Compartment‐based reconstruction of 3D acquisition‐weighted 31P cardiac magnetic resonance spectroscopic imaging at 7 T: A reproducibility study

Abstract Even at 7 T, cardiac 31P magnetic resonance spectroscopic imaging (MRSI) is fundamentally limited by low signal‐to‐noise ratio (SNR), leading to long scan times and poor temporal and spatial resolutions. Compartment‐based reconstruction algorithms such as magnetic resonance spectroscopy with linear algebraic modeling (SLAM) and spectral localization by imaging (SLIM) may improve SNR or reduce scan time without changes to acquisition. Here, we compare the repeatability and SNR performance of these compartment‐based methods, applied to three different acquisition schemes at 7 T. Twelve healthy volunteers were scanned twice. Each scan session consisted of a 6.5‐min 3D acquisition‐weighted (AW) cardiac 31P phase encode‐based MRSI acquisition and two 6.5‐min truncated k‐space acquisitions with increased averaging (4 × 4 × 4 central k‐space phase encodes and fractional SLAM [fSLAM] optimized k‐space phase encodes). Spectra were reconstructed using (i) AW Fourier reconstruction; (ii) AW SLAM; (iii) AW SLIM; (iv) 4 × 4 × 4 SLAM; (v) 4 × 4 × 4 SLIM; and (vi) fSLAM acquisition–reconstruction combinations. The phosphocreatine‐to‐adenosine triphosphate (PCr/ATP) ratio, the PCr SNR, and spatial response functions were computed, in addition to coefficients of reproducibility and variability. Using the compartment‐based reconstruction algorithms with the AW 31P acquisition resulted in a significant increase in SNR compared with previously published Fourier‐based MRSI reconstruction methods while maintaining the measured PCr/ATP ratio and improving interscan reproducibility. The alternative acquisition strategies with truncated k‐space performed no better than the common AW approach. Compartment‐based spectroscopy approaches provide an attractive reconstruction method for cardiac 31P spectroscopy at 7 T, improving reproducibility and SNR without the need for a dedicated k‐space sampling strategy.

adenosine triphosphate (PCr/ATP) ratio, the PCr SNR, and spatial response functions were computed, in addition to coefficients of reproducibility and variability. Using the compartment-based reconstruction algorithms with the AW 31 P acquisition resulted in a significant increase in SNR compared with previously published Fourier-based MRSI reconstruction methods while maintaining the measured PCr/ATP ratio and improving interscan reproducibility. The alternative acquisition strategies with truncated k-space performed no better than the common AW approach. Compartmentbased spectroscopy approaches provide an attractive reconstruction method for cardiac 31 P spectroscopy at 7 T, improving reproducibility and SNR without the need for a dedicated k-space sampling strategy.

| INTRODUCTION
Magnetic resonance spectroscopy (MRS) of high-energy phosphorus metabolites provides a unique insight into cardiac metabolism. 1 Of particular interest is the phosphocreatine-to-adenosine triphosphate metabolite ratio (PCr/ATP), which differs between healthy and diseased hearts. 2,3 Localization strategies include selective excitation of only the heart, for example, through coil design or sequences such as ISIS, 4 PRESS, 5 DRESS,6 or STEAM, 7 or most commonly for human cardiac studies, Fourier transform-based magnetic resonance spectroscopic imaging (FT-MRS) reconstruction of phase-encoded data. 8 Unfortunately, all 31 P MRS techniques are limited by the intrinsically low 31 P concentrations found in vivo, which makes it difficult to achieve adequate signal-to-noise ratio (SNR). Localization to small volumes can degrade the achievable SNR further, because the smaller volumes contain fewer 31 P nuclei for a given metabolite concentration. Additionally, in the case of single voxel techniques, there is the potential for a mismatch between the shape of the anatomical region of interest and the cuboidal voxel, and for Fourier transform (FT)-based methods, SNR losses can be associated with the shape of the spatial response function (SRF). Low SNR may reduce the reproducibility of 31 P MRS, because it negatively impacts the accuracy of fitting a model, in order to determine metabolite content, to the acquired spectra. Increasing the field strength of the magnet from 1.5 or 3 to 7 T can increase SNR and improve reproducibility 9,10 ; however, even at 7 T, 31 P spectroscopy is still an inherently SNRlimited technique.
Translation of 31 P magnetic resonance spectroscopic imaging (MRSI) techniques, from the development stage to clinical research, faces a number of challenges. Patient cohorts may produce significantly different results to healthy volunteers, because of physiological differences (age, heart rate, body mass index [BMI], etc.), making validation in patients essential; however, long scan times make it difficult to add experimental acquisitions to already long clinical research protocols. Therefore, improved reconstruction techniques, which could be applied to 31 P acquisitions already validated for use in clinical research, would be of great value as a means to improve SNR and repeatability without changing the clinical scan protocol, allowing the current reconstruction algorithm to serve as a fail-safe.
FT-based MRSI reconstruction 11 of acquisition-weighted (AW) data is commonly used at the higher field strengths of 3 and 7 T 8,10,12-14 for 31 P MRSI. However, alternative compartment-based reconstruction strategies 15-21 could potentially improve spectral SNR and PCr/ATP repeatability, without needing to change this already validated AW 31 P scan. These strategies, which separate out the 31 P signal into arbitrarily shaped user-defined compartments based on relatively homogeneous anatomical structures (e.g., heart, 22 liver, 23 skeletal muscle, 24 etc.), have been under development for several years. However, these techniques may be adversely affected by the increased inhomogeneities in the static (B 0 ) and RF (B 1 ) fields found at 7 T, compared with 1.5 or 3 T.
The magnetic resonance spectroscopy with linear algebraic modeling (SLAM) and spatial localization with optimal point spread function (SLOOP) compartment-based techniques have been validated for cardiac 31 P MRSI at the lower field strengths of 1.5 25 and 3 T 26,27 ; however, method-specific sequences were used. In this work, we compare the performance at 7 T of the compartment-based reconstruction algorithms used in these methods, the SLAM and spectral localization by imaging (SLIM) algorithms, respectively, against FT-based 31 P MRSI reconstruction when applied to AW, uniform, and per-volunteer optimized phase encode datasets.

| COMPARTMENT-BASED RECONSTRUCTION ALGORITHMS
The SLAM and SLIM algorithms formulate the problem of reconstructing spectra from C compartments as the solution to a system of M ≥ C equations that can be developed from the FT MRI signal equation. For one (spatial) dimension, the magnetic resonance spectroscopic imaging (MRSI) localization problem can be written as a continuous FT (CFT) 20 : where s k, t ð Þ is the received signal at k-space position k and time t and ρ x,f ð Þ is the spectral density at spatial position x (normalized by 1/field of view [FOV]) and frequency f. In the SLAM reconstruction, this equation is discretized to give a discrete FT, and a time-domain FT applied to give where M is the number of phase encodes and N the number of points recorded in the readout. As a further refinement to the SLAM technique, enabling the use of AW datasets, we introduce R m , the number of repeats of the mth phase encode here. Each row in the second matrix is weighted by R m to account for the effect on the signal magnitude of adding up the varying number of repeats for the different phase encodes.
For conventional SLAM implementations, R m ¼ 1.
The matrix equation can be written in short hand as S MÂN ¼ PE MÂM Â ρ MÂN , where S is the acquired data, PE is the phase encode operator defined by the second matrix in Equation (3), and ρ is the unencoded spectra for each spatial location x m .
The SLAM method reconstructs spectra from C ≤ M user-defined compartments by summing (with a normalization factor) the columns of the PE matrix, which correspond to x positions within each of the compartments, thereby reducing the dimensions of the reconstruction to Now that M ≥ C, and therefore PE MÂC is overdetermined, a reduced gradient set of M 0 ≥ C phase encodes may be used. To maximize SNR, the phase encodes are chosen from central k-space, and Equation (4) reduces to The SLIM reconstruction algorithm separates the MRSI equation (Equation 1) into C integrals, corresponding to C compartments, giving where ξ x, t ð Þ is defined as the time-domain FT of ρ x, f ð Þ.
If ξ x, t ð Þ depends only on t within each compartment (i.e., the compartment is homogeneous), we can substitute in the compartment signal ξ c t ð Þ for ξ x, t ð Þ, giving where M 0 is the number of phase encode steps recorded and g mc is defined in 1D as 15 Here, R m is the number of repeats of the m th phase-encoding step in order to account for acquisition weighting, k m is the k-space position of the m th phase-encoding step, and comp c is the spatial extent of the c th compartment. Because Equation (6) is at no point discretized, the CFT forms the basis of the SLIM equation, with the step between Equations (6) and (8) introducing a homogeneity condition for each of the compartment spectra.
This gives the matrix equation When solved by premultiplying both sides by H, where HG ¼ I, this gives the separated signal for each compartment, which can be Fourier transformed to give each compartment's spectra (ρ c ). Because this equation is overdetermined for M > C, the number of phase encodes (M) may be reduced to M 0 ≥ C as in SLAM.
The fractional SLAM (fSLAM) method improves the localization properties of SLAM by numerically optimizing the fractional k-space coordinates of the phase encodes 16,20 based on the compartmentalization mask. Because a localizer must be acquired to segment the compartments, this optimization must be done scanner-side at the start of the examination, increasing the time in the scanner.
An essential metric for understanding the SLAM and SLIM algorithms is the SRF, which gives the contribution of the metabolites at each point in the FOV to each compartment's spectra and allows the effect on the reconstruction of changing the acquired phase encodes to be quantified.
SRFs can be calculated for both the SLAM and SLIM reconstructions, by performing the reconstruction for each point in space, with the following formula: Here, β c x ð Þ is the c th compartment's SRF (with magnitude and phase components), h cm is the cm th element of H, and H is the pseudoinverse of PE or G for SLAM and SLIM, respectively. R m is the number of repeats of the m th phase encode, to allow for the effect of acquisition weighting to be calculated; for conventional SLAM and SLIM implementations, R m ¼ 1.
In a "perfect" theoretical experiment (with completely uniform B 0 , B 1 , and metabolite distribution within each compartment), the sum of a compartment's SRF over each of the other compartments multiplied by the signal distribution is exactly zero, providing perfect localization. In a real setting, however, the reconstructed spectra of a compartment, subject to B 0 , B 1 , and compartment inhomogeneity, can be calculated as where ρc is the reconstructed compartment spectrum for compartment c, ΔB 1 x ð Þ the relative B 1 distribution, ϕ x ð Þ the phase distribution, caused by B 0 inhomogeneity, and ρ x, f À Δf x ð Þ ð Þthe true, spatially inhomogeneous, spectra of the spins at spatial position x, with an added spatial component Δf x ð Þ to reflect B 0 inhomogeneity.
An SRF that assumes the "worst case" for spectral contamination from other compartments, caused by B0 inhomogeneity (when arg β c x ð Þ ½ ¼AEϕ x ð Þ), can be calculated by taking the magnitude of the SRF, which is used in the cost functions of SLOOP 16 and fSLAM 20 when optimizing the phase-encoding scheme. If the B 1 distribution is known, this magnitude SRF can then be weighted by B 1 to provide a guide to the likely signal contamination caused by field inhomogeneity.
The effect of inhomogeneity on the signal from within a compartment on that compartment's spectra is more difficult to quantify, with the major source of concern B 0 inhomogeneity. In the absence of B 0 inhomogeneity, the contribution of signal from within a compartment to the compartment's spectra will be a weighted sum of that signal. In the presence of severe B 0 inhomogeneity, however, there can also be signal cancellation, which could, if the metabolite distribution within the compartment is also highly heterogeneous, result in incorrect metabolite distributions, in addition to severely degraded SNR and poor spectral linewidths.
Numerous strategies to minimize the impact of inhomogeneity exist, with the optimal strategy depending largely on application and available SNR. These include the incorporation of a B 0 map for BSLIM, 18 per-volunteer phase encode optimization for fSLAM 20 and SLOOP, 16 compartment subdivision during reconstruction, 19 and higher resolution acquisitions.

| METHODS
Twelve healthy volunteers (six females/six males, age 29 ± 4 years, BMI 23 ± 3 kg/m 2 ) were recruited and gave informed consent for this study approved by our research ethics committee. The reproducibility of each method was quantified using a test-retest procedure, where each volunteer was scanned twice (each scan consisted of three 31 Table 1 and displayed graphically in Figure 1.

| SRF simulations
SRFs for a representative compartmentalization mask ( Figure 2A) were calculated for each method. The SRFs were calculated using the formulas described by Hu et al. 15 and Zhang et al. 20 and weighted with receive coil sensitivity calculated using the Biot-Savart law.

| Data acquisition
All acquisitions were undertaken on a Siemens MAGNETOM 7-T whole-body MRI scanner (Siemens Healthineers, Erlangen, Germany) with volunteers in the supine position, as previously described. 9 The RF coil used was a dual-tuned ( 31 P/ 1 H) transmit-receive quadrature, surface coil consisting of a 31 P coil (two 15-cm loops, with overlap decoupling), and a single 1 H loop (10 cm in diameter). 28 Two-chamber, four-chamber, shortaxis, and FT-MRS-matched 1 H images were acquired with a fast low-angle shot (FLASH) sequence. 29 The coil position was determined from images showing the location of four cod-liver oil capsules and one central phenylphosphonic acid fiducial embedded in the coil housing during imaging, and the transmit flip angle was calculated from a series of inversion recovery free induction decay acquisitions of the central fiducial, as previously described. 9 Three different 31 P acquisitions, corresponding to the AW, 4 Â 4 Â 4, and fSLAM phase-encoding schemes, were applied for the comparative studies. The 31 P acquisitions consisted of a 3D ultrashort echo time (UTE)-FT-MRSI sequence, 30 where the excitation was achieved with a shaped RF pulse, numerically optimized for excitation homogeneity 8 ; the excitation bandwidth was approximately 2 kHz. The excitation was placed with its center frequency at +266 Hz (towards the 2,3-diphosphoglycerate [2,3-DPG] resonance) relative to the PCr resonance. The repetition time (TR) was 1 s.
The phase-encoding schemes used in the acquisitions are summarized in   20 The acquisition order was fixed to allow the fSLAM coordinates to be calculated while the AW and 4 Â 4 Â 4 acquisitions ran and for scan planning to be accomplished using the AW sequence image overlay. The k-space coordinates used for each acquisition are detailed in Figure 3.
To enable calculation of the fSLAM coordinates while the AW and 4 Â 4 Â 4 phase-encoding scheme acquisitions completed, the degrees of freedom in the optimization were reduced to 12, after forming a 1D vector composed of the k-space coordinates in each orthogonal direction.

| Data Analysis
Reconstruction of the spectra was performed off-line using raw data (exported with the Siemens TWIX utility) for all methods. The SLIM and SLAM algorithms were implemented in the Matlab programming language according to the equations presented by Hu et al. 15 and Zhang et al., 20 respectively, in addition to a standard FT-MRS reconstruction. The SLAM and SLIM reconstructions included an additional weighting factor (R m ) F I G U R E 1 An overview of the data pathways for each acquisition/reconstruction combination considered from acquisition through reconstruction and fitting to results. AMARES, advanced method for accurate, robust, and efficient spectral fitting; AW, acquisition-weighted; fSLAM, fractional SLAM; FT-MRS, Fourier transform-based magnetic resonance spectroscopic imaging; IR FIDs, inversion recovery free induction decay acquisitions; SLAM, magnetic resonance spectroscopy with linear algebraic modeling; SLIM, spectral localization by imaging.
to allow the reconstruction of data with a nonuniform number of averages at each k-space point. An alternative method for solving the SLAM equation (SLAM*) 26 did not produce a significant improvement here, and was not used. For cardiac FT-MRS reconstructions, a single midseptal voxel spectrum was chosen, 8,10 because summing multiple voxels produced contamination artefacts (see the supporting information). The compartmentalization masks used in the reconstruction were defined during acquisition and consisted of three compartments: heart, chest wall, and other (see Figure 2A). Blood was not segmented because of the impact of cardiac and respiratory motion. If present, the heart compartment was removed from the final two slices (base of heart) to avoid aliasing of the heart compartment's SRF into the first slice. The raw reconstructed spectra from each coil element were combined with a noise-whitened singular value decomposition coil combination, 31 phased (zeroth and first order), apodized, and fitted with the Oxford spectroscopy analysis (OXSA) implementation of AMARES, 23,32 with an identical protocol used for all spectra. PCr/(γ-)ATP ratios were blood-and saturation-corrected, as previously reported. 9 The SNR of the PCr resonance and the Cramér-Rao lower bound (CRLB) of the PCr fit were calculated. The coefficient of reproducibility (CoR) was defined as 10 where SD intrasubject is the standard deviation (SD) of the signed difference in the PCr/ATP ratio between scans of the same subject with the same technique. The coefficient of variation (CoV) was defined as the SD of the PCr/ATP ratio across all datapoints for a technique (both repeats) divided by the mean value of the PCr/ATP ratio for that technique. Spectra from one participant with excessive movement (seen by the operator during the scan), and three of the remaining individual spectra, which showed very high liver contamination leading to the blood correction failing, were excluded from the analysis (the corresponding spectra from the other repeat were also excluded from paired analysis by necessity). Errors are reported as ±1 SD. All comparisons were with the AW FT-MRS method, which is considered the standard method at our center. Pairwise twotailed Wilcoxon signed-rank tests with 1% significance level (5% plus Bonferroni correction) were used for all comparisons.

| RESULTS
An overview of the results at each stage of the reconstruction process is given in Figure 2. It shows a representative segmented slice through the heart (a full segmentation stack is shown in the supporting information), the spectra generated by reconstructing the AW dataset corresponding to that volunteer using the SLAM algorithm, a sample fit, and comparison of cardiac spectra reconstructed from the same AW dataset using the SLAM, SLIM, and FT-MRS reconstruction algorithms. Clear spectral differences between the chest wall and heart compartments can be seen, in addition to an increase in SNR from the SLAM/SLIM algorithm compared with the FT-MRS algorithm.

| SRF simulations
Representative cardiac compartment SRFs for each of the compartment-based acquisitions are shown in Figure 4 with, and without, coil sensitivity weighting (an example B À 1 receive field map is shown in the supporting information). The segmentation of the heart as a spheroid resulted in the maximum of the SRF placed on the midseptal voxel, which has previously been shown to be optimal for the AW FT-MRS method, 10 and extending out to the perimeter of the heart. The AW acquisitions feature an elliptical SRF, with the semi-minor axis in the vertical y-direction, reflecting the higher number of phase encodes in the y-direction. The inclusion of coil sensitivity weighting to the theoretical (un-weighted) SRFs skews the AW SLAM and AW SLIM SRFs towards the anterior wall of the heart, in the direction of the coil. The 4 Â 4 Â 4 SLAM and 4 Â 4 Â 4 SLIM SRFs both show slightly greater overspill of the cardiac compartment's SRF into the chest wall than AW SLAM and AW SLIM and a more spherical SRF. The differences between 4 Â 4 Â 4 SLAM and fSLAM are subtle; however, fSLAM does appear to reduce some spillover of the SRF into the chest wall, particularly once coil sensitivity is taken into account.

| Truncated k-space acquisitions
The 4 Â 4 Â 4 SLAM and 4 Â 4 Â 4 SLIM reconstructions both had a significantly higher PCr SNR than the standard AW FT-MRS method; however, the CoV and CoR for the 4 Â 4 Â 4 SLAM and 4 Â 4 Â 4 SLIM methods were higher (worse) than for the AW FT-MRS method. The 4 Â 4 Â 4 SLAM and 4 Â 4 Â 4 SLIM PCr/ATP ratios, and PCr linewidths, were not significantly different from the AW FT-MRS method PCr/ATP ratio. The fSLAM method had a significantly higher PCr SNR than the AW FT-MRS method, and a lower (better) CoV, although CoR was higher.
The fSLAM method's PCr/ATP ratio and PCr linewidth were also not significantly different to the PCr/ATP ratio and PCr linewidth of the AW FT-MRS method.
F I G U R E 4 Simulated cardiac compartment SRF magnitude images for each SLAM/SLIM algorithm-based method for a representative compartmentalization mask (Figure 2A) and AW FT-MRS midseptal voxel PSF (same as Figure 2). Both theoretical and B 1 receive sensitivityweighted SRFs are shown. The MRSI grid is shown in green, and the midseptal voxel highlighted in white in the AW FT-MRS PSFs. The SRFs are displayed as a percentage of the maximum intensity. The coil sensitivity map applied to the SRFs was zeroed, close to the coil elements and within the coil housing, where the Biot-Savart simulation was not valid. AW, acquisition-weighted; fSLAM, fractional SLAM; FT-MRS, Fourier transform-based magnetic resonance spectroscopic imaging; PSF, point spread function; SLAM, magnetic resonance spectroscopy with linear algebraic modeling; SLIM, spectral localization by imaging; SRF, spatial response function.

| DISCUSSION
In this work, we investigated the use of SLAM and SLIM reconstruction algorithms to reconstruct a standard AW phase encode acquisition (AW SLAM and AW SLIM) at 7 T and compared this with a typical Fourier-based reconstruction of the same AW data (AW FT-MRS). We also evaluated two alternative acquisition phase encode schemes (with equal acquisition time) for reconstruction with the SLAM and SLIM algorithms: reasonable sample size. We found that all acquisition-reconstruction combinations delivered a significant improvement in SNR over AW FT-MRS and that AW SLAM also improved reproducibility while maintaining the PCr/ATP ratio. However, the increase in SNR, and consequently the lower CRLBs, did not uniformly translate into reduced variance across the three acquisition datasets, suggesting that the spread in PCr/ATP ratio is dominated by factors other than SNR, such as the impact of the shape of the SRF on sensitivity to motion artefacts and compartment heterogeneity, which is an area of future study, as well as the natural biological variation across a cohort. This finding mirrors the work of Ellis et al., 10 where multiple septal voxels, reconstructed using the AW FT-MRS method, were summed to improve SNR, which did not translate into improved reproducibility.

| AW k-space acquisitions
We recorded a PCr/ATP of 1.90 ± 0.92 using the AW FT-MRS method, which is in good agreement with the previous literature values at 1 Both of the AW SLAM and AW SLIM methods improve SNR over the AW FT-MRS method while improving reproducibility. The PCr/ATP ratio is also maintained for both methods with AW SLAM, achieving a ratio of 1.68 ± 0.50, close to the 1.71 ± 0.65 midseptal voxel value reported by Ellis et al., and AW SLIM returning 2.09 ± 0.43, which is within the range of literature values, [8][9][10]33 suggesting good rejection of out-ofcompartment signals. This is supported by the calculated SRFs, where the low magnitude of the AW SRFs inside of the chest wall suggests that inhomogeneity within the chest wall is unlikely to cause undue bias.
Because the same AW dataset is input into SLAM, SLIM and FT-MRS reconstruction algorithms, it is likely that the SNR increase seen here is caused by the increase in coil sensitivity-weighted tissue volume (see Figure 4) of the cardiac compartment over the single voxel used in the FT-MRS algorithm, particularly because the SLAM reconstruction simplifies to the FT-MRS if the compartment masks are set to the voxels of FT-MRS. The observed SNR increase over FT-MRS offers further reassurance to the application of SLAM and SLIM at 7 T, because if B 0 inhomogeneity within the heart was a serious concern, SNR would be impacted through signal cancellation. Furthermore, poor B 0 inhomogeneity within a compartment would be expected to cause poor spectral line shapes, 16 whereas the observed PCr linewidth for AW SLAM and AW SLIM is not significantly different to AW FT-MRS. The origin of the SNR increase seen here is different to the previous body of work on SLAM, 20 which used a reduced set of high SNR central k-space phase encodes to achieve a per unit volume increase in SNR over a higher resolution uniform set of kspace phase encodes. However, the increase in sensitive volume while maintaining the PCr/ATP ratio is only possible because of the compartment-based reconstruction producing a cardiac compartment SRF, which closely matches the shape of the heart, unlike the summation of AW FT-MRS voxels (see the supporting information).
These results, particularly when combined with the calculated AW SRFs, suggest that the AW SLAM and AW SLIM methods afford improved performance compared with AW FT-MRS for cardiac 31 P MRS at 7 T where B 0 and B 1 inhomogeneity are of greater concern than at 3 T.  Abbreviations: ATP, adenosine triphosphate; AW, acquisition-weighted; CoR, coefficient of reproducibility; CoV, coefficient of variation; CRLB, Cramér-Rao lower bound; fSLAM, fractional SLAM; FT-MRS, Fourier transform-based magnetic resonance spectroscopic imaging; PCr, phosphocreatine; SLAM, magnetic resonance spectroscopy with linear algebraic modeling; SLIM, spectral localization by imaging; SNR, signal-to-noise ratio.

| Truncated k-space acquisitions
scheme on a per-volunteer basis with the fSLAM optimization algorithm improved both CoV and CoR compared with the 4 Â 4 Â 4 SLAM and 4 Â 4 Â 4 SLIM methods; however, the CoV and CoR for fSLAM were still higher than for the AW SLAM method. Overall, we were not able to demonstrate a significant benefit to using 4 Â 4 Â 4 SLAM, 4 Â 4 Â 4 SLIM, or fSLAM methods rather than the AW SLAM method, because the improved SNR did not translate to improved reproducibility.

| SLAM and SLIM reconstruction algorithms
Both SLAM and SLIM reconstructions significantly increased PCr SNR compared to AW FT-MRS, but without significantly affecting either the PCr/ATP ratio or PCr resonance linewidth. This is probably because of an increase in the coil sensitivity-weighted tissue volume to incorporate a greater proportion of the heart, without including the chest wall, potentially increasing sensitivity to motion. The calculated SRFs suggest that cardiac/chest wall movement through breathing may be an important factor to consider for both SLAM and SLIM algorithm-based methods. Because the SRF is much larger than for a conventional Fourier-based reconstruction, it is more likely that the chest wall will stray into the cardiac SRF during breathing, potentially leading to contamination of the cardiac spectra. While this may present an issue, AW SLAM and AW SLIM each had a lower CoV and CoR than the AW FT-MRS method, and cardiac/respiratory gating, in addition to scanning in the prone position, may ameliorate this concern further.