Introduction

Diffusion-weighted MR spectroscopy (DWMRS) is a unique tool to non-invasively probe the microstructure of cell-specific intracellular spaces in vivo^[1]. Indeed, DWMRS has the potential to serve as a biomarker for neurodegeneration (neuronal-markers: tNAA, Glu) and neuroinflammation (astrocytic/glial-markers: tCho, Ins)^[2,3,4], where processing and fitting of DWMRS data significantly impacts the estimation of metabolites’ diffusion properties and, in turn, accurate modeling of cell-specific microstructural features. Although most processing and fitting steps overlap with methods used for standard MRS^[5], DWMRS also relies on the evaluation of multiple spectra acquired under different conditions (e.g. various b-values, encoding directions, etc.). Data at higher b-values are more prone to error, due to: higher sensitivity to motion, stronger individual phase drift, and lower SNR.
As the DWMRS community grows, we wanted to investigate how reproducible and reliable our research is, and possibly identify the most robust steps to process and fit DWMRS data. This was the aim of a workshop held at the Lorentz center in Leiden (NL) in September 2021. During the workshop, a processing and fitting challenge was proposed to the participants. We hereby present the setup of the challenge and its main results.

Materials & Methods

Deposited dataset: Simulated spectra of brain metabolites were generated using Matlab for a STEAM sequence with TE/TM=45/60ms at 7T (ideal RF-pulses, spectral width 3 kHz, 1024 complex points), and relying on published chemical shifts, J-coupling constants^[6], metabolite concentrations and T₂ relaxation values^[7]. A synthetic spectrum consisted of a weighted sum of 17 metabolites and included a macromolecular (MM) baseline (Fig.1A). Signal attenuation for individual metabolites at 9 different b-values (ranging from 0 to 50 ms/μm²) was generated using a model based on randomly oriented infinite cylinders^[8] and a mono-exponential decay for the MM (Fig.1B), resulting in 9 diffusion-weighted synthetic spectra (Fig.1C). Finally, to emulate in-vivo datasets, a total of 32 transients for each diffusion-weighting condition were imposed with independent line-broadening, random complex Gaussian noise, and frequency and phase drift (Fig.1D/E). The range for frequency drift across transients was similar for all b-values ([-4,4] Hz), but phase drift was increased with b-value. Ground-truth (GT) spectra (without noise and drifts) and synthetic spectra (with noise/frequency/phase drifts), sequence details, and a basis-set for fitting were shared on GitHub (https://github.com/dwmrshub/pregame-workshop-2021). The model used to generate signal attenuation across b-values and the GT on individual frequency and phase drifts were not shared with participants.
Challenge deliverables: The processing and fitting steps submitted by each participant (n=8) are summarized in Fig.2. To disentangle the effect of processing and fitting steps, we asked for the following deliverables:
(1) Processing alone: residuals between the GT and processed spectra at each b-value (Fig.3A) and the phase/frequency corrections applied at each transient.
(2) Fitting alone: signal decays (normalized to b=0ms/μm²) of 5 metabolites (tNAA, tCr, tCho, Glu, and Ins) and MM for the GT data. Not shown.
(3) Processing+fitting: signal decays (normalized to b=0ms/μm²) of 5 metabolites and MM for the processed and fitted synthetic spectra (Fig.3B).
Modeling: The original model (Fig.1B) was applied to the signal decays generated from the processed and fitted synthetic spectra to evaluate how much variability in the spectral fit affects the estimation of biophysical parameters (Fig.4C). Apparent diffusion coefficients (ADCs) were also computed with a mono-exponential fit applied between b=0 and b=3ms/μm² (Fig.4B).

Results & Discussion

As reported in Fig.2, participants used different platforms to process and fit the DWMRS data. All pipelines included individual phase-and-frequency correction. The various processing pipelines resulted in differences in residuals between processed spectra and GT (Fig.3), illustrating pipeline-specific peculiarities. This can also be appreciated from the difference in MSE between the fitted GT and synthetic data (Fig.3B). Fig.3C shows variability across participants due to both processing and fitting steps. Decays are relatively similar for tNAA and tCr and slightly more variable for tCho at higher b-values. Results for J-coupled metabolites and MM are more variable across b-values due to lower SNR and increased CRLBs at higher b-values (Fig.4). Such differences could potentially be alleviated by acquiring more transients at higher b-values. ADCs given in Fig.4A, and Fig.4C illustrate the variability in the modeling output (radii and D_intra).

Conclusion

This processing and fitting challenge is a first step towards understanding points of convergence and divergence in different processing and fitting pipelines. Most participants were already very familiar with DWMRS and yet results substantially vary as soon as conditions worsen (typically when CRLBs > ~5%). Further analyses will account for the fitting variability (typically, using the CRLBs) to report this uncertainty on the modeling. Dynamic fitting approaches (e.g. P1 and P7) fitting the spectra including a prior about the expected decay (here, bi-exponential) could help reduce the uncertainty at high b-values. The challenge is still open to everyone interested in DWMRS. Future steps will include recruiting more participants, and more realistic distorsions to match in-vivo conditions (eddy-currents, lineshape, amplitude variation, SNR, etc.).

Acknowledgements

This abstract was initiated by the Workshop ‘Best Practices & Tools for Diffusion MR Spectroscopy’, held at the Lorentz Center, Leiden, The Netherlands on 20-24 September 2021, and tremendously benefited from committed discussions among the attendees. The workshop has received funding from the Lorentz center and NeurATRIS (A Translational Research Infrastructure for Biotherapies in Neurosciences, “Investissements d'Avenir”, ANR-11-INBS-0011). HL and NJ have received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement No 804746). AD received funding from the Swiss National Science Foundation (SNSF #188142, #202962). GG received funding from the National Institutes of Health [grant numbers R01MH113700, R01AG039396, P41 EB015894, P30 NS076408] and the W.M. Keck Foundation