High-field downfield magnetic resonance spectroscopic imaging in the human brain

Abstract

Purpose:

To investigate the feasibility of downfield magnetic resonance spectroscopic imaging (DF-MRSI) in the human brain at 7T.

Methods:

A 7T DF-MRSI pulse sequence was implemented based on the previously described methodology at 3T, with 3D phase-encoding, $1\stackrel{-}{3}3\stackrel{-}{1}$ spectral-spatial excitation, and frequency selective refocusing. Data were pre-processed followed by analysis using the ‘LCModel’ software package, and metabolite maps created from the LCModel results. Total scan time, including brain MRI and a water-reference MRSI, was approximately 24 minutes. The sequence was tested in 10 normal volunteers. Estimated metabolite levels and uncertainty values (Cramer Rao Lower Bounds, CRLBs) for nine downfield peaks were compared between seven different brain regions, anterior cingulate cortex (ACC), centrum semiovale (CSO),corpus callosum(CC), cerebellar vermis (CV), dorsolateral prefrontal cortex(DLPFC), posterior cingulate cortex(PCC), and Thalamus(Thal).

Results:

DF peaks were relatively uniformly distributed throughout the brain, with only a small number of peaks showing any significant regional variations. Most DF peaks had average CRLB<25% in most brain regions. Average signal-to-noise ratio (SNR) values were higher for the brain regions ACC and DLPFC (~7±0.95, mean±sd) while in a range of 3.4 – 6.0 for other brain regions. Average linewidth (FWHM) values were greater than 35Hz in the ACC, CV, and Thal, and 22Hz in CC, CSO, DLPFC, and PCC.

Conclusion:

High-field DF-MRSI is able to spatially map exchangeable protons in the human brain at high resolution and with near whole-brain coverage in acceptable scan times, and in the future may be used to study metabolism of brain tumors or other neuropathological disorders.

Introduction

Despite the fact that several biomedically interesting compounds resonate downfield (DF) from the water resonance(1), the majority of proton magnetic resonance spectroscopy (MRS) studies of the human brain over the last three decades have focused on the upfield (UF) region. There are several reasons for this, but one of the major ones is that many of the DF functional groups are exchangeable with water (e.g. amides, hydroxyls, etc.). Since standard UF MRS methods nearly always employ a pre-saturation water suppression module before MRS data acquisition, exchangeable DF resonances will also be saturated, resulting in very low signal intensities(2,3). Another challenge is that DF resonances often also have short T₂ relaxation times, resulting in broad linewidths, overlapping peaks, and signal loss at longer echo times (TE)(2,4–8).

Over the last decade, methods have been developed to measure and characterize DF resonances in human brain using B₀ field strengths at 3, 7, and 9.4 Tesla (T)(2,3,9–21). An important principle is to avoid pre-saturation of water in order to maintain equilibrium magnetization of exchangeable DF peaks. An early study at 3T measured the DF T₁ relaxation times, that are important to know in order to optimize acquisition parameters such as TR and flip angle (2). Studies at 7 and 9.4 T reported exchange rates and T₂ relaxation times of DF resonances which provide knowledge about choosing optimal echo times(9–11). Higher magnetic field strengths are known to result in increased chemical shift dispersion (measured in Hz) and signal-to-noise ratios (SNR). Also, for spin systems in slow or intermediate exchange conditions, increasing the chemical shift dispersion between the exchanging peaks should improve peak visibility (22). However, increasing B₀ also has some disadvantages, such as increased linewidths (measured in Hz), potentially unfavorable changes in both T₁ and T₂, increased B₀ and B₁ field inhomogeneity, and other factors such as less efficient RF coil designs and increased RF power deposition (specific absorption rate, SAR)(22). Also, the chemical shift displacement error (CSDE) increases with higher field strength(22). Despite these potential disadvantages, multiple single voxel (SV) UF MRS studies have demonstrated improved SNR and lower uncertainty estimates (Cramer Rao Lower Bounds, CRLBs) at 7T compared to 3 or 4T(22–24). SV UF measurements at 7T have also been demonstrated to be very reproducible (25). However, there have been few studies directly comparing DF MRS measurements between different field strengths(26).

Recently, the spatial distribution of DF resonances in the human brain was mapped using a 2D DF MR spectroscopic imaging (MRSI) technique optimized for exchangeable protons(27), and further expanded to whole brain coverage using 3D phase encoding at 3T(28,29). The pulse sequence consists of a composite binomial pulse (90°) for spectral spatial excitation, and a 180° frequency refocusing pulse (Figure 1B) is designed to avoid any perturbation of the water resonance which could cause saturation of DF resonances via chemical exchange.

Figure 1.

(A) Sagittal brain MRI showing slab location and phase-encoding matrix used for 3D DF-MRSI. (B) Pulse sequence for 3D DF-MRSI implemented for 7T. The sequence consists of a $1\stackrel{-}{3}3\stackrel{-}{1}$ spectral-spatial excitation pulse, phase-encoding gradients in 3D dimensions, and a frequency selective 180° pulse for refocusing of DF signals only. Crusher gradients are applied on all 3 gradient directions to remove artifacts, including residual water signals. Note that there is no saturation of the water signal throughout the sequence to avoid attenuation of DF resonances via chemical exchange with water.

A primary motivation for the development of DF-MRSI measurements of exchangeable protons is to be able to map brain amide levels, similar to that which may be estimated using amide proton transfer - chemical exchange saturation transfer (APT-CEST) MRI techniques(30). MRSI may offer complimentary information on amide concentrations, since it is expected to be more sensitive to slower exchanging protons than CEST (which is most sensitive for intermediate exchange rates), and also potentially less susceptible to known CEST artifacts such as non-specific magnetization transfer (MT) effects, or contrast due to changes in brain water relaxation times (31). APT-CEST has been shown in multiple studies to provide information on brain tumors, stroke, and other neuropathologies, so DF-MRSI may also be useful for diagnostic or neuroscience studies of these conditions (32–36).

In addition to providing maps of exchangeable amide protons, there is also a potential other information available from DF MRS. For instance, depending on both the acquisition techniques used and study populations, a variety of other compounds have been detected in the DF region of the spectrum; these includes peaks from homocarnosine (an inhibitory neuromodulator consisting of GABA and L-histidine)(37), L-histidine itself (detected following oral loading)(15), phenylalanine (in patients with phenylketonuria)(14), and other compounds such as glutathione and nicotinamide adenine dinucleotide (NAD+)(6,8). The DF resonances of both homocarnosine and histidine are pH sensitive, so potentially can also be used to estimate brain pH. Finally, a recent study(20) using large voxel sizes also suggested that DF MRS may be able to detect the amino acid L-tryptophan, a precursor of both the neurotransmitter serotonin and the neurohormone melatonin.

The purpose of the current study was to develop methods to map the distribution of DF signals in the human brain at 7T with whole brain coverage. Regional 7T DF-MRSI measurements were made in seven brain regions in ten healthy volunteers, including estimated DF peak concentrations, uncertainty estimates (Cramer Rao, Lower Bounds, CRLB), signal-to-noise ratios (SNR) and linewidth values. Metabolite concentrations were examined for between region differences. In addition, metabolite maps were reconstructed to visualize the spatial distribution of DF metabolites.

Methods

All participants provided written informed consent approved by the Johns Hopkins Medicine Institutional Review Board (JHMIRB). All scans were performed on a Philips 7T ‘Achieva’ scanner equipped with an 8-channel transmit/32- channel receive RF head coil (Nova Medical), spherical harmonic shim coils up to 3^rd-order and a 16-channel local shim array (MRShim, GmbH, Tübingen, Germany) for optimization of B₀ field homogeneity. 3D DF-MRSI was implemented as described previously at 3T (27,28), with $1\stackrel{-}{3}3\stackrel{-}{1}$ spectral-spatial excitation and frequency selective refocusing. Briefly, as shown in Figure 1B, the $1\stackrel{-}{3}3\stackrel{-}{1}$ spectral-spatial pulse is a composite of four individual slice selective sinc pulses with flip angles in a binomial distribution; i.e. the “1” element corresponding to an 6.875° flip angle if the overall binomial pulse corresponds to a 55° rotation (38,39). The delays between the slice selective pulses makes the binomial pulse frequency selective, exciting off-resonance magnetization while avoid perturbation of on-resonance (i.e. water) magnetization (i.e. water magnetization is returned to the Z-axis at the end of the binomial pulse (34)). Frequency selective excitation and refocusing allows for the detection of downfield signals without the presence of large residual water signals, and since the water magnetization is not saturated, there is no signal loss due to chemical exchange between water and exchangeable protons; in fact, the DF signals benefit from the relaxation enhancement (‘RE’) effect whereby their apparent T₁ relaxation time is shortened by exchange with fully relaxed water (4,16,17). DF signals are mapped in 3 dimensions through the use of 3D phase-encoding. The frequency selective refocusing pulse only refocuses signals downfield from water. Crusher gradients (20 mT/m, 2 ms duration) are applied in all three spatial directions to eliminate any unwanted signals from the refocusing pulse.

Spatial resolution (i.e. field of view, voxel size, acquisition matrix size) and TR were the same as in the previous implementation at 3T (28). Because of the greater chemical shift dispersion at 7T, the $1\stackrel{-}{3}3\stackrel{-}{1}$ inter-pulse delay $(\mathrm{\delta },\mathrm{m}\mathrm{s})$ was shortened from 1.45 to 0.62 msec (i.e. maximum excitation at 7.4 ppm). The frequency offset $(\mathrm{\Delta },\mathrm{H}\mathrm{z})$ is related to δ according to:

$$\mathrm{\delta }=1000/(2^{\mathrm{*}}\mathrm{\Delta })$$

Further, the bandwidth of the selective refocusing pulse increased from 400 to 900 Hz (~3.0 ppm), which allows TE to be shortened to 15ms compared to 22ms at 3T. Because of the shorter inter-pulse delay of the $1\stackrel{-}{3}3\stackrel{-}{1}$ spectral-spatial excitation and slightly lower transmit B₁ on the 7T system (18 μT, vs. 22 μT at 3T), the number of lobes of the slice-select pulses was decreased from 5 to 3 in order to achieve similar flip angles on both systems. The pulse width of each individual component of the $1\stackrel{-}{3}3\stackrel{-}{1}$ spectral-spatial pulse was 0.314 ms, and their bandwidth was 13.3 kHz. A 120 mm excitation slab thickness covering from the cerebellum to the vertex was planned for each subject with a FOV of 200×180×120 mm and a matrix size 29×26×8, giving a nominal spatial resolution of ≈7×7×15 mm = 0.7 cm³. A circular region of k-space was covered using conventional, cartesian sampling (40). Circular sampling saves about 25% scan time compared to a square and also gives a spherically symmetric point-spread function (41). Other scan parameters were 1 transient, flip angle 55 °, TR 282 ms. A SENSE acceleration factor of 1.5 was used in both left-right and anterior-posterior directions resulting in a scan time of 10m 37s. All shim currents (both LC coil, and scanner spherical harmonics up to 3^rd order) were calculated based on field maps over the entire volume of the brain using the ‘Arche’ software provided by MRShim. In addition to 3D DF-MRSI, an unsuppressed water MRSI was also recorded with the same voxel size and geometry as the DF sequence using a FID-MRSI sequence with flip angle 30°, TR 208 ms, TE 1 ms with a scan time of 7m 52s. Proton density (PD) localizer images, and a 3D T₁w Magnetization Prepared - RApid Gradient Echo (MPRAGE) scans (42) were also performed. Total scan time was approximately 24 minutes. Finally, an imaging sequence implementing the DREAM(43,44) method for B1 mapping was performed with a nominal flip angle of 7°, a TE of 1.40 ms and TR of 3.7 ms. The protocol was tested in ten normal volunteers (5F, age 42±10 yrs).

All spectral data were exported from the scanner after channel combination and SENSE-MRSI reconstruction performed by the scanner software. Time-domain MRSI data were then processed in an automated pipeline using custom MATLAB (Mathworks, MA) software. In a pre-processing step, 4 Hz exponential filtering was applied, and DF-MRSI data were frequency corrected on a voxel-by-voxel basis using the water frequency determined from the H₂0-MRSI scan. Residual water was then removed using an HLSVD filter (45) in a range of 4.1 to 5.2 ppm (i.e. a frequency range of ~328 Hz) (46). DF-MRSI data were analyzed using ‘LCModel’ (47) fitting with a basis set consisting of nine Gaussian peaks ranging in frequency from 6.83 to 8.49 ppm as described previously (27,28). Water scaling was used for the LCModel concentration estimates. Due to the lack of the knowledge about the number of protons per function group, exchange rates, T₁ or T₂ relaxation times, no attempt was made to correct for sequence timings (TR, TE) or estimate millimolar concentrations, and results are therefore presented in ‘institutional units’ (i.u.). In addition to reporting individual values, the 3 peaks in the amide region of the spectra (8.18 to 8.37 ppm) were summed together (‘8.x’). SNR and CRLB maps were reconstructed from the LCModel output. DF maps were created using the amplitude estimates from the water-scaled LCModel output for each voxel, except when a CRLB of ‘999’ was reported, in which case the amplitude was set to zero. The LCModel control file is provided in the supplemental information.

For the reconstruction of the metabolite maps, proton density (PD) images were segmented using SPM12 (48). Metabolite maps were only computed for voxels included in the brain mask. As suggested in a recent consensus paper (49), visual interpretation of metabolite images was improved by using image interpolation factor of 8.

In each subject, up to four voxels were selected from seven representative brain regions at different locations throughout the brain: anterior cingulate cortex (ACC), dorsolateral prefrontal cortex (DLPFC), and posterior cingulate cortex (PCC) in gray matter rich regions as well as corpus collosum (CC) and centrum semiovale (CSO) in white matter, and finally cerebellar vermis (CV) and thalamus (Thal) in inferior regions. The ‘R’ Statistical Analysis Package version 3.5.1 was used for regional brain analysis (50). After log transformation, the ‘R’ generalized linear model (‘glm’) function was used to compare brain metabolite levels (i.u.) between ACC, CSO, CC, CV, DLPFC, PCC, and Thal. The independent variable is brain regions, and the dependent variable is level of metabolites. The Benjamini-Hochberg (BH) procedure (51) was used for multiple comparison correction. P values corrected with the BH procedure were presented as p-adjusted and considered significant if less than 0.05.

Results

Figure 1A shows the slab location chosen for 3D DF-MRSI in one subject, while Figure 2 shows voxel locations and representative spectra from the 7 brain regions chosen for quantitative analysis.

Figure 2.

Proton density localizer images and voxel locations of the 7 brain regions selected for quantitative analysis: cerebellar vermis (CV), left thalamus (Thal), anterior cingulate cortex (ACC), body of the corpus callosum (CC), posterior cingulate cortex (PCC), right dorsolateral prefrontal cortex (DLPFC), and left centrum semiovale (CSO) white matter. Spectra from each of these regions are plotted for 1 subject. The region analyzed by LCModel (see figure 3) is highlighted between approximately 6.5 to 9.0 ppm. The residual water signal (after removal by the HLSVD filter) can be seen at 4.7 ppm.

Figure 3 shows results of LCModel fitting for the brain regions and spectra shown in Figure 2. Figure 4 shows PD localizer images, 8.x ppm amide maps (8.1–8.3 ppm) from five slices in one subject. Spectra from 5 voxels in each slice are shown to illustrate near uniform spectral quality over this range of brain regions.

Figure 3.

LCModel fits of the 7 brain regions (figure 2) selected for quantitative analysis using nine Gaussian peaks and a baseline constrained to be relatively flat.

Figure 4.

Example amide (8.x ppm, 8.1–8.3 ppm) concentration maps and corresponding PD localizer images from five slices in one subject, with five spectra plotted from different anatomical regions from each slice. Note that this volunteer has a benign right parietal epidermal inclusion cyst (slice 7).

Figure 5 shows proton density images, amide maps and corresponding CRLB maps, SNR, LW, B₀ and B₁ maps over all 8 slices from one subject. Amide maps from all 10 subjects are provided in Figure S1.

Figure 5.

Proton density (PD) and amide (8.x) images, and maps of amide CRLB (range 0–20%), SNR (from LCModel, range 0–10), linewidth (LW, from water scan, range 0 to 60 Hz), B₀ (range −50 to + 50 Hz) and transmit B₁ maps (0 to 120% of nominal B₁) from all 8 slices in one subject. More caudal slice locations are characterized by worse B₀ field homogeneity and lower transmit B₁, which results in lower SNR and increased linewidths and CRLB values.

Figure 6 shows estimated concentrations maps reconstructed for all nine DF peaks at 8.49, 8.37, 8.24, 8.18, 7.48, 7.30, 7.09, and 6.83 ppm in one subject.

Figure 6.

Estimated concentrations maps reconstructed for all nine DF peaks at 8.49, 8.37, 8.24, 8.18, 7.48, 7.30, 7.09, and 6.83 ppm in one subject. Note that different brightness/contrast settings are applied for each DF peak.

Figure 7 shows descriptive statistics for concentration estimates from LCModel as metabolite levels in “institutional units” and CRLB values as percentage for the nine downfield peaks and combined amide resonances (8.18–8.37 ppm) in all ten subjects from selected voxels in seven different brain regions. SNR and FWHM values from LCModel are also given for all 7 brain regions. There were no statistically significant differences between brain regions for any peak, except for DF6.83 which was significantly lower in the ACC, DLPFC, PCC, CV and CSO compared to Thal, and significantly lower in ACC than CSO (p<0.05) as well as DF8.18 was significantly different between ACC and CC, CC and DLPFC, and CC and Thal. Table S1 gives the results of statistical tests between brain regions using amide (8.x ppm) estimations from LCModel. The uncertainty estimates from LCModel (CRLB) were less than 5% for amides (8.x), DF8.18, DF7.90, and 6.83 in all brain regions except CC and CV that were less than 10% while all other CRLB values were mostly less than 25% in all brain regions except CV. Average SNR and FWHM values are also depicted in Figure 7. Average SNR, values were higher for the ACC and DLPFC regions (~7.0±0.9, mean ± s.d.) while the average SNR range was 3.4 – 6.0 with a standard deviation range of 0.97 – 1.27 for all other brain regions. Average linewidth (FWHM) values were greater than 35Hz in the ACC, CV, and Thal, and 22Hz in CC, CSO, DLPFC, and PCC.

Figure 7.

Metabolite levels (‘institutional units’) and % CRLB values for the nine downfield peaks and combined amide resonances (8.18–8.37 ppm) in all ten subjects for the seven different brain regions selected for quantitative analysis. Also shown are SNR and linewidth values for all 10 subjects as reported by LCModel.

Head specific absorption rate (SAR) and peripheral nerve stimulation (PNS) levels were <48% and 59% for the 7T DF-MRSI sequence respectively. ‘MRSinMRS’ consensus parameters are also given in the supplemental material.

Discussion

This study demonstrates that 3D DF-MRSI with 0.7 cm³ nominal voxel resolution is feasible at 7T. DF spectra have similar appearance to those previously published at 3T, however metabolite maps are somewhat less homogeneous mainly due to difficulties in achieving uniform B₀ and B₁ field homogeneity over the volume of the brain at high fields. Also the slab excitation profile is expected to be less rectangular than in the previous study at 3T(28) due to the reduced number of sinc lobes used in the current implementation of the $1\stackrel{-}{3}3\stackrel{-}{1}$ excitation pulses. Regional brain analysis showed similar or lower CRLB values compared to the previous study at 3T for all DF metabolites. However, in future work, a systematic comparison between 3T and 7T should be performed with matched methodology and the same volunteers at both field strengths.

In some of the spectra, especially close to the air cavities, but also sporadically in the middle slices, we observed failed LCModel fitting with a CRLB=999 which resulted as ‘black holes’ in the final metabolite maps. As it can be seen in Figure 6, DF maps DF6.83, DF7.48, DF7.90 and DF8.18 have more black holes in lower slices (#1-#4) than the higher slices (#5-#8) as expected to due to worse field homogeneity and lower transmit B₁ levels in the inferior brain regions.

Water scaling was used and maps were created relative to the brain water signal, but there are many correction factors that would need to be applied before the LCModel output could be regarded as even approximately mM concentration units; these would include correcting for T₁ and T₂ relaxation times, taking into account different sequence timings used for water and metabolites, and segmenting brain MRI to account for gray white, white matter, and CSF water contents in each voxel (which also affects the water relaxation times) on a voxel-by-voxel basis. Also, for molecular concentrations to be estimated, a knowledge of peak assignments and number of protons per functional group would be required.

As is well known for high field human MR systems, transmit B₁ is less homogeneous than at lower field strengths (e.g. 3T or lower) and conventional amplitude-modulated RF pulses will not able to achieve a uniform flip angle for all regions of the brain. With the current 8-channel transmit head coil, B₁ maps (as shown in Figure 5), indicate some central regions with higher than nominal flip angle, with lower flip angles in lateral, posterior and particularly inferior (i.e. cerebellar) brain regions. This causes signal losses in these areas in both MRI and MRSI data. In future studies, this could be corrected by the using transmit B₁ maps together with knowledge of signal dependance on flip angle (52) to perform post-hoc intensity corrections.

The current study also used a 16-channel local coil (LC) shim array (in addition to the standard spherical harmonic shims on the scanner (up to 3^rd order)) to improve B₀ field homogeneity. Spectral quality including better linewidths and less residual water signals, particularly in more peripheral brain regions (data not shown) was clearly improved using the LC shim array, and less artifactual hyper- or hypo-intensity was noted in the amide maps. In one subject, the average magnitude water linewidth decreased from 43.1 ± 18.6 to 35.4 ± 12.9 Hz over the volume of the supratentorial brain using the LC shim array, and improvement of ~18%. Similarly, the distribution (standard deviation_ of water frequencies over the same volume decreased from 19.5 to 11.1 Hz, an improvement of ≈43%.

The present study found DF peaks to be relatively uniformly distributed throughout the brain, with only a small number of peaks showing any significant regional variations. Visual inspection of metabolite maps only shows clear reductions in signal intensity in CSF regions such as the lateral ventricles (e.g. Figures 5, 6). The lack of regional brain variations may in part be due to relatively large variance in the measurements, particularly for some of the smaller DF peaks, but also is consistent (for instance) with APT-CEST measurements of human brain protein amides, which in the normal brain are quite homogeneous and do not typically show significant regional or tissue-type variations(53). Also, maps of N-acetylaspartate (NAA, which also has a DF amide resonance) created from conventional upfield MRSI measurements are typically fairly uniform in normal brain and show little contrast between white matter and gray matter (54).

While the current study demonstrates 3D DF-MRSI at 7T with acceptable SNR, PNS and scan time, it does also have several limitations. These include a relatively small sample size, and no reproducibility measurements. In the future, accelerations methods such as sparse k-space sampling and low rank image reconstruction (55) may also be useful in further reducing scan time beyond the SENSE acceleration used here.

In conclusion, high-field DF-MRSI is able to spatially map exchangeable protons in the human brain at high resolution and with near whole-brain coverage in acceptable scan times, and in the future may be used to study metabolism of brain tumors or other neuropathological disorders.

Supplementary Material

Tab S1

Table S1. Table shows the results of statistical tests between brain regions using amide (8.x ppm) estimations from LCModel.

Fig S2

Figure S1. Reconstructed amide (8.x) maps for all ten subjects.