Comparison of different linear-combination modeling algorithms for short-TE proton spectra

Area of Science:

• Neuroscience
• Biochemistry
• Medical Imaging

Background:

• Proton magnetic resonance spectroscopy (MRS) with a short echo time (TE) is crucial for studying human brain metabolism.
• Linear combination modeling (LCM) is a common method for analyzing MRS data, modeling spectra as a combination of metabolite basis spectra.

Purpose of the Study:

• To compare metabolite quantification in short-TE brain MRS spectra using three LCM algorithms: Osprey, Tarquin, and LCModel.
• To assess the agreement and variability of metabolite estimates across different algorithms and vendor-specific basis sets.

Main Methods:

• Analyzed 277 medial parietal lobe short-TE PRESS spectra (TE=35 ms) from a multi-site study, preprocessed with Osprey.
• Modeled spectra using Osprey, Tarquin, and LCModel with identical vendor-specific basis sets (GE, Philips, Siemens).
• Quantified levels of total N-acetylaspartate (tNAA), total choline (tCho), myo-inositol (mI), and glutamate + glutamine (Glx) relative to total creatine (tCr).

Main Results:

• Good agreement in group means and coefficients of variation for tNAA and tCho across algorithms and vendors.
• Substantial disagreement for Glx and mI, with Tarquin systematically underestimating mI.
• Moderate agreement (mean R²=0.39) between individual metabolite estimates from different LCM algorithms.
• Significant correlation between local baseline amplitude and metabolite estimates (mean R²=0.10).

Conclusions:

• While mean metabolite levels show broad agreement at the group level, individual estimates from LCM algorithms show only weak-to-moderate correlations.
• Standardized preprocessing and high spectral quality did not ensure strong agreement between algorithms.
• Findings raise concerns about the comparability of brain MRS studies due to variations in LCM software and smaller sample sizes.