I’m running quality checks and seeing spurious echoes affecting the 3.75 ppm Glx signal. I’m trying to retain as much data as possible, should I exclude all the Glx signals that show a spurious echo? As far as I understand, ideally yes, however as per Kreis 2004: “the resulting metabolite contents from metabolites with spectral components in the distorted region should be considered with caution”. Thus, they may still be considered.
So, I’m trying to get an idea of what can be kept or not. For that, I’m considering keeping those which show the expected Glx signal with 2 peaks and excluding those with one large bump. For example, including examples 1, 2, and 3, while excluding examples 4 and 5. Does this make sense?
Thank you for your reply. That was indeed my concern. This project has now been completed, so I won’t be able to obtain more data to address the artefact during the acquisition.
Unfortunately, I believe there is no straightforward method for removing it during post-processing?
I would not judge these out-of-voxel echoes so hard here. Least-squares models tend to be not too bad at ‘fitting right through them’ since they’re basically high-frequency oscillations. You also have the other Glx/GSH signals at 2.25 ppm, and they seem… OK. This isn’t perfect data, but I would be careful to exclude too aggressively.
No straightforward method is established for such kind of postprocessing (yet).
Since the data acquisition is already over, I would recommend limiting the processing range up to ~3.5 ppm, thus removing the corrupted region. This is also a tricky thing, but as Georg noticed, the 2.25 ppm region may still give you the Glx numbers. Cutting the range prevents spoiling them by the imperfect fitting through the artifact.
If I cut the ppm range in some datasets but not others, will the metabolite estimates still be comparable across datasets? This is a repeated visits study, where in some cases, V1 is fine but, V2 has the ghosts - when I run stats (both within and between) will that be problematic?
I would have gone for cutting the range in both V1 and V2 when looking for the differences between V1 and V2, since same processing for compared datasets is more scientifically accurate.
Also I would have been careful when grouping the cut and uncut subjects for further analysis (maybe some tests to show that the V1 → V2 changes are consistent between cut and uncut?). But of course this advise may be not applicable.
I would not use different fitting ranges across subjects. We’ve shown that the choice of fit range, unsurprisingly, causes quite some analytical variability (and that a larger fit range is generally better).
The other thing is that Osprey currently uses the same fit range to model the edit-OFF spectra as well.
