Fast, regional three-dimensional hybrid (1D-Hadamard 2D-rosette) proton MR spectroscopic imaging in the human temporal lobes

Abstract

¹H-MRSI is commonly performed with gradient phase encoding, due to its simplicity and minimal radio frequency (RF) heating (specific absorption rate). Its two well-known main problems—(i) “voxel bleed” due to the intrinsic point-spread function, and (ii) chemical shift displacement error (CSDE) when slice-selective RF pulses are used, which worsens with increasing volume of interest (VOI) size—have long become accepted as unavoidable. Both problems can be mitigated with Hadamard multislice RF encoding. This is demonstrated and quantified with numerical simulations, in a multislice phantom and in five healthy young adult volunteers at 3 T, targeting a 2-cm thick temporal lobe VOI through the bilateral hippocampus. This frequently targeted region (e.g. in epilepsy and Alzheimer’s disease) is subject to strong, 1–2 ppm.cm⁻¹ regional B0, susceptibility gradients that can dramatically reduce the signal-to-noise ratio (SNR) and water suppression effectiveness. The chemical shift imaging (CSI) sequence used a 3-ms Shinnar–Le Roux (SLR) 90° RF pulse, acquiring eight steps in the slice direction. The Hadamard sequence acquired two overlapping slices using the same SLR 90° pulses, under twofold stronger gradients that proportionally halved the CSDE. Both sequences used 2D 20 × 20 rosette spectroscopic imaging (RSI) for in-plane spatial localization and both used RF and gradient performance characteristics that are easily met by all modern MRI instruments. The results show that Hadamard spectroscopic imaging (HSI) suffered dramatically less signal bleed within the VOI compared with CSI (<1% vs. approximately 26% in simulations; and 5%–8% vs. >50%) in a phantom specifically designed to test these effects. The voxels’ SNR per unit volume per unit time was also 40% higher for HSI. In a group of five healthy volunteers, we show that HSI with in-plane 2D-RSI facilitates fast, 3D multivoxel encoding at submilliliter spatial resolution, over the bilateral human hippocampus, in under 10 min, with negligible CSDE, spectral and spatial contamination and more than 6% improved SNR per unit time per unit volume.

1 |. INTRODUCTION

Proton magnetic resonance spectroscopic imaging (MRSI) of the human brain is almost universally spatially encoded using field gradients. Existing sequences for covering the whole brain often excite a large slab and use a combination of phase encoding along the slices’ direction,^1,2 and various k-space readouts in-plane. For example, echo-planar spectroscopic imaging, spirals, concentric rings,^3–5 rosettes and their variants^6–8 facilitate partial to whole brain coverage at submilliliter resolutions in under 20 min.^9–11 However, often only regional coverage is needed, such as when imaging the temporal lobes in search of an epileptogenic focus,^12,13 or imaging peritumoral tissue prior to a surgical intervention or following radiation therapy.^14,15 In these cases, obtaining high quality spectroscopic data from a small number of slices, quickly, may be more important than extended or even whole-brain coverage.

Unfortunately, slab excitation, combined with phase encoding of a few (i.e. <8) slices degrades the MRSI data and its localization accuracy in several ways: first, radio frequency (RF) slab-profile imperfections and finite transition bands excite spins outside the slab. To avoid signals from those spins aliasing into the slab, the field of view (FOV) must be larger than the slab thickness. This becomes an issue due to the “SINC-like” nonlocal nature of the spatial response—that is, point spread function (PSF)—associated with the small number of phase-encoding steps. The PSF causes (i) the signals excited outside the prescribed slab to bleed into it; and (ii) mixes signals between slices within the slab. Because the PSF mostly depends on the thickness of the slices, increasing the number of phase-encoding steps and FOV along the slice direction, while keeping the slice thickness fixed, does little to ameliorate this.^16–18 Second, another source of errors comes from the limited transmit field (B₁₊) of clinical imagers, which restricts the bandwidth (BW) of the slab-selective RF pulses. Low BW pulses are problematic for selecting thick slabs comprising more than one slice, as their associated chemical shift displacement error (CSDE) can be substantial and lead to localization errors of well over 10%.¹⁹

Both issues can be addressed with Hadamard spectroscopic imaging (HSI) along the slice direction.¹⁹ Unlike phase encoding, which uses a single slab-selective pulse, HSI employs several pulses that excite separate slices and can be cascaded (i.e. played sequentially).¹⁹ Because each cascaded pulse excites just one slice, its selection gradient can be kept as high as the B₁₊ permits, regardless of the volume of interest (VOI) thickness or number of slices,¹⁹ minimizing the CSDE, leading to sharper slice profiles that improve the signal-to-noise ratio (SNR) and spatial localization^20,21 and avoids the nonlinear RF interactions incurred when the HSI pulses are superposed (i.e. applied simultaneously).²² In this paper we demonstrate and quantify these advantages by comparing two sequences for ¹H-MRSI: cascaded, transverse HSI along the slice direction,¹⁹ combined with 2D 20 × 20 rosette spectroscopic imaging (RSI) in-plane²³; and phase-encoded MRSI with an equivalent slab-selective excitation and identical in-plane encoding. Both approaches yield fast 3D ¹H-MRSI data in under 10 min. They are compared in a phantom and the bilateral human temporal lobes, a common target in epilepsy and Alzheimer’s disease studies,^24,25 but which suffer severe B₀ susceptibility gradients,²⁶ hence can benefit from a limited shimming volume making it well suited for Hadamard encoding.

2 |. METHODS

2.1 |. 3D MRSI sequences

Two ¹H-MRSI sequences were compared (TR/TI/TE = 2000/240/40 ms), as shown in Figure 1. They differ in localization along the direction of the VOI’s thin aspect. First, a two-slice HSI, with two cascaded 3-ms, 3.2-kHz BW, Shinnar–Le Roux (SLR) 90° RF pulses,^27,28 shown in Figure 1A¹⁹ and subsequently referred to as HSI. The two HSI-encoded slices are slightly offset (approximately 10% of their width) to further reduce crosstalk.²⁹ The second sequence excites a slab with the same SLR 90° with a thickness equal to that of both HSI slices (2 cm); consequently, the pulse employs only half the gradient strength. This sequence also uses chemical shift imaging (CSI) across it over a FOV with ×8 phase-encoding steps, chosen to yield the same nominal slice width, as shown in Figure 1B. This sequence will subsequently be referred to as CSI. Acquisitions were 8 and 5 min long for the CSI and HSI with one and two averages, respectively. Total acquisition times were chosen to ensure quality spectra; however, because our comparisons are in terms of SNR per unit time per unit volume (SNR_VT), no attempt was made to match acquisition times. Each sequence was rerun with identical parameters but without water suppression and IR pulses, at minimal TR = 390 ms to obtain water reference maps in less than 2 min.

FIGURE 1.

The two TI/TR/TE = 240/2000/40 ms 3D ¹H-MRSI hybrid sequences compared (A) 1D Hadamard spectroscopic imaging (HSI). A cascade of two selective 3-ms long, 90° Shinnar–Le Roux (SLR) RF pulses of B₁^max = 18 μT,^19,27,28 encoding a second Hadamard order (two 8-mm thick slices 2 mm apart to minimize crosstalk) in the volume of interest (VOI)’s short direction.²⁹ (B) 1D chemical shift imaging (CSI) over a FOV yielding the same slices’ thickness. The VOI is excited with one 3-ms selective SLR 90° RF pulse (half the gradient strength of (A). (C) Common to both: 16 ms adiabatic inversion (B₁^max = 18 μT) for lipids suppression; B₁₊ and T₁ insensitive water-suppression: six 12-ms, 70-Hz BW Gaussians (τ_i = 34.6, 29.7, 40.0, 29.5, 15.3 and 15.0 ms, flip angles indicated). A 246642 binominal 180° (1.9-ms interpulses delay, 300 Hz BW) to create a spin-echo and additional water and lipids rejection.³⁰ Finally, 320 ms, 1250-Hz BW 20 × 20 2D-RSI (4.7 mTm⁻¹ peak gradient, 40 mTm⁻¹ ms⁻¹ slew, 20 × 20 cm FOV) in the other two directions. Note: (i) peak RF and gradient amplitudes are met by all scanners; (ii) gray gradients are crushers, black are selection or refocusing; (iii) the MR signal in (a) (solid line) is a superposition of echoes from the first and second (dotted, dashed lines) HSI 90°s; and (iv) the strong, 8.4-mTm⁻¹ HSI gradients limit the Lac-to-Cho chemical shift displacement error (CSDE) (approximately 2 ppm, 250 Hz at 3 T) to less than 6% slice width

Common to both sequences in Figure 1C, are: (i) a 16-ms hyperbolic secant RF pulse³¹ for global lipid suppression,³² yielding less than 98% inversion with 0.8 kHz ± 20% peak B₁₊; (ii) a 180-ms water-suppression train of six 12-ms, 70-Hz BW Gaussian RF pulses that performs well for water T₁s of less than 500 ms at up to ±15% B₁₊ variations, regardless if the magnetization is inverted or not; (iii) a 136631 180° RF pulse (1.9 ms delays, 300 Hz BW) to produce an echo with extra water and lipids rejection.³⁰ Because it is not spatially selective, it suffers no CSDE; finally, (iv) 2D-RSI: 4.7 mTm⁻¹ peak, 40 mTm⁻¹ ms⁻¹ slew rate, encoding a 20 × 20 matrix, over a 20 × 20 cm² FOV in the left–right (LR) × anterior–posterior (AP) directions for 320 ms, at 1.25 kHz BW.^8,23 With the circular trajectory that repeats with a spectral dwell time of 800 μs, the RSI suffers minimal in-plane spatial distortion due to susceptibility gradients, despite the long sampling window.

The ¹H-MRSI data were processed with in-house software (MATLAB, MathWorks Inc., Natick, MA, USA). The 2D-RSI data were reconstructed independently for each coil using k_x-k_y-time regridding on a twofold oversampled grid, with a Kaiser–Bessel kernel (W = 4 window) and postgridding density compensation³³ and zero-filled to 32 × 32. For each voxel and coil, the water sensitivity map was used for phase correction and amplitude weighting for coil combination.¹³

2.2 |. Acquisitions in a phantom

SNR, SNR_VT and the localization performance of HSI were compared with CSI. Both MRSI sequences were run in the eight-slice cylinder (125 × 80 mm² diameter × thickness, 0.5 mm walls) shown in Figure 2A. Its four central slices each contained a different 0.1 M metabolite, yielding a singlet at a distinct chemical shift: methanol (Met), sodium acetate (NaAc), tert-butanol (t-Bu) and tri-methyl-silyl propanoic acid (TMSP) at 3.4, 1.9, 1.2 and 0.0 ppm, as shown in Figure 2B. Two slices on each side contained 0.1 M NaCl to load the coil and lower air-water susceptibility. Both CSI and HSI of Figure 1 were used to acquire 3D data, with HSI exciting the two innermost NaAc and t-Bu compartments (slice thickness = 0.9 cm, gap = 0.2 cm), and CSI using a slab-selective pulse to excite the same two compartments (Figure 2A) but encode all eight phantom compartments (FOV = 8 cm, slice thickness = 1 cm, VOI = 2 cm). The spectra from each acquisition are shown in Figure 2C,D.

FIGURE 2.

(A) Sagittal T1-weighted MP-RAGE MRI of the phantom, showing the partitions containing 100 mM methanol (Met), sodium acetate (NaAc), tert-butanol (t-Bu) and tri-methyl-silyl propanoic acid (TMSP) and the 2-cm (IS) volume of interest (VOI) (yellow frame). Two partitions on each side are filled with 100 mM NaCl to reduce air-water susceptibility and load the coil. (B) ¹H-MRS from a 1 × 1 × 4 cm³ (LR × AP × IS) voxel showing the four singlets from the central partitions, at 3.4, 1.9, 1.2 and 0.0 ppm. Note that because all the metabolites are at an equal concentration of 0.1 M, the TMSP and t-Bu, which have three −CH₃ functional groups, have resonances that are threefold larger than the NaAc singlet with only one methyl group. Spectra from a voxels column along the IS center axis, encoded with (C) chemical shift imaging (CSI) and (D) Hadamard spectroscopic imaging (HSI). Note that: (i) there is substantially worse bleed-in and -out of the VOI (red and black arrows) in the former vs. a near absence in the latter; and (ii) there is chemical shift displacement error (CSDE) bleed from slice #3 (Met) outside into #4 (NaAc) inside the VOI (dotted ellipse and arrow) for CSI, which is absent in HSI due to an approximately twofold stronger gradient of 8.4 vs. 3.8 mTm⁻¹

For each sequence, the total signal s_jk of the jth metabolite (j = 1, …,4) in the kth compartment (k = 1, …,8) was quantified via direct integration of the phased, real part. These integrals were used to assess the bleed of each metabolite, defined as the amount of signal observed outside that slice, divided by that metabolite’s total signal:

$${\text{f}}_j=\frac{{\sum }_{\text{k}\ne {\text{k}}^{(j)}}{\text{s}}_{jk}}{{\sum }_{\text{k}}{\text{s}}_{jk}}(\text{j}=1,\dots ,4),$$ 1

where k^(j) is the compartment of the ith metabolite. This was repeated for each in-plane voxel within the phantom and the results were averaged to yield the mean bleed.

The SNR of each metabolite was calculated as its peak height in its appropriate compartment, divided by the RMS of the noise.³⁴ The SNR_VT was calculated by further normalizing the raw SNR by the nominal voxel volume and square root of the acquisition time.

2.3 |. Simulations

The phantom results were compared and contrasted with 1D simulations, performed using in-house MATLAB software that numerically solves the Bloch equations with relaxation T₁/T₂ = 1 s/0.25 s (the average values of these parameters reported for the major brain metabolites^35,36). Two sets of simulations were run. First, the PSF was calculated for both methods. For HSI, two excitations were played out, as performed experimentally, and the resulting transverse magnetization was then Hadamard-transformed to yield the transverse magnetization of each slice, $M_{k,xy}^{(HSI)}(x)$. Note that the magnetization reconstructed for each excitation might slightly differ due to temporal cascading of the excitation pulses and the non-negligible T₁ and T₂ effects. For CSI, eight excitations were played out, which included the slab-selective pulse followed by phase encoding. The resulting transverse magnetization profiles were then transformed using a discrete Fourier transform to yield the transverse magnetization of each slice, $M_{k,xy}^{(CSI)}(x)$(k = 1, …,8). These were used to calculate the bleed of each voxel, for each method:

$$b_k^{(\text{m})}=\frac{{\int }_{x\notin (\text{slice }k)}{M}_{k,xy}dx}{{\int }_{-\infty }^{\infty }{M}_{k,xy}dx}\left(m=CSI,HSI\right)$$ 2

While examining the magnetization profiles yields interesting insights about each method’s spatial response, this does not account for CSDE effects or sample geometry. To include those, a second set of 1D simulations generated a numerical phantom with the same composition as the physical phantom, using the same pulses, spatial encoding and reconstruction of either sequence along the slice direction. This provided a more direct comparison between simulation and experiment, and enabled simulating the spectrum from each compartment and calculating the bleed in an entirely analogous manner to the phantom experiments (Equation 1). Note that all bleed differences, in the phantom, simulation and in vivo, are attributed to the localization method, CSI or HSI, along the VOI. This is because both sequences use the same in-plane RSI localization (see Figure 1), that is, they share the same shape PSF in these two directions.

2.4 |. In vivo experiments

Five young healthy adult volunteers (four males, one female) aged 19–39 (29.8 ± 7.3 mean ± standard deviation [SD]) years were prospectively recruited. Their health status was established by negative answers to excluding neurological conditions before the scan and unremarkable MRI, as determined by a neuroradiologist afterwards. All gave IRB-approved written informed consent.

Experiments were performed in a 3-T scanner (Prisma, Siemens AG, Erlangen, Germany) with a 20-channel receive-only head-array. The body coil, capable of peak B₁₊ of 24 μT, provided RF excitation. We limited B₁₊ to ≤18 μT (γB₁₊ = 0.8 kHz) and gradients to ≤18 mTm⁻¹, to conform with the capabilities of any clinical scanner. The phantom (or subject) was placed into the magnet head-first supine and scanned with sagittal 3D T₁-weighted MP-RAGE: TE/TI/TR = 2.6/800/1360 ms, 256 × 256 matrix, 256 × 256 mm² FOV, and 160 1-mm thick slices. For VOI guidance these were reformatted into axial, coronal and sagittal stacks angled along the long axis of the bilateral hippocampus. Our noniterative, B₀ map-based, BOLERO (B₀ loop-encoded readout) auto-shim software optimized B₀ homogeneity over the hippocampi.³⁷ Once completed, each subject was scanned using both CSI and HSI protocols, as described in detail above.

The SNRs of the metabolites in each voxel, in both phantom and human subjects, were defined as peak height (after the baseline estimate, also produced by the SiTTools FITT spectral modeling package,³⁸ was subtracted from the raw data) divided by RMS noise.³⁴ In the subjects, axial MP-RAGE MRI were transformed to MNI space with SPM12,³⁸ and the bilateral hippocampi outlined using the Harvard–Oxford subcortical atlas.³⁹ SPM12 then inverse-transformed the hippocampi outlines back to the subject’s space. SNR_VT of the major metabolites (choline [Cho], creatine [Cr], N-acetyl-aspartate [NAA]) was then assessed and averaged for all voxels whose center fell inside the hippocampal mask, and whose Cramer–Rao lower bound (CRLB) was less than 20, as estimated using Soher et al.’s parametric spectral modeling and least-squares optimization SiTTools FITT software package,⁴⁰ with an aspartate, Cho, Cr, glutamate, glutamine, myo-Inositol, NAA and taurine metabolites’ basis set. Note that because NAA-glutamate (NAAG) resonates at just 0.036 ppm (4 Hz at 3 T) from the NAA, their approximately 8 Hz linewidths make it impractical to constrain the fit to resolve them. Consequently, the fitted NAA values here also comprise a small contribution of NAAG weighting.

3 |. RESULTS

3.1 |. Phantom

Our auto-shim software optimized B₀ homogeneity over the phantom or hippocampi³⁷ to 5 Hz water linewidth in the phantom. Spectra from slices encoded along the inferior–superior axis for both methods, acquired back-to-back with the same acquisition parameters as used in the simulations, are shown in Figure 2C,D. Because each slice contains a single spectrally nonoverlapping metabolite, any peaks at different chemical shifts are bleed-in and their chemical shift identifies their spatial source. Figure 3 shows in-plane spectroscopic imaging results for the two slices and two metabolites (t-Bu, NaAc) within the VOI.

FIGURE 3.

Quantitative comparison of the volume of interest (VOI) (slices #4 and #5, Figure 2A) interslice bleeds between the chemical shift imaging (CSI) (A, B) and Hadamard spectroscopic imaging (HSI) (C, D), for the sodium acetate (NaAc) (A, C) and tert-butanol (t-Bu) (B, D). Bleed is quantified at each in-plane voxel, as described by Equation 1. Histograms of voxel bleeds from the circular regions of interest (ROIs) (white outline: defined to exclude air-water susceptibility artifacts near the edges) on the bleed maps for the two methods for (E) NaAc and (F) t-Bu. Dashed vertical lines in each histogram indicate median bleed percentage. Note (i) the dramatic nearly eightfold worse bleed for CSI compared with the approximately 5% median of HSI on the images and histograms; and (ii) the slightly larger bleed for t-Bu acquired with HSI due to its larger chemical shift displacement error (CSDE)²⁹

3.1.1 |. Signal bleed

Experimentally, the Met (slice #3), NaAc (slice #4), t-BU (slice #5) and TMSP (slice #6) have nearest-neighbor chemical shift offsets of 1.5, 0.7 and 1.2 ppm, respectively, leading to 0.50 mm (Met to NaAc), 0.24 mm (NaAc to t-BU) and 0.40 mm (t-BU to TMSP) CSDEs under the HSI-selective gradients. Consequently, no bleed is seen between the two VOI partitions (#4 and #5, Figure 2D), nor observed from external partitions (slices #3 and #6) due to the small CSDE and localized slice profile. The bleed extent (Equation 1) is quantified for each voxel in-plane for both NaAc and t-Bu and is shown in Figure 3C,D. The distribution of in-plane bleeds (Figure 3E,F) exhibits a median of 5%–8% for HSI. Note that this distribution is almost entirely due to interslice bleeds because both sequences otherwise use exactly the same RSI for in-plane localization (whose PSF is similar to rectilinear k-space sampling, as shown by Schirda et al.).

For CSI under × twofold weaker gradient, the CSDEs double to 1.00, 0.54 and 0.92 mm. Together with the broader slice profile, this leads to a Met (slice #3) leak into NaAc (slice #4) in Figure 2C, and between slices #4 and #5. Figure 3A,B shows the in-plane bleed for both metabolites. The distribution of in-plane bleeds (Figure 3E,F) exhibits a median of 51%–60% for CSI, much worse than HSI.

3.1.2 |. SNR

The SNRs of the metabolites in each voxel in the human subjects were defined as peak height (after the baseline estimate, also produced by the SiTTools FITT spectral modeling package,⁴⁰ was subtracted from the raw data) divided by RMS noise.³⁴ The dramatic bleed differences between CSI and HSI should be reflected in their metabolites’ SNR_VT. Non-normalized SNRs over all voxels within the phantom were 1209 ± 66 and 452 ± 57 for t-Bu and NaAc in the HSI vs. 1472 ± 98 and 450 ± 34 for CSI. However, normalizing for voxel volume and acquisition time differences (thinner HSI slices [9 vs. 10 mm] and acquisition time shorter [5 vs. 8 min] compared with CSI) yields (10/9) × (8/5)^½ = 1.4, that is, a substantial SNR_VT advantage of approximately 15% for t-Bu and approximately 40% for NaAc, over CSI, consistent with the bleed loss. As shown by the simulations below, the smaller SNR_VT increase for t-Bu reflects its greater CSDE.

3.2 |. Simulations

The PSF for HSI, based on the geometry of Figure 2A, is shown in Figure 4A. It is compared with the PSF of a 1D CSI with eight phase-encoding steps in Figure 4B. Based solely on the PSF, the fractional bleed for the phantom in Figure 2, as estimated using Equation 2 and averaged over both voxels, is 0.3% for HSI vs. 26% for CSI.

FIGURE 4.

Point spread function (PSF) bleeds, chemical shift displacement error (CSDE) and slice profile simulations: Hadamard spectroscopic imaging (HSI) and chemical shift imaging (CSI) comparison. (A) HSI (black solid lines) for two 0.9-cm thick slices, 0.2 cm gap, 8.4 mTm⁻¹ gradient, 3 ms 90° Shinnar–Le Roux (SLR) of peak γB₁₊ = 0.8 kHz, defining a 2-cm volume of interest (VOI) (vertical dashed gray lines). Percentages indicate bleed (Equation 1): yellow zone within, red outside. (B) CSI over 8-cm field of view (FOV) (1 cm nominal width) slice profiles (solid gray lines) for the two slices inside the 2-cm VOI (solid black line) excited with the same SLR and B₁₊ as (A), but with a ×2 weaker, 3.8 mTm⁻¹, gradient. Note: (i) sharp HSI slice profiles, minimizing VOI bleed-in and -out (yellow and red zones) for maximal signal-to-noise ratio (SNR) and SNR per unit volume and unit time (SNR_VT)²¹; (ii) extension of the CSI PSF beyond the VOI, bleeding its signals out for SNR loss; (iii) sampling of the CSI PSFs of the slab edges outside the VOI (red zones); and (iv) CSI’s VOI interslice bleed (yellow zone) reducing localization accuracy. None of (ii)–(iv) is suffered by HSI. (C) HSI cascade first radio frequency (RF) pulse of Figure 2A (black lines). Colored zones represent the metabolites’ partition and plastic walls. Bleed % values are defined by Equation 1. Note minimal bleeds, due to sharp slice profiles falling on partition walls. (D) Same for the second HSI pulse in the cascade. Note the asymmetrical (minimal) bleed due to the sodium acetate (NaAc) CSDE, as shown by the spin packets |M_XY| (red circles). (E) Real part of the simulated spectra from the VOI (blue and brown lines). The four colored regions on the frequency scale show the peak integration windows used to estimate a signal’s area within its partition (indicated by the spectral line color) and in the other (outside) partitions used to obtain the % bleed as described by Equation 1. (F) Same as (C, D) for CSI. Same RF pulse, at half the gradient strength: Note: (i) broader slice profiles (black line), extending into nearby (methanol [Met] and tri-methyl-silyl propanoic acid [TMSP]) partitions; (ii) bleed is exacerbated by ×2 CSDE, depicted by spin packets in their partitions. (G) Same as (E). Note the dramatically larger Met and TMSP signals in the integration intervals, representing slab profiles and CSDE in (F) and PSF in (B). Finally, note that the simulated TMSP bleed is not observed in the phantom (Figure 2C) due to suppression by the sequence’s $1\overline{3}6\overline{6}3\overline{1}$ refocusing pulse (Figure 2B)

The simulated bleeds obtained by applying Equation 1 to the simulated spectra is a negligible less than 4% for HSI. It exhibits no bleed from outside the VOI (Figure 4C). By contrast, CSI suffers approximately 37% bleed from the selected (VOI) slices and approximately 62% from slices outside, as shown in Figure 4D,F. The superior performance of HSI is due to its reduced CSDE and substantially better slice profile and PSF. The disparity between simulated and measured VOI bleeds is due to each voxel’s signal comprising an integrated PSF shifted in space; thus, different bleeds result from different MRSI grid placements.

3.3 |. In vivo comparisons

MRSI from the 5 × 7 cm² matrix in Figure 5A, containing the bilateral hippocampi, are shown in Figure 5D,E. Note the absence of linear phase shift in the HSI spectra, despite its two echoes coming at different times (see Figure 1A). Because HSI reconstruction separates the two echoes, the later one (from the first 90°) can be shifted temporally so that both coincide. Also, note the spectra quality in terms of SNR and resolution: 8 ± 2 Hz linewidth for the HSI; 10 ± 2 Hz for CSI, as shown in Figure 5D,E in the VOI, and from the entire slice in Figure 5F,F’, also obtained with the SiTTools FITT package.⁴⁰ These reflect the performance of the shimming routine that yielded a less than 20 Hz water linewidth over this region, as shown in Figure 5G,G’, in under 5 min.

FIGURE 5.

(A–C) Axial sagittal and coronal MP-RAGE MRI of a 19-year-old subject’s brain superimposed with the positions of the volume of interest (VOI), field of view (FOV) (thick red and dashed white frames) and slices over both hippocampi and the left hippocampus mask (black trace). (D, E) Real part of the spectra matrix from a 7 × 5 cm volume (red frame over Figure 5A–C) from the inferior slice (marked by the red arrow on Figure 5B,C) covering both hippocampi and obtained with Hadamard spectroscopic imaging (HSI) from Figure 2A (D) and from the same slice obtained with chemical shift imaging (CSI) of Figure 2B (E). The black traces on (D) and (E) reflect the hippocampus mask on (A), and the black squares on the bilateral hippocampi mark the spectra expanded for greater detail, together with the fit, residual and baseline performance, in Figure 6A,A’. (F, F’ and G, G’) Linewidth (T₂*) and field (ΔB₀) distribution maps from the entire slice (A), demonstrating the performance of the BOLERO auto-shim software. Note (i) the lack of lipids contamination in the spectra; (ii) signal-to-noise ratio (SNR); (iii) spectral resolution: (8 ± 2 Hz linewidth for the HSI; and 10 ± 2 Hz for CSI, especially at the hippocampi anterior); and (iv) performance of the auto-shim procedure in this difficult temporal lobe region

Especially noteworthy is the spectra quality in the anterior hippocampus (Figure 5D,E, expanded in Figure 6A,A’), which usually suffers more B₀ susceptibility distortions from the air-filled sinuses below. This advantage is due, in part, to shimming a relatively thin, small axial VOI, which would otherwise suffer if larger brain volume is included in the shim process. The absence of strong lipid signals from the optic chiasm just below the VOI is due to the sharp slab profiles of the CSI and especially the HSI. Also, note the sometime undulating baseline of some of the spectra (e.g. Figure 6A,A’). This is likely due to incomplete suppression of the macromolecules’ resonances with a single TI, due to their 200–300 ms T₁s,⁴¹ in a short TE acquisition. These incompletely suppressed unwanted signals most likely get spread around the slices planed by the RSI PSF. It is noteworthy that while these spurious signals could be removed by a four-step phase-cycling scheme, which would also improve the SNR by approximately 40%, this would come at the cost of double the measurement time and consequently an increased risk of motion errors, neither of which are welcome in clinical applications, especially because these baseline undulations can be successfully removed in postprocessing by the SiTTools FITT package, minimizing their effect on both the major and minor metabolites quantification, as shown for NAA and glutamate (Glu) in Figure 6B,B’ and 6C,C’. Together with shimming, ¹H-MRSI and water reference acquisition, the procedures require less than 13 and 10 min, respectively, for the CSI and HSI that it is practical to incorporate into clinical studies of the temporal lobes.

FIGURE 6.

(A, A’) Real part of the Hadamard spectroscopic imaging (HSI) (left) and chemical shift imaging (CSI) (right) spectra from the bilateral (left hippocampus (HC), right hippocampus (HC)) four spectra from the frames on Figure 5D,E expanded for greater detail. Each spectrum frame contains the raw spectrum (black lines), superimposed with their spectral fits (baseline + metabolites, thick gray lines). Below each spectrum is the residual (raw spectrum – fit) and below that is the (raw spectrum – baseline estimate) to demonstrate the performance of the baseline estimation. All were obtained with the SiTTools FITT software package.⁴⁰ Note (i) the excellent fit fidelity of these TE = 40 ms spectra, reflected by the vanishing residual; (ii) The excellent flat baseline of the raw-baseline spectrum, used for the metabolites’ signal-to-noise ratio (SNR) estimates reported; (iii) the left–right spectral similarity; and finally (iv) linewidth at the hippocampi, reflecting BOLERO’s regional shim performance. (B, C and B’, C’) Major (N-acetyl-aspartate [NAA]) and minor glutamate (Glu) metabolic maps obtained with the SiTTools FITT package over the entire slice in Figure 5A (duplicated above each maps for more detailed anatomic reference), for HSI and CSI encoding. Note the differences between the maps between the two methods, which are largely due to: (i) concentrations shown are not corrected for the thinner HSI slices, nor to their shorter acquisition time. Consequently, metabolites’ maps acquired with HSI are not as intense in color; (ii) voxels with metabolite’s Cramer–Rao lower bounds (CRLBs) of less than 20% are rejected from the analyses, so the lower raw SNR of the HSI acquisition leads to more voxels getting rejected (particularly for Glu). (E, E and D’, E’) Corresponding CRLB maps of the NAA and Glu metabolic maps above, also obtained with the SiTTools FITT package. Note the distributions of these CRLBs in voxels within the brain: NAA: 10.4 ± 5.0 (12.0 ± 4.6); Glu: 11.2 ± 4.8 (14.1 ± 3.6), acquired with CSI (HSI)

3.3.1 |. SNR and SNR_VT

The metabolite SNR of the hippocampi voxels, averaged across all volunteers, are within ±5% for both methods: 18 ± 4 (25 ± 3), 14 ± 2 (19 ± 2) and 11 ± 3 (14 ± 2) for NAA, Cr and Cho for CSI (HSI), respectively. Their corresponding whole-slice CRLBs are 10.4 ± 5.0 (12.0 ± 4.6), 8.6 ± 4.9 (11.7 ± 4.7) and 10.3 ± 5.0 (12.7 ± 4.6), respectively, as shown in Figure 6D,E,D’,E’. Scaling the HSI SNRs for thinner slices and shorter acquisition, (10/9) × (8/5)^½ = 1.4, yields a statistically significant (paired t-test) 6% SNR_VT advantage for all singlets, substantially less compared with the simulation and phantom results. We attribute this difference to the fact that brain metabolic content is similar everywhere, therefore: (i) within the VOI, slice bleed-in is similar to bleed-out, unlike in the phantom; and (ii) tissue outside the VOI, excited by the imperfect slab profile, contributes approximately 20% to the VOI slices, as shown in Figure 4F.

4 |. DISCUSSION

Phase encoding has gained universal prominence due to four attractive attributes: (i) simple implementation; (ii) no specific absorption rate (SAR) contribution; (iii) immunity to CSDE; and (iv) an ability to accelerate k-space coverage.^6–8 It is not surprising, therefore, that many variants have been devised since their early 1980s inception,^1,2 including rectilinear echo planar ¹H-MRSI^10,42,43 and nonuniform k-space schemes.^3,5,8,44 However, CSI’s two main shortcomings remain: first, signal bleed due to PSF and intravoxel dephasing is inherent.^16–18 Second, VOIs defined with selective RF pulse(s), for example, PRESS or STEAM, incur CSDE at their edges²⁰ that can exceed a full voxel for larger VOIs, higher B₀s and larger chemical shifts.

Both issues can be mitigated by using RF encoding. For the large matrix sizes often used in-plane, these approaches become impractical due to SAR and relaxation effects.^20,45 However, ¹H-MRSI often employs a smaller number of voxels in one direction, rendering it an ideal candidate for HSI,⁴⁶ motivating hybrid approaches, as demonstrated here. The cascaded excitation (two slices per TR here) avoids the nonlinear RF interactions incurred when HSI pulses are superposed, that is, applied simultaneously,²² and is more efficient than multislice acquisition schemes, for example, as first proposed by Duyn and Moonen.⁴⁷ This is due to the global (whole head) inversion-recovery lipids suppression (and global frequency elective 180° refocusing) applied to facilitate ¹H-MRSI to the brain’s edges. These force a long, less than 1.5-s TR, making slice-interleaving schemes slow (i.e. time-inefficient), whereas cascading produces N (here two) slices every TR and thus “averages” each slice N time (here twice) more for a substantial, 40% SNR_VT advantage.⁴⁸

Moreover, the add-subtract nature of the Hadamard transform eliminates, by destructive interference, out-of-VOI signals (contamination) created by outside spins excited by the RF pulse imperfections,⁴⁹ a mechanism not available to multislice variants. It is noteworthy that while the lipids and macromolecules suppression elements (IR + frequency-selective 180°) in the sequence achieves that task well, a ×4 1D cascaded HSI hybrid with 2D CSI sequence without them was recently demonstrated for FID direct detection at 7 T by Hangel et al., which they call FID-MRSI.⁵⁰ In this short TR = 300 ms, the lipids’ signal was suppressed using a combination of very high spatial resolution (<2 mm in-plane, 23 μL voxels) in the acquisition, together with L2-regularization in postprocessing. It is noteworthy that the spatial resolution needed to achieve sufficient lipids suppression mandated a long, approximately 17-min acquisition compared with approximately 5 min here. Finally, because each of the cascade 90°s is pulsed separately, and their 10% offset slices have no spatial overlap, as shown in Figure 3A,C,D, the hybrid spectra are no different from simple TE = 40 ms, 90°–180° SE with respect to J-coupled metabolites signals and the macromolecular baseline.

Mitigating the CSDE in the slice-selection direction(s) for CSI with stronger selective gradients is impractical because B₁₊ of body coils at 3 T is typically less than 1 kHz. Consequently, the slab profiles shown here are close to optimal. Because the ¹H-MRSI voxel size is determined by the available time, slices much thinner than approximately 1 cm are unlikely in approximately 10-min acquisition. Consequently, even the thinnest slab and minimal 3D acquisition (two slices) already suffers approximately 20% bleed from a nonideal RF slab-profile. While this and the CSDE progressively worsen for thicker slabs in CSI, HSI remains unaffected, because its single slice-selecting gradients are independent of the VOI width, retaining near-ideal profiles, less than 4% bleed and very small, less than 5%, CSDE.

An additional “free” advantage of the reduced interslice and external “bleed in” of HSI is improved spectral resolution. Specifically, note that in our in vivo results, the voxel linewidth is 25% narrower for the HSI than CSI acquisition, averaged across thousands of voxels, despite both having the same global VOI shim value in every case. This we attribute to “bleed in” for peaks from voxels outside that slice that may resonate at slightly different frequencies, reflecting their local susceptibility. (Note that this effects only the in vivo spectra, where all slices have similar metabolites, but not in the phantom where each slice has a single unique constituent.) The HSI acquisition, which suffers negligible such bleed, consequently enjoys narrower lines (i.e. improved spectral resolution for the same global shim). In addition, the shaper HSI slab profile, together with its well-controlled slices’ PSF, combined to reduce: (i) extraneous lipids contamination into the VOI; and (ii) the bleed of unwanted signal between the VOI slices, compared with CSI. This advantage can lead to a flatter baseline, hence better spectral quantification.

While the presence of bleed contamination and its effect on SNR is known, the comparison of the phantom and in vivo data is of importance for recognizing its magnitude in the CSI. This is not always evident due to the fact that in the brain, interslice bleeds are very likely to be of comparable magnitude and spectral content, due to similar VOI tissue composition. Thus, even though the in vivo SNR_VT is similar between the two methods, with the acquisition parameters used here, the CSI slices are spatially contaminated to approximately 50% with inappropriate signals arising from regions both internal and external to the selected VOI. At these levels, the likelihood of distortion of results is of concern. Because spatial localization of spectra is MRSI’s raison d’être, it has already been shown that both these CSI shortcomings can be nearly eliminated using the high-BW RF pulses of HSI.^19,20 This advantage, however, comes at several costs, the first of which is the increased SAR (from multiple RF pulses). Second is a progressive T₂ weighting for earlier pulses in longer cascades (although the latter amount to a relatively small <10 ms for two or four slices). Third, because Hadamard reconstruction relies on linear combinations of the acquired FIDs, it is more susceptible to patient motion, which increases for higher order encodings.²¹ While this is a limitation in applications where patient motion may be an issue, such as epilepsy,⁵¹ Alzheimer’s disease, attention deficit hyperactivity disorder and Parkinson’s disease, the relatively short, approximately 5-min acquisition reduces the sequence exposure to subject movements. Nevertheless, HSI hybrids with CSI⁴⁸ can be conceived to leverage each method’s strengths and minimize the impact of its weaknesses.

Given nearly 4 decades of MRSI, it is not surprising that remedies have been offered to reduce the extent of CSI voxel bleed. These fall into two categories: (i) prospective tailoring of the RF excitation or phase-encoding strategy^52,53; and (ii) retrospective k-space filtering.^54,55 The first requires careful, time-consuming preplanning for VOI placement and size, as well as complicated reconstruction. The second modifies the PSF shape, and therefore invariably increases the voxels’ volume (i.e. reduces the spatial resolution) and consequently may confound metabolic quantification. The nearly ideal PSF of HSI, in comparison, requires none of these and can be improved further simply with more B₁₊ (e.g. with parallel transmission⁵⁶).

Abbreviations used:

BW

bandwidth

CSDE

chemical shift displacement error

CSI

chemical shift imaging

Cho

choline

CRLB

Cramer–Rao lower bound

Cr

creatine

FOV

field of view

HSI

Hadamard spectroscopic imaging

Met

methanol

NAA

N-acetyl-aspartate

NaAc

sodium acetate

PSF

point spread function

¹H-MRSI

proton MR spectroscopic imaging

RF

radio frequency

RSI

rosette spectroscopic imaging

SAR

specific absorption rate

SLR

Shinnar–Le Roux

SNR

signal-to-noise ratio

SNR_VT

signal-to-noise ratio per unit volume and unit time

t-Bu

tert-butanol

TMSP

tri-methyl-silyl propanoic acid

VOI

volume of interest