I am interested in analyzing whether concentration of some metabolite (say, mI) in an MRS voxel is correlated with the fraction of WM+GM within that same voxel, and draw conclusions about the relationships between the concentration of the metabolite and subsequent atrophy. However, I worry that this may be circular, and that the concentration will naturally have some dependency on the relative tissue volume in the voxel and any conclusions drawn will be due to that relationship, and not relationship between the presence of the metabolite and atrophy. Is there a way to do an analysis like this in a sound way? Maybe a partial correlation, correcting for WM+GM fraction?
Right now, I am using the ‘Conc.’ output from LCModel with water scaling turned on, so that concentration is relative to water. I am not correcting this ‘Conc.’ value for CSF (i.e., by dividing Conc. by each individual’s WM+GM tissue fraction within the voxel) since that would cause the concentration to be dependent on WM+GM fraction by definition.
With current methodology, water-scaled concentration estimates require knowledge of the tissue composition, so I agree with you that this problem is somewhat circular until we figure out
a much better way of doing water-scaled concentration estimation. As you say correctly, the water signal you’re using as concentration reference is inevitably coming from different tissue classes. GM, WM and CSF not only have different concentrations of tissue water, but also differ quite considerably in terms of water T_1 and T_2. This is why simply normalizing your LCModel outputs by a factor of (1 - f_{CSF}) is not sufficient: your water signal will still be weighted by the differences between GM and WM. Probably the most sophisticated method has been proposed by Gasparovic et al, but it requires plugging in literature values for water concentrations and relaxation times for all three compartments and for metabolite relaxation times, since measuring them in individual participants is time-prohibitive.
I will also add that LCModel by default does a really crude relaxation correction that a) assumes pure white matter, i.e., a water concentration of ~ 36M, and b) is not adjusted for the TE of the dataset you put in (but rather assumes a TE of about 30 ms depending on field strength), but it is rather controlled by a separate control parameter. See this thread for more information. We circumvent this problem in Osprey by setting these control file parameters to 1 and then conduct the full tissue and relaxation correction outside of LCModel.
Finally, the uncertainties that we gain from not knowing the relaxation parameters and concentration values more accurately are rarely properly propagated into the uncertainty estimate of our final concentration estimate. That’s because error propagation on complicated terms gets really ugly really fast, and you basically need Monte Carlo simulations like Ron Instrella showed in a paper that deserves so much attention IMO.
So, TLDR - I think the Gasparovic method is still the best we have available, but it’s not perfect. We’ll need to be able to measure metabolite and water relaxation times much quicker (MRS fingerprinting comes to mind). I’ve come around to regard tCr ratios as quite useful - yes, they are sensitive to changes in tCr, but at least they are automatically at least partially self-normalized in terms of the tissue composition, and they don’t need to deal with the variability of the water relaxation corrections. I think they might be useful to correlate against atrophy.
I know this probably doesn’t answer your question right out of the gate, but conveys that this is a tricky issue and is still a bit helpful.
Thanks so much for the detailed response, it adds a lot of context for me. It does seem tricky to come to a consensus - I’ll take a look at the resources you’ve linked. My data is from the PCC and from people of all ages, so I am definitely seeing an increase in tCr with age, making it tough to use for normalization. I’ll try and see if there is a way to correctly regress out that age effect first before taking the ratios.
Cool, glad to hear it’s helpful. I think the first step should be to do a full Gasparovic-style tissue/relaxation correction instead of the (too crude) one that LCModel gives you.