Or “What you need to know about correction for relaxation effects when fitting MRS spectroscopy using automated software”
So I’m finally getting around to writing this blog post, or at least attempting it. This blog post will be about some of the default assumptions two common pieces of MRS fitting software make with regard to relaxation effects (T2 mostly) on magnetic resonance spectroscopy quantification of neurochemistry in vivo and how researchers can/should approach or deal with it.
When fitting proton MRS data and using water as a concentration reference it is important that you are aware of all assumptions and default corrections applied by the software you are using (or indeed by your own calculations) – and that you apply appropriate correction for the tissue content of your voxel, as well as differences in water content and relaxation for different tissue types and relaxation effects of the metabolites of interest. I also think that as spectroscopists, we need to get better at reporting all the assumptions and corrections we are applying. At the very least I think we should state if our LC-Model or Tarquin (or other) fits have been done using the default water concentration (white matter) and attenuation correction (for 30 ms TE at 1.5 T), or if we have set these to other values (like pure water, or grey matter, longer TE and different B0). Sometimes this is done, but a lot of the time it is left out. Why do I think we should make sure it is included? Well it would help with making proper comparisons across the literature to start with, and would help us know if our concentration estimates make sense. Reporting this stuff would also provide useful guides for new researchers entering the field, providing useful guidelines about what exactly goes into generating useful MRS results.
A lot of this particular post has been prompted by questions I keep seeing raised on the MRS forums I am on about appropriate referencing for concentrations in proton MRS and by this recent paper “Influence of echo time in quantitative proton MR spectroscopy using LCModel”,( T Yamamoto, T Isobe, H Akutsu, T Masumoto… – Magnetic Resonance in Medicine, 2015 – Elsevier). Basically, Yamamoto and co report that if they use LC Model with out changing the default settings to fit longish echo MRS data set, they get much higher metabolite concentration estimates then if they use jMRUI and correct for the effect of T2 relaxation on their longer echo MRS data. (They actually collect three TE’s 72, 144 and 288 ms so they can estimate T2 relaxation for their metabolites). Their conclusion is that “if TE is long, LCModel overestimates the quantitative value since it cannot compensate for signal attenuation, and this effect is different for each metabolite and condition. Therefore if TE is longer than recommended, it is necessary to account for the possibly reduced reliability of quantitative values calculated using LCModel.”
These last statements, while not wrong, are actually the result of an unfair comparison between the two methods used. While LC Model is able to compensate for T2 related signal attenuation (to a degree) and does so through the use of a correction factor to account for water attenuation due to T2 relaxation (when using the water reference method) it is set to a default value. This is clearly stated in the manual, and Yamamoto et al actually mention this “correction” factor in the paper. However, as pointed out in the LC Model manual this correction factor assumes a TE of 30 ms (at a field strength of 1.5 T in a voxel predominantly composed of white matter) – and so should be changed when using longer TE’s (or collecting from regions of mixed tissue content). This can be easily done using through the use of an appropriate control file (see the LC Model manual) or through the use of appropriate post processing and correction.
While LC Model can account for differences in water relaxation, Yamamoto et al are correct in mentioning that LC-Model does not correct for T2 relaxation effects for each metabolite, however, this is true for any fitting technique currently available (including jMRUI). So any water referenced results obtained by such methods should be appropriately corrected for potential T2 relaxation effects in a post processing step as a matter of course, and not just taken directly from the fitting program. (Actually, there probably is a case for correcting for relaxation effects even when using ratio’s to Cre if the TE is longer then 40 ms, as the difference in relaxation effects for Cre to NAA, Cho, Glu and Ins is larger then 5% beyond this echo time). Yamamoto et al are aware of this need for appropriate relaxation correction – as a matter of fact they give a nice description of why it is needed, and apply it to their jMRUI results – so why they don’t similarly apply it to their LC Model results is bit of a mystery. (Perhaps they wanted the difference to ensure they could get a publication?). It should be pointed out that if Yamamoto et al had used another more recent piece of fitting software, Tarquin (without changing defaults), they would have found similar increases for the ratios they where reporting at longer echoes. On the face of it this paper is an example of user error – where the authors did not fully understand the tools (LC Model) they were using, leading to the report of an error that would not be present, if those tools had been used correctly. However it does highlight how the use of automated MRS software can get you into trouble if you do not know what the software is doing and what assumptions are used in calculation of your results.
A good description for the need for relaxation correction is provided by Gasparovic et al (2006), where Chuck takes it a step further then Yamamoto et al and brings in the problems associated with inaccurate estimation of water content and water relaxation as well as metabolite relaxation, something that Yamamoto et al do not consider. The role water content plays in the correction calculation is perhaps more pertinent to the patient cohort Yamamoto et al included in their study, patients with glioma, in which the water content and water relaxation values will also likely vary due to differing tissue composition for different glioma types. The fact they do not seem to consider water relaxation changes, or water content, is a little disappointing. For a good discussion of why it is important to consider differences in tissue content from your voxel of interest when trying to get “concentration estimates” see Gussew et al (2012) as well as the early work on using water referencing for “absolute concentrations” (Ernst, Kreis and Ross, 1993, and Kreis, Ernst and Ross 1993).
T2 relaxation has the effect of decreasing the signal seen both for a metabolite of interest, and for the unsupressed water as echo time increases, however as most metabolites have a much longer T2 in tissue then water does (unless it is in CSF or pure water), the effect is greater on the water. Of course, there is still an effect due to metabolite T2 relaxation, but the water will show the greatest effect. If you don’t correct for this loss of signal (greater for water then for metabolites) you will get different values for your concentrations at short TE versus long TE, especially if you reference to water. This occurs because the water signal decreases faster then the metabolite signal as TE increases (e.g T2 of water in grey matter is estimated at 63 ms, while for NAA it is around 250 – 301 ms depending on tissue type (average of a lot of literature values – I have a student writing a review so keep your eyes peeled for that)). Therefore if you are using the unsuppressed water as a concentration reference (as both LC Model and Tarquin will do for you) and you want to get as accurate an estimate as possible (considering all factors that affect signal), you need to correct for a lot of things.
How would you do this for LC Model (and Tarquin)? Well one easy way is to set the default values in the LC Model control file or tarquin parameter file (read the manuals to find out how to do this – I’m not going to do everything for you 😉 ) for water concentration to 55550 mM (approximate concentration of pure water – feel free to correct if you think I should use another value) and attenuation to 1. You can then use the segmentation package of your choice (SPM works well, but so does FSL) and figure out how to calculate the tissue makeup of your voxel (usually by making a mask of your voxels location on the anatomical T1 image and applying this to the segmented images to get CSF and grey and white matter content). You can then use this information to correct your LC Model/Tarquin provided metabolite results for tissue content, water content (using estimates from the literature for each tissue type), water relaxation variations due to tissue type (again use literature estimates or measure it) and metabolite relaxation effects (again literature values or measure them as Yamamoto et al, and others suggest). if you don’t want to mess about with changing the default water concentration and attenuation settings in LC Model (or Tarquin), then you can always correct your results by multiplying them by 2.211 before (or after) you do partial volume and relaxation correction. (In case you are wondering where that number comes from it is the concentration of pure water at room temp (55550 mM) divided by (the default water concentration (35880 mM) times the water attenuation (0.7)), this will then remove the effect of the default water concentration and attenuation from the final result.) NB, if you don’t change the default water correction values you really do need to do this correction if you plan on doing partial volume correction.
“But that seems like a lot of work? do I have to do all that?” I hear you say – well, you don’t have to, of course, you can do what you want, but you might get pulled up by reviewers (like me) who ask you to do it – so may as well do it before you submit your paper. Why might reviewers insist on these steps? Well, differences in tissue composition (and therefore water content and relaxation) can have a real impact on the estimates of your metabolite concentrations for different subjects. This is especially important when you are doing something like trying to correlate performance, or BOLD effects with metabolites. You want to know the effect is a result of the metabolite being correlated, and not just because in some subjects you had more (or less) grey matter. Note, due to differences in Creatine content between Grey and white matter, using ratios to Creatine may not get away from this potential confound. There are some cases were it really isn’t necessary, like dynamic scanning in an individual where you are looking at changes over one long MRS session. You would be applying the same correction across all measures, so it’s kind of a moot point. However if you wanted to make comparisons with performance etc you should either use a delta measure (eg. size of change in metabolite A) or do the tissue and relaxation correction appropriately. (Even if just to stop me asking you to go back and do it if I happen to review your paper).
By the way, if you do do this sort of correction and reference to water, your results will be in mM – not “Institutional Units” (although I have seen, and used, “institutional units of mM”) or “arbitrary units”. If you do the calculations correctly, as the values for water content you use in the referencing calculation are in mM (or should be), your results will likewise be in mM – that’s the way the maths works (units carry through). Your results may be only be an estimate due to several assumptions, but the unit of your estimate is still in mM (note this is molal – not molar).
I have mentioned partial volume correction, or tissue content correction, and suggested you make a mask of your voxel location to apply to your tissue segmented anatomical images. This is an easy thing for me to type, but not everyone knows how to do this. I’m lucky, I’ve worked with a few people who have figured out how to make a voxel mask in the same space as the anatomical image – but this is not a trivial task, and even the bits of code I have don’t aways work for all occasions. I have previously shared some of the mat-lab based code we use for Philips single voxel MRS (at bottom of the page), but if you are reading this and you have some code for other vendors (or an improvement on our Philips code) please post it in the comments or let me know if I can share it and I will update this post.
Okay, that’s about it for relaxation correction for now. I will add some more on T2 relaxation and it’s potential effects on metabolite visibility in the future, but hopefully some of you find this useful and it reminds you of some of the details you should pay attention to when analysing your data.