I’m new to fsl_mrs and am running into some problems getting my neurotransmitter of interest, GABA, to show up.
Data is collected as a single 2x2x2 voxel on a 7T Siemens TERRA scanner using sLaser. Files are saved as dicom which I’ve converted using spec2nii without any real issues. I then went through the standard pipeline using fsl_mrs_preproc and fsl_mrs to fit my data and for some participants it seemed to have worked fine, but for the majority no GABA shows up.
I don’t think it’s a data issue as I’ve also checked with a control subject from someone else’s study that is supposed to have great levels of GABA previously analysed using LCmodel.
I assume I’m going wrong somewhere, but I wouldn’t know where to start looking. Any help or ideas would be greatly appreciated!
GABA is a small signal and difficult to reliably and unequivocally resolve without spectral editing, even at 7T.
LCModel uses internal soft constraints and expectation values for certain low-level metabolites like GABA (see section 11.8 in the manual), which encourages GABA to be modeled at relatively consistent levels (low variance) at the risk of providing a bias. I don’t think the effects of the strength of the soft constraint have been thoroughly studied.
I’m not sure FSL-MRS uses similar soft constraints (Will will hopefully clarify), so I wouldn’t be surprised if that is the main reason for the difference, i.e., you did nothing wrong.
Thanks for giving FSL a try. The fit you show looks mostly good (what are the settings for the macromolecules?), so you should be able to get an estimate for GABA.
Georg is of course right, LCModel uses some prior knowledge on the concentration of GABA to initialise the fit with some GABA concentration. FSL doesn’t, trying to avoid biasing the outcome. However, this does seem to mean that sometimes using the default (quick) Newton solver [GABA] comes out as zero. To solve this, the best approach is to run using the full Metropolis Hastings solver. You can do this by using the --algo MH flag (at the same time use --mh_samples 2000 to get a decent number of samples, and I’d suggest increasing this to ~10000 once you want to do the final fit). The MH solver is able to estimate the posterior distribution of [GABA]. You will see this in the output HTML report as a histogram of values. A mean GABA conc will be given. Let me know how that goes, and we can always refine from there.
Could you add --ind_scale Mac? I expect the MM basis set is scaled massively differently to the other basis set spectra. And if you aren’t already, use --metab_groups Mac. Your MM’s look reasonable for the echo time.
Yes, all looks good. Definitely check in with Chris and Carina about the processing of the results. I’ve been meaning to add their method into FSL, but haven’t got around to it yet.
