, overwhelming increase in scan time hampers the clinical use of these techniques. The requirements to acquire data in 3D spatial and 2D spectral dimensions can be expressed in five k-space axes, namely, kf1, kf2, kx, ky, and kz. Conventional phase-encoded MRSI methods are too inefficient for most applications, although in some cases, longer scan times are inevitable due to the signal to noise ratio (SNR) requirements . An alternative method incorporates fast imaging features, thereby reducing the overall minimum scan time [9-13].
We investigated fast imaging techniques combined with 2D spectroscopic imaging. Using a k-space sampling efficiency criterion, we found that oscillating readout gradients, especially those generating a spiral k-space trajectory, had relatively short minimum scan time requirements. This feature makes it possible to perform clinical in vivo multi-voxel 2D spectroscopic imaging within acceptable scan times. Expanding on our previous results [9, 11], we applied the technique to various new applications including volumetric TE-averaged MRSI of brain glutamate levels, volumetric multi-TE MRSI for T2 quantification, and high spatial resolution multi-voxel 2D-J resolved MRSI to detect citrate and polyamines in the prostate at 3T.
The k-space sampling efficiency is determined by the minimum number of repetition time (TR) steps required for a scan protocol. Given an imaging matrix size of N × N, the minimum number of TR steps for conventional phase-encoded MRSI and echo-planar MRSI (EPSI) will be N2 and N, respectively. For 2D spectroscopy requiring multiple TE acquisitions, the total scan time is linearly proportional to the number of TE steps. Spectroscopic U-FLARE approaches have total scan time that is largely independent of the matrix size N, but dependent on the spectral bandwidth and resolution in f2 .
To evaluate the minimum number of TR steps for the spiral protocol, we used a trajectory design method presented by Glover  and simulated the required length of the spirals for various values of N. Simulations were performed assuming a 1000 Hz spectral bandwidth in f2. For spirals, the number of TR steps is governed by the number of spatial interleaving steps required. Spatial interleaving is needed to reduce the length of each spiral to fulfill the spectral bandwidth requirements. Each spiral during a single readout (TR period) is connected via rewinder gradients that are designed to relocate the k-space trajectory back to its origin (i.e., kxy = 0). These additional rewinder gradients also cause a reduction in spectral bandwidth. To reduce the length of these rewinders as much as possible, we used a gradient design algorithm that operates at the gradient amplifier slew-rate limit. A second order polynomial fit of the minimum number of TR steps required was calculated for the given values of N.
Spiral readout gradients with 2D spectroscopy features were added to a PRESS sequence as shown in Fig. 1. Three different protocols were used to validate the sequence for specific applications. All scans were conducted on a 3T EXCITE scanner (General Electric Healthcare Technologies, Milwaukee, WI).
The TE-averaging approach was used with nine different TEs chosen to maximize Glu C4 proton peak detection at 2.35 ppm [16, 17]. Single voxel spectroscopy was performed using a phantom containing Glu at various echo times with high SNR to estimate the Glu C4 amplitude. Echo times in regions with relatively big Glu C4 amplitudes were densely selected. As a result, the TE steps chosen had a sampling density that was proportional to the amplitude of the Glu C4 proton peak. Specifically, the TEs chosen for this study were 40-50-65-85-110-140-170-200-220 ms. The spectroscopic imaging parameters were as follows: 16×16×8 spatial coverage over a 16×16×12 cm FOV (nominal spatial resolution of 1.5cc), 1.7 second TR, 2 signal averaging, CHESS water suppression, approximately 1000 Hz spectral bandwidth in f2 acquired with 512 points, and 16:30 minute scan time using an 8 channel phased array head receiver coil. Phase encoding in z was performed to obtain resolution in the through-plane direction. The spatial selection of the PRESS box was a rectangular region within the parenchyma of the brain typically of the size ~ 12×12×8 = 1152cc. Spatial saturation pulses were applied to eliminate lipid contamination and aliasing due to the relatively small FOV. An Eddy current correction routine as outlined in Ref.  was used to reduce the effects of gradient imperfections for this particular spiral trajectory. For experiments 2) and 3), no correction routine was used since the spiral trajectories did not cause detectable amounts of error. Raw data collected from the different echo times were first added to obtain TE-averaged raw data. The data were then reconstructed by applying gridding in kx and ky. Following gridding, a 4D Fast Fourier Transform (FFT) was performed to convert data from the kx, ky, kz, and kf2 (=t2) domain to the spatial x,y,z and spectral f2 domain, respectively. Apodization using a 4 Hz Lorenztian filter was applied prior to the FFT in the t2 domain. A repeatability study was also performed on a volunteer to evaluate the potential for usage in clinical settings.
For this application, the imaging parameters were slightly modified from those given above: 8×8×8 spatial coverage over a 16×16×8 cm FOV (nominal spatial resolution of 4 cc), 2 second TR, 1 signal averaging, CHESS water suppression, approximately 1000 Hz spectral bandwidth in f2 acquired with 512 points, TE from 40ms to 190ms with 10ms intervals for 16 different TE step acquisitions, and 17 minute scan time. A spiral out-in trajectory was used for the readout in this case, which reduces phase errors due to gradient moment buildup during the readout . Phase encoding in z was used for through-plane coverage. The spatial selection of the PRESS box was normally chosen to be approximately 12×12×6 = 864cc and the positioning of the spatial saturations pulses were similar to experiment 1). This study was also performed using an 8 channel phased array coil. For the studies using multi-channel receivers, data were processed for individual coils separately and then combined. Prior to combination, each voxel from each coil was multiplied by a weighting factor, proportional to the amplitude of the phased water spectra. For experiments 1) and 2), higher order shimming over the selected region was performed to improve field homogeneity . For this experiment, data were reconstructed separately for each TE acquisition much like the procedure outlined in experiment 1). From each reconstructed echo time data, the areas of NAA, Cr, and Cho were calculated and curve fitting of the area values was performed using a single exponential decay model to estimate the T2 values for every voxel.
Spiral trajectories are characterized as having a scan time that is largely independent of the imaging matrix and field of view. The protocol was modified for obtaining fine spatial resolution without an increase in the minimum scan time or decrease in field of view . The modified imaging parameters for single slice high spatial resolution 2D-J resolved MRSI were: 32×32 in-plane matrix size over a 16 cm FOV with 1cm slice thickness (nominal spatial resolution of 0.25 cc), 16 echo time steps starting from 35ms to 285ms with 15.6ms intervals (bandwidth 64 Hz and resolution 4 Hz in f1), approximately 1000 Hz spectral bandwidth in f2 acquired with 256 points, 2 signal averaging, 2 sec TR, and 17 minute scan time. An endorectal coil was used for signal reception. Following gridding in kx and ky, a 5D FFT was performed to convert data from the kx, ky, kz, kf1 (=t1), and kf2 (=t2) domains to the spatial x,y,z and spectral f1 and f2 domains, respectively. Data from the 5D reconstructed data sets were visualized for particular f1 values. In particular, spectra corresponding to the f1=0 and -8 Hz lines (J(0) Hz and J(-8) Hz lines) were extracted and phased manually to investigate the presence of the metabolites of interest.
Figure 2 shows the minimum required TR steps for different MRSI sequence protocols assuming a single TE acquisition. Note that further reductions of this value could be achieved by using other approaches such as parallel imaging schemes or reduced k-space sampling trajectories [21-23]. The resulting polynomial fit indicated that the number of TR steps required for a given imaging matrix size of N×N for spiral acquisitions was 0.012×N2+1. Compared to EPSI acquisitions, spirals require approximately half the scan time for the given matrix sizes.
A volumetric TE-averaged MRSI example is shown in Fig. 3 from a patient with suspected amyotrophic lateral sclerosis. Representative spectra from the 16×16×8 TE-averaged data set are shown covering the corticospinal tract. The largest Glu concentrations were observed in the gray matter regions (indicated by ‘v’). A repeatability study performed on a normal volunteer resulted in a coefficient of variation (CV) for NAA ≈ 7%, Cr ≈ 10.8%, and Cho ≈ 10.6%. For Glu, the CV was approximately 17% in regions where Glu values were clearly visible such as that shown in the left column of Fig. 3. These regions were mostly located in the gray matter region as indicated in Fig. 3. Note that reduced B0 homogeneity resulted in broadened linewidths in the inferior and anterior voxels.
Figure 4 shows an example of metabolite T2 quantification via spiral 2D MRSI from a normal volunteer. From the processed data sets, the voxels containing the most gray and white matter areas were selected and are shown with a) corresponding TE spectra, b) summed (TE-averaged) spectra, and c) T2 fitting example. The T2 estimates in the gray and white matter were NAA = 199 and 278 ms, Cr = 144 and 140 ms, and Cho = 299 and 397 ms, respectively .
Finally, Fig. 5 displays an example of high-resolution 2D-J resolved prostate MRSI from a subject with suspicion of prostate cancer. It has previously been shown that the addition of a J-resolved dimension allows observation of the J-modulation of citrate, as well as the resolution of polyamines from overlapping Cho and Cr signals . These J-modulated signals can be clearly seen in this example from the J(-8) Hz line spectra, which shows the polyamine and doublet citrate metabolite. In addition, the spatial distribution of these signal components is well resolved at a high spatial resolution of 0.25 cc. It is worth noting that applying the same imaging parameters would have necessitated a scan time of 1092 min and 34 min for conventional phase-encoded and echo-planar MRSI, respectively.
We implemented and demonstrated 3T multi-voxel 2D MRSI sequences using spiral readout gradients for several applications. The application of spiral k-space trajectories, which use two simultaneously oscillating readout gradients, reduced the minimum number of TR steps needed as compared to other acquisition methods. This method can be effectively added to 2D spectroscopic acquisitions while maintaining high spectral quality and reasonable scan times. Our methods were all devised to be executed within a reasonable scan time of approximately 17 minutes with TR values of 1.7 or 2 seconds. Given this protocol (matrix size N=16), the imaging time required for conventional phase-encoded MRSI and EPSI would be found to be ×64 and ×4, respectively, from that of the spiral MRSI as shown in Fig. 2. Previous studies of 2D MRSI using spiral readouts were limited in spatial coverage or spatial resolution [9, 11]. We have extended these studies by gathering volumetric 3D data or high spatial resolution single-slice data sets. Imaging on a 3T scanner using multiple receiver coils facilitated the utility of this technique by providing improved SNR.
The ability to simultaneously perform both 2D spectroscopy and spectroscopic imaging within a reasonable scan time has great potential for clinical applications. For TE-averaged volumetric MRSI, we demonstrated the measurement of the spatial distribution of brain Glu. In this implementation, we have used a variable density filter function approach in which the TE steps were optimized for Glu C4 detection. Different filter functions could also be used to enhance detection of other metabolites such as GABA or myo-Inositol . Although the CVs of NAA, Cr, and Cho were within reasonable values, the CV for Glu was relatively high owing to the low SNR achieved for this metabolite, which is dominated by the voxel size prescribed. Another reason for the high CV of Glu is the low spatial resolution of the multi-voxel acquisition, which leads to ringing effects. The remaining lipids after imperfect suppression became a large source of contamination due to ringing. Lipids themselves are coupled resonances leading to a strong echo time dependency that can overwhelm Glu behavior. Improved lipid suppression pulses need to be developed for volumetric coverage applications.
The ability to quantify metabolite T2 from various regions of the brain simultaneously in a single scan can reduce systematic problems in multiple studies while greatly increasing the scan efficiency. Volumetrically resolved T2 quantification can be very useful in providing not only new information regarding local magnetic environments but also improved absolute metabolite quantification. While most quantification approaches rely on reported T1 and T2 values, applications such as pediatrics where dramatic changes in the parenchymal environment occur during the early stages of life, could incur substantial errors when using assumed, fixed relaxation values . The capability to measure T2 spatial variations directly will certainly help in the quantification process. It is worth noting that T2 information can be obtained automatically in both TE-averaged and J-resolved 2D spectroscopy since raw data is stored separately for each TE acquisition.
One limitation of our procedure for multi-voxel T2 measurements as illustrated in the example shown in Fig. 4 is the relatively big voxel size, which leads to inaccurate quantification due to signal contamination from adjacent voxels. Although we selected the voxels with the purest gray and white matter components, contributions from outside the target voxels are unavoidable. Another difficulty is the long process time needed to calculate the area of each metabolite and estimation of the T2 values for every voxel, thereby limiting the number of case studies. At the time of this study, we did not have an automated procedure for this routine. We are currently incorporating a multi-voxel multi-echo LC-model fitting routine  to automate the whole reconstruction and fitting procedure. In this example, the scatter of the Cho peak should be noted. Although we are not sure of the source of this scatter, a potential explanation is that the peak at 3.2 ppm reflects the contribution of a multitude of metabolites, including mI and taurine, in addition to Cho . The amount of scatter would then be dependent on the selected echo times due to the strong coupling effects of both mI and taurine.
The utility of multi-voxel 2D MRSI is not limited to the functions presented here. Indeed, given the flexible scan time management of the spirals, other applications are also possible. In the examples presented, we have generally used a small number of TE steps (16 or below), which undersamples the f1 spectra and results in a small spectral bandwidth. Others have shown that oversampled J-resolved acquisitions can be very useful for reducing sidelobes arising from unsuppressed water or lipid resonances [29, 30]. Our approach could be extended to perform oversampling of the f1 domain. In this case however, decreased spatial resolution may be unavoidable to reduce the overall scan time to within clinically acceptable limits.
In practice, several major obstacles can arise during actual implementation of multi-voxel 2D MRSI. High gradient performance is particularly important. The use of oscillating gradients in EPSI or spiral base MRSI can be demanding on the gradient coils. For experiment 1), we used an additional eddy current correction routine in addition to the pre-emphasis routine provided by the scanner itself to further mitigate these effects. For the other experiments, our eddy current correction routine did not help, as explained in Ref. . The collection of multi-TE data directly impacts the amount of acquired data with typical raw data file sizes often exceeding 1 gigabyte, which can be demanding on the scanner acquisition and reconstruction hardware. In fact, this was the reason why we could not gather additional echo time data for experiment 1), where we opted to use 2 signal averages instead due to raw data size limit.
Finally, there are general differences in the performance of spiral-based MRSI, EPSI MRSI, and conventional phase encoded MRSI. As mentioned, EPSI and spiral trajectories demand high performance gradients. Depending on gradient fidelity and specific spiral trajectories, we have seen spectral differences of up to ~10% in previous studies . For EPSI, gradient coil dependence can be somewhat alleviated by using fly-back echo-planar trajectories . The relative sensitivity to lipids or inhomogeneity is governed by the impulse response of these trajectories. While MRSI itself will resolve for any inhomogeneities, the resulting end image will be different for these trajectories, which depend on the spatial point spread function . In relation to motion sensitivity, spirals are considered advantageous due to their sampling pattern which starts at the k-space origin. More comprehensive studies comparing these schemes have previously been published showing small SNR differences between EPSI and phase-encoded methods . We have recently added the spiral trajectory to the aforementioned study and results indicate that the spiral trajectory perform similar to EPSI trajectories (data not shown).
The utility of a variety of multi-voxel 2D MRSI sequences has been investigated. The use of spiral readout gradients leads to a small number of TR steps, enabling clinically feasible scan times. Applications of this technique include brain Glu mapping, metabolite T2 mapping, and improved prostate MRSI.
Grant sponsors: NIH CA48269, RR09784.
Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final citable form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.