Synopsis

Keywords: Data Processing, Spectroscopy, Spectra Registration

Diffusion-weighted MRS (DW-MRS) suffers from low signal intensity and large phase errors when applying strong diffusion weighting. Coil combination and spectral registration are vital as spectra with high b-values are usually poorly reconstructed. This work aims at the optimization of post-processing pipeline with singular value decomposition (SVD)-based spectral registration and the comparison of different regimes of coil combination and spectral registration for in-vivo DW-MRS post-processing. The SNR of post-processed spectra and the metabolite diffusivities are assessed. The statistical outcomes suggest that the SVD-based spectral registration could improve the SNR of DW-MRS signal at high b-values and therefore alleviate diffusivity overestimations.

Introduction

DW-MRS allows for studying molecular diffusivity to reveal the microstructural and microenvironmental properties in vivo. However, the phase/frequency errors among multiple acquisitions in DW-MRS are non-negligible, especially at large b-values. Many efforts have been made in the past to improve MRS post-processing for a better SNR in the coil combination and spectral registration process. The MRS data processing toolkit FID-A¹, GANNET² and Osprey³ all adopted the time-domain coil combination which calculates the phase and amplitude weighting using the first time point data of FID^4,5 (referred to as tdCoilComb). While an SVD-based method was proposed for coil combination⁶ (svdCoilComb) without referencing to any specific time point. For spectral registration, the robust spectral registration (rSR) method was proposed to align individual transients using an optimal reference and nonlinear least-squares optimization algorithm⁷. In this work, we proposed an SVD-based spectral registration (svdSR) to improve the quality for DW-MRS. This work compared four different regimes for DW-MRS post-processing. The spectral SNR at different b-values and the apparent diffusion coefficients (ADCs) of different metabolites were assessed.

Methods

Instrument:Experiments were performed on a 3T Siemens MAGNETOM Prisma MRI scanner, equipped with 80mT/m gradients and 64-channel head phased-array receive coils.
In vivo Experiment: Five healthy volunteers (male/female=1/4; 22-25 years old) participated in this study. The DW-MRS acquisitions were performed using a diffusion-weighted STEAM sequence (TE/TM/TR=62/45.6/1500 ms) in the frontal lobe. The scan parameters included: voxel size = 24*60*25 mm³; spectral width = 4 kHz and 2176 complex datapoints. Diffusion-weighting was applied in three orthogonal directions with diffusion gradient duration (δ) = 20 ms, diffusion time (Δ) = 80 ms. Seven b-values ranging from 0 to 5000 s/mm² (0, 833, 1667, 2500, 3334, 5000) with 20 shots were acquired.
Data Processing:The schematic diagram of DW-MRS processing pipeline was shown in Figure 1. The raw data were reconfigured by down-sampling in the time domain. Two coil combination methods, tdCoilComb⁴ and svdCoilComb⁶ were used to combine the individual channels with amplitude and phase weighting respectively.
For spectral registration, we proposed a method based on SVD algorithm, which extracts the principle component to produce the aligned signal in all transients according to the algorithms described below: $${{S}_{unaligned}}=U\Sigma {{V}^{H}}$$$${{S}_{aligned}}={{\Sigma }_{1,1}}{{u}_{n,1}}\left| v_{m,1}^{H} \right|\times {{exp}({\phi {{({{v}_{m,1}})}_{median}}})}$$
where the unaligned signal matrix is denoted as $$${{S}_{unaligned}}\in {{\mathbb{C}}^{n\times m}}$$$; the numbers of complex points and shots are denoted as $$$n, m$$$; $$$H$$$ is the Hermitian transpose. After SVD decomposition, the signal were aligned as product of the maximum singular value $$${{\Sigma }_{1,1}}$$$ and the corresponding left singular vectors $$${{u}_{n,1}}$$$weighted by the amplitude and median phase of right singular vectors $$${{v}_{m,1}}$$$.
At the same time, the rSR method⁷ was adopted as a comparison in the process of frequency and phase correction before the final averaging step. The averaged spectra was then processed for spectral fitting and ADC calculations. AMARES (jMRUI-v6) algorithm was used to quantify metabolite signals. The metabolite ADCs were calculated from a mono-exponential model (fits up to b₄ and b₆). The spectral SNR was calculated according to the NAA signal over the standard deviation of spectral noise.
Statistical Analysis:The paired t-test was applied to assess four different regimes for the spectral SNRs and ADC measurements. P < 0.05 was considered statistically significant.

Results

Figure 2 presented the alignment using svdSR and rSR after tdCoilComb at b₀ and b₅. At b₅, the svdSR method showed better performance in recovering the highly diffusion-weighted signals.
Figure 3a showed the spectral SNRs of all the post-processed spectra at 7 b-values in 3 directions from five volunteers. The SNR decreased globally as the b-value increased. The SNRs were similar between coil combination methods (tdCoilComb vs. svdCoilComb) while were slightly but significantly differed between spectral registration methods, that was, svdSR method showed a better SNR than rSR at higher b-values, especially when b > 3000 s/mm2. Figure 3b,c presented the trace ADC (arithmetic mean of three directions’ ADCs) of 3 metabolites with different b-ranges in five volunteers. Among four proposed regimes, the trace ADCs calculated from the post-processing using time domain coil combination (tdCoilComb) plus SVD-based spectral registration (svdSR) were the lowest among all pipelines.

Discussion

In this study, SVD-based spectral registration was found to be beneficial to improve in-vivo DW-MRS measurement. The spectra registration using SVD extracts the major component in all spectra, which achieves the denoising and phase/frequency alignment of individual spectra at the same time. However, it remains unclear whether multiple SVD procedures or different numbers of major components extracted would influence the ADC calculation - which needed to be evaluated in the future. Moreover, there are some other coil combine methods such as WSVD⁸, GLS⁹, as well as other spectral registration methods such as tdSR¹⁰, RATS¹¹, which were also used for MRS post-processing. In the future, these methods could all be parallelly compared and assessed with a larger sample size. Test-retest should be reported to verify which post-processing pipeline is optimal for edited MRS, DW-MRS as well as conventional MRS.

Conclusion

This study has examined different coil combination and spectral registration methods for DW-MRS post-processing. The tdCoilComb-svdSR pipeline achieved best post-processing of in-vivo DW-MRS data. The proposed SVD spectral registration method improved the spectral quality but needs some further verifications.

Acknowledgements

No acknowledgement found.