Synopsis

Keywords: Image Reconstruction, Spectroscopy

We utilized a quintuple-tuned RF head coil array by using the high-SNR ²³Na signals from the brain to optimize the weighting for combining signals from the same coil array elements for the low-concentrated ³¹P metabolites. ²³Na-weighted Roemer combination of ³¹P MRSI signals is verified on EM simulations and MR experiments. Comparing to ³¹P self-weighted combination, ²³Na-weighted combination shows higher SNR and better-combined spectra in regions with low intrinsic SNR. It also shows potential of mitigating the signal contaminations when using ³¹P-self-weights for ³¹P data acquired with large voxels.

Introduction

³¹P MRSI is a powerful tool to study energy metabolism¹ and cell proliferation and therefore being used in cancer treatment efficacy monitoring². However, ³¹P MRSI is not widely used because of the low ³¹P SNR, (caused by low natural abundance and low gyromagnetic ratio compared to ¹H). To harvest all potential SNR, some researchers use close-fit loop arrays for signal reception. However, the multi-channel signal combination³ can be challenging at low-SNR regions because coil sensitivity distributions are difficult to measure. Self-weighted techniques like WSVD⁴ combination and PCA based denoising have made significant progress in this area⁵. In this study, we explore the feasibility of another approach enabled by a coil array design which contains 15 loop-coils tuned to phosphorous, sodium and carbon⁶. We will investigate combining multi-channel ³¹P signals using the sensitivity distributions of ²³Na which is expected to be similar because they share the same coil array geometry and the frequencies of operation of both nuclear species are still in near-field regime.
Self-weighted Roemer combination method works reasonably well if SNR is sufficiently large (e.g. brain). However, when it comes to the body region, or regions surrounded by highly concentrated ³¹P signals (i.e. muscle), the low-SNR makes self-weighted combination more problematic, while on top of this the signals contain a significant portion of voxel bleeding from neighboring voxels and therefore no longer represent the coil sensitivity.
This study will investigate the potential for using ²³Na sensitivity maps for signal combination of ³¹P MRSI. We will test the applicability by EM simulations and brain MRSI scans. As the self-weighted Roemer combination is still robust in most brain regions, our experiment with ²³Na-weights can be validated.

Methods

Simulations: We performed EM simulations on Sim4Life (Zurich Med Tech, CH). We simulated the B₁^- fields at the frequencies of ³¹P and ²³Na on DUKE⁷. Figure 1a shows the simulated model. By approximating ³¹P signals with simulated B₁^-*, we compared self-weighted and ²³Na-weighted Roemer combinations by evaluating the theoretical SNR. Equation 1 is used for SNR evaluation³, omitting all global scaling factors:
$$SNR^2=\frac{(\omega MV)^2 \sum_{i=1}^N \sum_{k=1}^N n_i n_k B_{ti} B_{tk}}{4KT\Delta f \sum_{i=1}^N \sum_{k=1}^N n_i n_k R_{ik}\cos(\theta_i - \theta_k)} \qquad \text{(Eq1)}$$
MR experiments: The MR experiments were performed on a 7T MRI system (Philips Healthcare, Best, the Netherlands). An embedded-in-bore birdcage coil is used for ³¹P transmission. A Helmholtz coil is used for ²³Na transmission. A quintuple-tuned head coil⁶ is used for reception. Figure 1b-c show the coil. ³¹P MRSI is acquired with the following specs: FOV = 320(FH)x260(AP)x240(RL) mm³, resolution = 20x20x20 mm³, TR = 300 ms, TE = 0.5 ms, FA = 20°, readout bandwidth=5000 Hz, 18 averages. ²³Na MRI was acquired with the same resolution.
Data processing: Data processing is performed in Python for simulations and on MATLAB (The MathWorks, Inc.) for MR data. PCA-based denoising⁵ is performed after signal combination. Roemer’s optimal-SNR method is used for multi-channel signal combination (Eq2 below). We implement the ²³Na weights by replacing the $$$b$$$ in Eq2 with ²³Na MRI of identical resolution. The channel-wise constant ³¹P/²³Na phase offset due to arbitrary phase shifts in the Rx chain is calibrated according to the ³¹P/²³Na phase discrepancy at the maximum-²³Na-signal locations.
$$P=C \frac{p^T R^{-1} b}{\sqrt{b^T R^{-1}b^*}} \qquad \text{(Eq2)}$$

Results

B₁^- maps comparison: Figure 2 shows the simulation-based comparison between the two combination methods on two sampled slices. The SNR can drop by 10% to 20% when using ²³Na weights.
In vivo MR: Figure 3 shows the SNR analysis of combined ³¹P MRSI from an in vivo scan. The second and the third row correspond to the SNR of the combined signals with ³¹P-self-weights and with ²³Na weights respectively. Figure 4 compares these two methods by presenting the combined MRSI of slice 4. Figure 5 presents combined spectra comparisons of more sampled voxels.

Discussion

Simulations demonstrate that the weights for ³¹P coil element combinations can be correctly determined from ²³Na signals of the same coil elements to maintain highest sensitivity. Also for most in vivo slices the ²³Na weights are well assessed. However, looking at the first three columns of Figure 3, it seems like self-weighted Roemer combination generates undoubtably higher SNR for ³¹P MRSI. In fact, in center voxels the ²³Na-weighted method shows almost no signal where self-weights combination shows outstanding SNR. It should be noted however that these center voxels are mainly in CSF where ³¹P concentration is low and sodium content high. Moreover due to the large voxels, there will be substantial point-spread contributions from higher signals from neighboring voxels towards the coils. Consequently, the observed signal will not provide the correct amplitude and phase weighting information for coil combinations. In the ³¹P-(self)-weighted reconstruction, SNR seems high, but this may be severely biased by optimizing the weights for the point-spread contribution. This is confirmed by substantial artefacts in ³¹P spectra from areas in the head close to tissue interfaces when using self-weighting while the spectra have good quality when using ²³Na weighting.

Conclusion

By using the same receive elements, ²³Na-weighted Roemer combination of ³¹P CSI performs well in brain, and it even shows potential advantage over ³¹P-self-weights.

Acknowledgements

Great thanks to Lieke van den Wildenberg and Ayhan Gursan for their guide in data precessing!