High-resolution metabolic mapping of the cerebellum using 2D zoom magnetic resonance spectroscopic imaging

Abstract

Purpose:

The human cerebellum plays an important role in the functional activity of the cerebrum, ranging from motor to cognitive systems given its relaying role between the spinal cord and cerebrum. The cerebellum poses many challenges to Magnetic Resonance Spectroscopic Imaging (MRSI) due to its caudal location, susceptibility to physiological artifacts, and partial volume artifacts resulting from its complex anatomical structure. Thus, in the present study, we propose a high-resolution MRSI acquisition scheme for the cerebellum.

Methods:

A zoom or reduced field of view (rFOV) metabolite-cycled MRSI acquisition at 3 Tesla, with a grid of 48 × 48, was developed to achieve a nominal resolution of 62.5 μL. Single-slice rFOV MRSI data were acquired from the cerebellum of 5 healthy subjects with a nominal resolution of 2.5 × 2.5 × 10 mm³ in 9.6 min. Spectra were quantified using the LCModel package. A spatially unbiased atlas template of the cerebellum was used to analyze metabolite distributions in the cerebellum.

Results:

The superior quality of the achieved spectra-enabled generation of high-resolution metabolic maps of total N-acetylaspartate, total Creatine (tCr), total Choline (tCho), glutamate+glutamine, and myo-inositol, with Cramér-Rao lower bounds below 50%. A template-based regions of interest (ROI) analysis resulted in spatially dependent metabolite distributions in 9 ROIs. The group-averaged high-resolution metabolite maps across subjects increased the contrast-to-noise ratio between cerebellum regions.

Conclusion:

These findings indicate that very high-resolution metabolite probing of the cerebellum is feasible using rFOV or zoomed MRSI at 3 Tesla.

1 |. INTRODUCTION

Compared to the cerebrum, the cerebellum constitutes only 10% of total brain volume; however, it has 3 times more neurons than the cerebrum¹ and consumes around 20% of the whole energy utilization of the brain.² The human cerebellum plays an important role in functional activity of the brain, which ranges from motor to cognitive systems due to its relaying role between the spinal cord and cerebrum. With the development of neuroimaging modalities, the cerebellum has been a focus of interest, but its complex structure and caudal location pose technical challenges for imaging modalities.³ Increasingly, advanced MRI technologies, including anatomical and functional modalities, have facilitated the evaluation of the cerebellum.^4,5 This is of great importance because abnormalities in the cerebellum are associated with many neurological, neurodevelopmental, and psychiatric disorders.^6–9

Magnetic Resonance Spectroscopy (MRS) techniques are increasingly used to enable the neurochemical probing of brain tissue.¹⁰ Single voxel (SV) MRS has been the method of choice to study the cerebellum, with a nominal voxel dimension ranging from 4090 μL¹¹ to 15620 μL¹² for different neurological conditions. However, spectra in single SV-MRS acquisitions are derived from a mixture of different tissue compartments, which results in a partial volume effect. This partial volume effect is particularly problematic in the cerebellum due to its complex structures.

Alternatively, MRS imaging (MRSI) methods record multiple spectra from different regions simultaneously. The widespread use of MRSI has been limited by several challenges, including low signal to noise ratio (SNR) due to suboptimal sequences, inhomogeneity of the main (B₀) and radiofrequency (B₁) magnetic fields, and long acquisition times.¹³ As increasingly advanced MRI acquisition and reconstruction approaches have been developed, sensitivity and resolution of MRSI have drastically improved. Recently, a nominal resolution of 110 μL has been reported for 3D-MRSI of the whole-cerebrum at 3 Tesla (T).¹⁴ Further improvements in resolution up to 36 μL have been reported with the use of constrained reconstruction of 3D-MRSI at 3T.¹⁵ Simultaneously, efforts at ultrahigh field (UHF ≥ 7T) MRI have resulted in further improvements in resolution both for 2D and 3D MRSI. A nominal resolution of 24 μL with an acquisition duration of 46 min has been reported at 9.4T using 2D MRSI.¹⁶ As for the 3D-MRSI, a nominal resolution of 20 μL has been achieved within 17 min at 7T.¹⁷ Although there have been numerous achievements for neurochemical mapping cerebrum using MRSI both at 3T and UHF, the cerebellum has not been a focus of interest due to the caudal location and the susceptibility to physiological artefacts. Only very few studies have been performed using MRSI in the cerebellum, with coarse nominal resolutions ranging from 310 μL¹⁸ to 1210 μL.¹⁹

Zoom-MRI or reduced field of view (rFOV) acquisition schemes overcome the limitations of spatial resolution and low SNR for MRI. This has been achieved using outer volume suppression²⁰ or a well-defined spatial excitation.²¹ Similar to zoom-MRI, 3-pulse localization of MRSI suppresses the signal outside of the FOV and is used for rFOV MRSI measurements to reduce acquisition duration by reducing the number of phase-encoding steps.^22,23 In contrast to previous rFOV MRSI implementations, the present study sought to develop an accelerated rFOV acquisition scheme without reducing phase encoding steps for high-resolution short TE semi-LASER MRSI at 3T with a nominal resolution of 62.5 μL. In order to increase the SNR of the zoomed acquisition, k-space data were acquired with a sampling density proportional to a Hanning window.²⁴ Further SNR gain was achieved using a narrower acquisition bandwidth. In 5 healthy subjects, we demonstrated that the proposed rFOV-MRSI with accelerated k-space trajectory generates high-resolution metabolic maps at 3T across an entire cerebellum slice with a thickness of 10 mm. Finally, we explored neurochemical alterations in regions of interest (ROIs) of the cerebellum using the spatially unbiased atlas template of the cerebellum (SUIT atlas).⁴

2 |. METHODS

Five healthy subjects (3 females, 25.5 ± 7.07 years [mean ± standard deviation (SD)]) participated. All subjects provided informed consent before the in vivo MRI exam, which was approved by the Purdue University Institutional Review Board. Before being scanned, informed written assent was obtained from all the subjects.

2.1 |. MRI data acquisition

The data were acquired using a Siemens Prisma 3T MR system (Siemens, Germany) with a 64-channel (N_channels) head array receive coil. A T1-weighted MPRAGE dataset (TR = 1900 ms, TE = 2.13 ms, TI = 900 ms, flip angle = 8°, 176 transverse slices, 0.9 × 0.9 × 1 mm³ voxels) was acquired for each subject for MRSI acquisition planning. B₀ shimming was achieved using the vendor-provided procedure, gradient echo shimming. The vendor-provided “patient-specific” B₁ shimming option was used.

2.2 |. Reduced FOV or zoom metabolite-cycled semi-laser MRSI

Metabolite-cycling MRSI was acquired using the same parameters as described in Emir et al.²⁵ Briefly, before the semi-LASER localization, 2 asymmetric inversion RF pulses in alternating TRs were used for downfield/upfield (N_directions = 2 [upfield/downfield]) measurements.

The semi-LASER localization (Figure 1A) (Supporting Information Figure S1) was positioned in the central cerebellar slice in the transverse plane, without any angulation, to include right crus I, left crus I, and dentate nuclei (Figure 1B). The high in-plane nominal resolution (2.5 mm × 2.5 mm) with a thickness 10 mm was achieved using rFOV = 120 mm × 120 mm × 10 mm, semi-LASER localization = 100-90 mm × 50-40 mm × 10 mm, TR = 1500 ms, and TE = 32 ms. The semi-LASER volume of interest was placed asymmetrical in the rFOV.

FIGURE 1.

(A) Representative semi-LASER localization (red boxes) and resulting water image of MRSI acquisition (gray shade) overlaid on top of high-resolution MPRAGE images. (B) ROI analysis is illustrated on the same subjects’ MPRAGE axial image. Numbers from 1 to 18 represent ROIs in the cerebellar MNI FNIRT Maxprob thr25-2 mm SUIT atlas. MRSI, MRS imaging; ROI, region of interest

To cover the 48 × 48 grid in a rFOV, we modified our 2D-density weighted concentric ring trajectory (DW-CRT) design.²⁴ Each ring in DW-CRT consisting of 64 points (number of points per ring, N_{p_ring} = 64) was collected with an analog-digital-converter bandwidth of 40 kHz. This resulted in 128 spectral points (N_sp = 128). Two temporal interleaves (number of temporal interleaves, N_ti = 2) with inverted readout gradients were collected with an spectral bandwidth of 625 Hz.²⁴ The total acquisition window was 205 ms for 256 spectral points. The rest of the DW-CRT parameters were identical to the previous 3T and 7T implementations^24,26: 4 spatial interleaves (N_si = 4 and N_avg = 1) of 24 DW concentric rings (N_ring = 24) were acquired with α = 1. The total duration for the DW-CRT was 9 min 36 s ((N_si = 4) × (N_directions = 2) × (N_ring = 24) × (N_ti = 2) × (TR = 1500 ms) = 576 s).

2.3 |. Postprocessing

MatLab (R2016b, MathWorks, Natick, MA) was used for reconstruction. To achieve the full spectral bandwidth (1250 Hz), non-uniform fourier transform²⁷ was used to calculate the Fourier transform of the acquired k-space data.²⁴ The metabolite-cycled spectra were processed as described in Emir et al.²⁵ A Gaussian filter of 250-ms timing parameters was applied to the FIDs. The non–water-suppressed spectra-enabled frequency and phase corrections for each voxel with a dimension of 62.5 μL. The metabolite spectra were obtained via subtracting the upfield and downfield FIDs, whereas summing them generated the water-only spectra. The remaining water signal in the metabolite spectrum was removed using the Hankel-Lanczos singular value decomposition.²⁸ A lipid-basis penalty algorithm was used to minimize the lipid contamination.²⁹ Water-only and metabolite-only whole cerebellum slice images were used to generate a brain and lipid mask for the lipid-basis penalty algorithm.

2.4 |. Metabolite quantification

Metabolite quantification was performed using LCModel package (Version 6.3-1M).³⁰ The model spectra including 8 LCModel-simulated macromolecule resonances (positioned at 0.91, 1.21, 1.43, 1.67, 1.95, 2.08, 2.25, and 3 ppm) was used as a basis set for the LCModel analysis, as described in Steel et al.²⁶ The basis set included the model spectra of alanine, aspartate, ascorbate/vitamin C, glycerophosphocholine, phosphocholine (PCho), creatine (Cr), phosphocreatine (PCr), GABA, glucose, glutamine (Gln), glutamate (Glu), glutathione, lactate, myo-inositol (myo-ins), N-acetyl-aspartate (NAA), N-acetylaspartylglutamate (NAAG), phosphoethanolamine, scyllo-Inositol, and taurine. The LCModel fitting was performed over the spectral range from 0.5 to 4.2 ppm. The resulting Cramér-Rao lower bounds of estimated metabolites less than 50% were reported. Summed metabolite estimations were reported for total NAA (NAA+NAAG, tNAA), total Cr (Cr + PCr, tCr), total Cho (GPC + PCho, tCho), and Glu and Gln (Glu + Gln, Glx) because their pairwise-correlation estimates were high (correlation coefficient < −0.5). The normalized metabolite levels were reported for each subject using the water-scaled signal intensity of each metabolite (raw):

$$\text{Metabolite – normalized}=\text{Metabolite – raw}/\left({\text{tCho}}_{\text{raw}}+{\text{tCr}}_{\text{raw}}+{\text{tNAA}}_{\text{raw}}+myo-{\text{Ins}}_{\text{raw}}+{\text{Glx}}_{\text{raw}}\right).$$

Metabolite levels were also reported using water scaling and relative to tCr approaches. These metabolite levels were not corrected for T₁ and T₂ effects, partial volume, and water content differences.

2.5 |. Regional distributions of neurochemical in the cerebellum

The cerebellar ROIs were obtained using the Cerebellar MNI FNIRT Maxprob thr25-2 mm of SUIT Atlas (http://www.diedrichsenlab.org/imaging/propatlas.htm). Each subject’s MRSI slice and metabolite maps was determined in the MPRAGE image space using statistical parametric mapping.³¹ The resulting MRSI resolution was 0.9 × 0.9 × 1 mm³. Then, ROIs of the SUIT atlas were warped into each subject’s MPRAGE image space in the following manner (Supporting Information Figure S2): each subject’s MPRAGE image was nonlinearly registered to the Montreal Neurological Institute-152 template using the FMRIB nonlinear registration tool (FNIRT); afterward, the resulting warp field was applied to the Cerebellar MNI FNIRT Maxprob thr25-2 mm of the SUIT atlas to transform them to each subject’s MPRAGE image. Mean metabolite levels in cerebellar ROIs were calculated using fslstats³² in the center slice of MRSI in the transverse plane.

To generate group-averaged metabolite spatial maps, metabolite maps were transformed to the MNI-152 template by applying the resulting warp field; afterward, metabolite maps were averaged across the 5 subjects.

2.6 |. Statistical analysis

Only the ROIs that were consistently detected across subjects (in at least half of the subjects) were included in statistical analyses. As a secondary filter to select reliable ROIs, the mean number of voxels in each ROIs had to be 1 SD above the mean SD.

The mean regional metabolite-normalized levels from the aforementioned reliable ROIs of cerebellar atlas were used for statistical analyses using SAS software v.9.4 (SAS Institute, Cary, NC). MRS data from different ROIs were compared using a 1-way repeated measures analysis of variance separately for each reported normalized metabolites (tNAA, tCr, tCho, Glx, and myo-ins) across each target brain region (within-subject factor). Given the occurrence of missing data due to exclusion for unreliability, we utilized restricted maximum likelihood to fit the models. Each analysis examined the differences in MRS data across 9 ROIs, resulting in 36 possible pairwise comparisons. We employed a Benjamini-Hochberg correction within each family of comparisons in order to control the false discovery rate.³³ The same statistical analysis was performed for the water-scaled and relative to tCr approaches.

3 |. RESULTS

Figure 1A shows coronal, sagittal, and axial images of an anatomical scan, including the placement of the MRSI slice and semi-LASER localization. The image derived from the non-water suppressed MRSI data shows the signals from the cerebellum. ROIs of the SUIT atlas encompassed by the MRSI localization are illustrated in Figure 1B. The list of ROIs and the mean number of voxels for each metabolite that met our criteria for reliable quantification in the SUIT atlas are listed in Table 1.

TABLE 1.

The list of ROIs and the mean number of voxels across subjects for each metabolite calculated in the center slice of semi-LASER VOI transformed to MPRAGE dimensions

		Mean Number of Voxel ± SD
		tNAA	tCr	tCho	Glx	Myo-ins
ROIs	Dentate nucleus	150 ± 44	150 ± 44	150 ± 44	131 ± 45	148 ± 41
	Right crus I	961 ± 89	961 ± 89	926 ± 83	915 ± 109	945 ± 94
	Left crus I	817 ± 193	817 ± 193	774 ± 191	761 ± 188	794 ± 198
	Right VI	488 ± 181	488 ± 181	484 ± 179	471 ± 176	485 ± 183
	Vermis crus II	91 ± 27	91 ± 27	91 ± 27	91 ± 27	91 ± 27
	Vermis VIIB	45 ± 17	45 ± 17	45 ± 17	45 ± 17	45 ± 17
	Vermis VIIIa	91 ± 7	91 ± 7	91 ± 7	91 ± 7	91 ± 7
	Vermis IX	54 ± 20	54 ± 20	54 ± 20	51 ± 21	54 ± 20
	Vermis X	38 ± 14	38 ± 14	33 ± 16	35 ± 15	38 ± 15

Note: The resulted resolution of the center slice was 0.9 × 0.9 mm². ROIs are in rows and metabolites are in columns.

Abbreviations: Glx, glutamate + glutamine levels; myo-ins, myo-inositol; ROI, region of interest; tCho, total Cho; tCr, total Cr; tNAA, total NAA; VOI, volume of interest.

Figure 2A,B show the zoomed MRSI spectra with LCModel fits. Even at a nominal resolution of 62.5 μL at 3T, spectra from 5 different volumes of interest are of high quality enough for reliable quantitation. Although LCModel-simulated macromolecule resonances account for most of residual lipid signal in the spectral range of 0.5 to 1.8 ppm, some voxels have negative peaks at 1.8 ppm (gray box in Figure 2B) due to regularization error of the L2 lipid removal procedure.

FIGURE 2.

(A) Extracted spectra (blue) from 9 voxels (2.5 mm × 2.5 mm × 10 mm each) of 5 different volumes of interests with LCModel fit (black). The semi-LASER localization (yellow box) and resulting water image of MRSI acquisition (gray shade) overlaid on top of high-resolution MP-RAGE image. (B) A single spectrum (red box in (A)) illustrates the quality of the achieved spectrum (blue) and LCModel fit (black). Gray box shows residual lipid artifact due to the L2 lipid removal procedure

The mean CRLB values for each ROI across subjects are reported in Table 2. In most of the voxels, the mean CRLB values after the Gaussian filter resulted in values less than 30% for the metabolites tNAA, tCr, tCho, myo-ins, and Glx. The resulting normalized metabolite maps and corresponding CRLB maps of a healthy subject are illustrated in Figure 3A,B, respectively.

TABLE 2.

The list of ROIs and the mean CRLBs across subjects for each metabolite

		Mean CRLB ± SDs
		tNAA	tCr	tCho	Glx	Myo-ins
ROIs	Dentate nucleus	9.0 ± 1.9	8.3 ± 2.4	10.3 ± 2.0	26.3 ± 8.5	10.0 ± 3.7
	Right crus I	9.7 ± 1.9	6.4 ± 0.9	11.9 ± 1.7	15.1 ± 1.7	10.8 ± 1.6
	Left crus I	11.4 ± 1.6	6.7 ± 1.0	13.0 ± 2.5	17.4 ± 3.8	11.3 ± 2.2
	Right VI	7.2 ± 0.8	4.7 ± 1.1	8.7 ± 0.7	12.6 ± 0.9	6.7 ± 0.3
	Vermis crus II	6.5 ± 0.3	3.9 ± 0.4	8.4 ± 1.2	9.0 ± 1.2	6.9 ± 1.3
	Vermis. VIIB	6.2 ± 1.2	3.5 ± 0.5	7.2 ± 1.2	11.7 ± 3.1	5.8 ± 0.8
	Vermis VIIIa	7.2 ± 2.2	4.5 ± 0.9	8.3 ± 1.2	12.6 ± 4.5	6.6 ± 1.2
	Vermis IX	11.7 ± 5.0	6.2 ± 1.4	10.7 ± 1.9	18.3 ± 7.6	8.3 ± 1.3
	Vermis X	22.4 ± 7.0	16.25 ± 11.0	29.0 ± 12.2	22.4 ± 9.7	17.0 ± 11.2

Note: ROIs are in rows, and metabolites are in columns.

Abbreviations: CRLB, Cramer-Rao lower bound.

FIGURE 3.

Metabolite and CRLB distribution maps after the Gaussian filter obtained with rFOV MRSI from a subject. Normalized metabolite level maps from a nominal voxel dimension of 62.5 μL for tNAA, tCr, tCho, Glx, and myo-ins overlaid on an anatomical image. a.u., arbitrary units; CRLB, Cramer-Rao lower bound; Glx, glutamate + glutamine levels; myo-ins, myo-inositol; rFOV, reduced FOV; tCho, total Cho; tCr, total Cr; tNAA, total NAA

Figure 4 shows the group-average metabolite maps of normalized metabolite levels in the MNI-152 template space. The mean normalized metabolite levels that met our criteria for reliable quantification in cerebellar ROIs (9 cerebellar ROIs) of the SUIT atlas are shown in Figure 5 (Supporting Information Table S1). The normalized metabolite levels revealed different neurochemical profiles that may characterize different cerebellar regions (Supporting Information Table S4). For instance, the tCho, tNAA, and myo-ins levels in the dentate nuclei are significantly higher than the right and left crus I and vermis ROIs, whereas tCr is significantly lower compared to the right and left Crus I and vermis ROIs. Similarly, dentate nuclei has higher tCho, tNAA, and myo-ins levels compared to the vermis ROIs. Similar findings were also observed for water-scaled metabolite levels and metabolite ratios (Supporting Information Tables S2 and S3 for metabolite levels; Supporting Information Tables S5 and S6 for statistical findings). However, statistical significance degraded for water-scaled metabolite levels (Supporting Information Table S6).

FIGURE 4.

Group-average normalized metabolite maps overlaid on Montreal Neurological Institute-152 template

FIGURE 5.

Neurochemical profiles from different brain regions determined by LCModel fitting of rFOV MRSI with a nominal resolution of 2.5 mm × 2.5 mm × 10 mm acquired at 3T. Only metabolites and ROIs that were quantified reliably are shown. Metabolites significantly different between regions (P < .05, ANOVA) were reported in the Supporting Information Table S4. Error bars: intersubject SD (N = 5 for all brain regions). ANOVA, analysis of variance; T, Tesla

4 |. DISCUSSION

This study has demonstrated the advantage of rFOV (zoom) acquisition, zoomed, with a DW-CRT short TE semi-laser localization to improve the resolution and accuracy of whole cerebellum slice metabolic maps at 3T. The improved resolution and spectral quality-enabled reliable quantification of tNAA, tCr, tCho, Glx, and myo-ins from a nominal voxel volume of 62.5 μL. The metabolite distribution maps for different ROIs in the cerebellum were consistent with the literature (see below). To our knowledge, this is the first study to have demonstrated a higher nominal resolution of 62.5 μL at 3T for the cerebellum.

Implementing the rFOV acquisition at 3T for high-resolution MRSI faces several challenges caused by hardware and SNR ratio limitations. Mainly, high-resolution MRSI puts demands on maximum gradient amplitudes, rise time, and slew rate. To decrease the FOV while maintaining the acquisition matrix size, one might choose to increase the gradient amplitude and slew rate or to reduce the spectral bandwidth resulting in lower spectral sampling. In this study, to reach the nominal in-plane resolution of 2.5 mm × 2.5 mm, we decided to utilize reduced spectral bandwidth (625 Hz vs. 1250 Hz) compared to previous implementations.²⁶ The spectral bandwidth was doubled (1250 Hz) using an inverted readout gradient trajectory without causing any additional noise.²⁴ The use of DW-CRT and its reconstruction pipeline further contributed the SNR improvement, as discussed previously by Chiew et al.²⁴ Besides, a lower acquisition bandwidth of 40 kHz compared to the previous implementation of 80 kHz^25,26,34 minimized a larger amount of noise sampled due to the narrower frequency range. Moreover, the 64-Channel head/neck coil allowed the incremental SNR gains in the peripheral ROIs closest to the coil elements such as the cerebellum.³⁵ Finally, additional SNR improvement was achieved through the use of metabolite cycling, allowing voxel-vise preprocessing steps.²⁵

The present study did not directly compare this acquisition to others at 3T in which the nominal resolution ranged from 310 μL¹⁸ to 1210 μL¹⁹ for the cerebellum. We believe that our strategy offers substantial resolution improvement for 2D MRSI compared to the other implementations of the cerebrum at 3T and UHF. When the acquisition duration and magnetic field difference were considered, the achieved nominal resolution of 62.5 μL and provided metabolite maps within 9 min 36 s in the cerebellum are comparable with the 3T and the UHF 2D MRSI in the cerebrum. For instance, a nominal voxel size of 24 μL at 9.4T¹⁶ and of 29 μL at 7T³⁶ were achieved with acquisition durations of 46 and 42 min, respectively. With the use of constrained reconstruction methods, the proposed strategy is expected to provide higher resolution.^15,37

The rFOV MRSI acquisition with a resolution of 62.5 μL allowed reliable quantification of major brain metabolites in the cerebellum (Figures 2 and 3). Concentration distributions in the cerebellum of the reported metabolites revealed significant variations between the cerebellar vermis, white, and gray matter of both cerebellar hemispheres, and these were in line with previous MRSI and SV-MRS studies. For example, the high cerebellar total creatine level in the gray matter agreed with previous reports.¹⁹ In addition, the higher Glx level obtained in cerebellar gray matter was consistent with the previous SV-MRS studies.³⁸ The highest total choline levels were observed in the cerebellar white matter, also in agreement with SV-MRS.³⁸

The voxel-based comparisons of the SUIT atlas ROI⁴ analyses of whole cerebellum slice metabolite maps demonstrated that metabolite distributions are spatially dependent. The dentate nuclei, the link in the cortico-cerebellar closed loop circuits,³⁹ had higher tNAA, tCho, and myo-ins levels. Because NAA was extensively distributed in the cerebellum and NAAG was very high in the deep cerebellar nuclei, higher tNAA levels observed in the dentate nuclei are in agreement with previous studies.⁴⁰ Because the dentate nuclei are embedded in the white matter having a higher glia-to-neuron ratio than gray matter,⁴¹ myo-ins and tCho as markers of glial cells resulted in higher levels in the dentate nuclei. In line with the cyto and receptor architectonic mapping of glutamate receptors in the human cerebellum,⁴² Glx levels were higher in the cerebellar cortex (right crus I, left crus I, and vermis) than in the dentate nuclei. Finally, higher creatine levels in the cerebellar cortex might suggest energy demand of Purkinje cells.¹⁹ With the accurate coregistration between the MRSI and MNI template, high-quality group-average metabolite distributions of tNAA, tCr, tCho, Glx, and myo-ins were achieved.

The number of subjects is a limitation, and further experiments are required for the clinical validity of the proposed method. In addition, a single slice acquisition is not sufficient to map the entire cerebellum. Although the use of DW-CRT with α = 1 resulted in an improved SNR,²⁴ it deteriorates the point spread function and dilutes metabolite levels in small ROIs such as the dentate nuclei. Moreover, limited coverage of the single-slice measurement of zoom MRSI can be overcome using multi-slice stacks with different slice orientations, providing improved coverage and resolution in the cerebellum.⁴³ The use of a long TR of 1.5 s and the cost of increased total acquisition duration due to the choice of adiabatic pulses can be mitigated by gradient-modulated offset-independent adiabaticity pulses.⁴⁴ Together with temporal interleaves (N_ti = 2), total acquisition was 9 min 36 s. Recently, we have reduced acquisition duration to 4 min 48 s by reducing the spectral bandwidth to 893 Hz (Nti = 1) with the use of maximum slew rate of the gradient system (data were not shown). Although multiplet structures can still be resolved with acquisition window of 205 ms (corresponding to a spectral sampling resolution of 5 Hz), maintaining an high spectral sampling resolution is recommended for better separation.⁴⁵ As in all spectral subtraction-based techniques, the metabolite-cycling technique relies on homogenous B₁ and B₀ fields for an effective inversion profile. Although it had reported previously that the metabolite-cycling technique resulted in higher metabolite levels compared to the water suppression techniques,^25,46 the metabolite signals in the spectral range between 4.2 and 3.8 ppm might be affected by the adiabatic inversion pulses. Thus, further advances in hardware such as multi-transmit and shim coils are expected to improve the performance the metabolite cycling by providing homogeneous B₁ and B₀ fields in the cerebellum. Finally, the exact quantification of peaks might be affected by the use of simulated macromolecules and lipid removal procedures.

5 |. CONCLUSION

In conclusion, the rFOV 2D MRSI resulted in very high-resolution metabolite maps of the cerebellum. These initial findings indicate that optimal methods can be developed for the future to generate a probabilistic metabolic atlas of the cerebellum.

Supplementary Material

2

FIGURE S1 The MRSI volume of interest positioning for all subjects. The semi-LASER localizations (red boxes) and resulting water image of MRSI acquisitions (gray shade) overlaid on top of high-resolution MP-RAGE image

FIGURE S2 Each subject’s MRSI slice and metabolite maps were determined in the MP-RAGE image space using SPM. The Cerebellar MNI FNIRT Maxprob thr25-2mm of the SUIT atlas was transformed to each subjec’s MPRAGE image space. Mean metabolite levels in cerebellar ROIs were calculated in the center slice of MRSI in the transverse plane. The representative ROI was illustrated for dentate nuclei (white mask in the bottom row)

TABLE S1 Neurochemical profiles from different brain regions determined by LCModel fitting of rFOV MRSI with a nominal resolution of 2.5 mm × 2.5 mm × 10 mm acquired at 3T. Mean ± standard deviation of normalized metabolite levels is in arbitrary units

TABLE S2 Neurochemical profiles from different brain regions determined by LCModel fitting of rFOV MRSI with a nominal resolution of 2.5 mm × 2.5 mm × 10 mm acquired at 3T. Mean ± standard deviation of metabolite ratios (tNAA/tCr, tCho/tCr, Glx/tCr and myo-ins/tCr) is in μmol/g assuming tCr concentration to be 8 μmol/g

TABLE S3 Neurochemical profiles from different brain regions determined by LCModel fitting of rFOV MRSI with a nominal resolution of 2.5 mm × 2.5 mm × 10 mm acquired at 3T. Mean ± standard deviation of water scaled metabolite levels is in μmol/g

TABLE S4 Pairwise contrasts for normalized metabolite levels with explicit p-values and Benjamini-Hochberg corrected versions of the differences in MRS data across 9 ROIs, resulting in 36 possible pairwise comparisons. We employed a Benjamini-Hochberg correction within each family of comparisons in order to control the false discovery rate of 5%. Statistically significant P-values are highlighted in bold font

TABLE S5 Pairwise contrasts for metabolite ratios (tNAA/tCr, tCho/tCr, Glx/tCr and myo-ins/tCr) with explicit P-values and Benjamini-Hochberg corrected versions of the differences in MRS data across 9 ROIs, resulting in 36 possible pairwise comparisons. We employed a Benjamini-Hochberg correction within each family of comparisons in order to control the false discovery rate of 5%. Statistically significant P-values are highlighted in bold font

TABLE S6 Pairwise contrasts for water scaled metabolite levels with explicit p-values and Benjamini-Hochberg corrected versions of the differences in MRS data across 9 ROIs, resulting in 36 possible pairwise comparisons. We employed a Benjamini-Hochberg correction within each family of comparisons in order to control the false discovery rate of 5%. Statistically significant P-values are highlighted in bold font

DATA AVAILABILITY STATEMENT

The data that support the findings of this study are available from the corresponding author upon reasonable request.