Choosing the most appropriate Osprey/LCModel quantification strategy for MRS data in clinical pediatric cohorts

Ask Anything!
1

Hello,

I am currently completing my final-year internship at Sainte-Justine University Hospital in Montreal, where I am working on proton MRS data analysis using Osprey.

I am currently working with two different datasets:

  • A dataset acquired at a hospital in Paris from patients with anorexia nervosa and ARFID, in whom we expect potential brain metabolic alterations that we hope to characterize using MRS.
  • A dataset acquired at Sainte-Justine Hospital that includes patients with anorexia nervosa, ARFID, and healthy controls.

Both studies include short-TE and long-TE SVS acquisitions.

For the Sainte-Justine dataset, we plan to add MRI sequences to acquire T1 maps, T2 maps, and Proton Density maps. This will allow us to estimate subject-specific T1 and T2 values for WM, GM, and CSF, instead of relying on literature values, which may not be appropriate for our pediatric clinical population.

Because of this, I thought that using the Osprey quantification method in the GUI (rather than LCModel output) and reporting the TissCorrWaterScaled concentrations would be the most appropriate approach, since these values account for tissue fractions (WM/GM/CSF) as well as subject-specific tissue relaxation corrections.

Importantly, we are not planning to directly compare the two cohorts with each other. The two datasets will be analyzed separately, with different objectives and study designs. Therefore, our main concern is to select the most appropriate quantification strategy for each dataset independently, rather than necessarily applying the exact same processing pipeline across both cohorts.

However, for the Paris dataset, we do not have T1 or T2 maps. Our initial plan was to use literature-based pediatric T1 and T2 relaxation values, such as those reported by Lee et al. (2018), to perform tissue relaxation correction and study the concentrations from TissCorrWaterScaled. However, since our participants are patients with eating disorders rather than healthy children, I am unsure whether this would be the most appropriate approach.

An alternative I have been considering would be to use the LCModel output instead, apply only a manual CSF correction, and then include the GM fraction as a covariate in the statistical analysis. This approach has been used in several publications, such as Castro et al. (2010), Goldwska et al. (2017), and Perdue et al. (2024).

My reasoning is that this might avoid introducing an additional source of bias from tissue relaxation corrections based on literature values that may not accurately reflect our clinical population. However, I am not sure whether this is the most appropriate strategy, and I would greatly appreciate your opinion.

More generally, I am also not entirely sure that I understand the practical difference between selecting the LCModel or Osprey quantification methods in the Osprey job file. My understanding is that the fitting algorithms differ, leading to different estimated metabolite concentrations, but I am uncertain whether one approach would be preferable over the other in our context.

Finally, I have one additional question regarding the output units. Are the metabolite concentrations reported in TissCorrWaterScaled expressed in mol/kg or mmol/kg?

Overall, given our study design, would you recommend using the TissCorrWaterScaled, the LCModel output, or perhaps another output CSV altogether?

Thank you very much for your time and for developing such a valuable tool. I would greatly appreciate any advice or recommendations you may have.

Kind regards,
Nina Cordier

2

Dear Nina,

We are doing something similar to what you describe in terms of using Literature and/or Real qMRI water relaxation times. There is a stat.csv file in Osprey, where you could provide some data which is missing from the input files (for example qMRI values for every voxel) and with a small code modification these values would be used. Do you already have a technique to extract the qMRI values from the voxels? We have just released a tool which can do this (if you use Siemens or Philips).

The quantification questions are not related to the “Osprey” or “LCModel” options in the job file at all. These two are just the algorithms which will be used solely for the spectral fitting (i.e. obtaining intensities of the for the metabolites and reference water). You should choose between them depending on how they (and you) manage. I believe that for the spectra measured in the normal-appearing cerebral tissue (even if the participants are patients and not healthy subjects) Osprey could be a better choice, it’s easier. For the difficult cases for example High-Grade tumors LCModel might be a better option (my experience). It has to be checked with your data.

All the quantification is performed later at the Quantify step. It’s hard to speak about the possible bias in your study cohort without having more information. Since everything (CSF-corrected, TissCorrWaterScaled and so on) is quantified in one pipeline, you can get all the concentrations quantified differently, and then after careful thinking address the quantification/statistical approach if you have some suspicions.

Best regards,
Andrei

P.S. Just in case: I’m not an Osprey developer

3

As we have always done in Gannet, the quantified values in Osprey are reported in “institutional units (i.u.)”. The philosophy here is that even with all the tissue corrections applied to the TissCorrWaterScaled values to obtain a (pseudo-)absolute biochemical measurement, several data-dependent scaling factors remain uncorrected/unaccounted for. Using i.u. is a simple way around the issue of what units MRS measurements should be reported in.

4

If doing it this way, I believe you should normalize fractional GM by the total proportion of brain tissue. That is, use fGM / (fGM + fWM) as your covariate. Since you’re already applying a CSF correction. This gives you fractional GM per unit of tissue.