Hi all!
What are the advantages of using the group normed metrics over the ordinary alpha correction in water-scaled GABA and Glx/glutamate estimates? I have tried to read up on this, but I do not really understand.
/Linda
Hi @Linda_S ,
For many metabolites, we expect different concentrations in gray matter vs white matter, which may confound findings. Alpha correction aims to account for this difference, by adjusting measurements to some nominated standard tissue composition. This can be quite effective, particularly in cohorts with age- or disease-related atrophy; see for example: Impact of tissue correction strategy on GABA-edited MRS findings
The āoriginalā alpha correction aims to normalise GABA estimates to an equivalent pure-GM voxel (ie, 100% GM, 0% WM, 0% CSF). This is straightforward and convenient, especially if comparing across groups/locations with different compositions. However, the efficacy can be limited by the accuracy of the underlying segmentation, and moreover knowledge/assumptions on the relative concentration of the metabolites in different tissue. Often the assumption is simply that GABA concentration in WM is half that of GM; this works fairly well in most cases, but might not be accurate in every scenario (and almost certainly wonāt translate to other metabolites). Especially if your actual voxel has a relatively large WM fraction, minor inaccuracies in the segmentation or the underlying assumptions can amplify to give larger variability in the alpha-corrected (pure GM) estimate. Figure 1 of that paper nicely illustrates one aspect of that risk.
Alpha correction to a group norm aims to mitigate these issues, by normalising to the typical voxel composition in your data, rather than a pure GM voxel ā meaning the adjustments are generally on a more moderate range, hence thereās less risk of amplifying errors.
In summary:
- If you have a high proportion of GM in your voxels and/or want to maximise comparability across regions etc, ordinary alpha correction to pure GM is a good choice.
- If your GM fraction is a bit lower, group-normalised alpha correction may be a safer option.
3 Likes
Thank you so much, @alex! Your explanation was very helpful.
Hi @alex,
I was also confused about these two correction methods, thank you for your explanation! Iād like to give two examples: do you think that a voxel with 65% GM, 20% WM, and 15% CSF in the PCC is āa high proportion of GMā? Or that a voxel with 45% GM, 45% WM, and 10% CSF in the unilateral parietal lobe is āa bit lower GM fractionā?
I also want to ask whether the metab/tCr ratio requires ordinary or group-normalized alpha correction?
Yushan
Hi @Yushan,
There isnāt really a rule for this (as far as Iām aware), so Iām hesitant to specify levels ā I expect the more appropriate choice would depend a bit on the dataset, the study design, and the research question.
One way to assess this may be to look at the outcomes from each approach; if alpha correction to pure GM appears to have introduced more variance, then alpha-correction to group normalised GM may be a better choice. If in doubt, I would lean towards the more conservative correction (group norm), although keeping in mind that this may compromise comparability of absolute values.
Note also that you donāt have to use alpha correction in every study; if you expect systematic differences in tissue content (eg, age-related) itās a good idea, but in many other settings it can be sufficient to use tissue-corrected values without alpha correction then just check for GM fraction as a possible confound in your statistics.
As far as Iām aware, alpha correction is typically only applied for water-referenced GABA and Glx estimates.
In principle other metabolite concentrations (including the Creatine reference) may also differ across tissue types, so similar techniques could be relevant ā but this isnāt commonly done and would depend on having a reliable estimate of the relative content of each metabolite in each tissue class. So for creatine-referenced values, again I would just consider tissue fraction in the statistical model as a possible confound.
Dear Alex,
I just saw the reply, thank you very much for your detailed suggestions!
Yushan