Comparison of algorithms for linear-combination modelling of short-echo-time magnetic resonance spectra [article]

Helge J Zollner, Michal Povazan, Steve Hui, Sofie Tapper, Richard AE Edden, Georg Oeltzschner
2020 bioRxiv   pre-print
Short-TE proton magnetic resonance spectroscopy is commonly used to study metabolism in the human brain. Spectral modelling is a crucial analysis step, yielding quantitative estimates of metabolite levels. Commonly used quantification methods model the data as a linear combination of metabolite basis spectra, maximizing the use of prior knowledge to constrain the model solution. Var-ious linear-combination modelling (LCM) algorithms have been integrated into widely used commercial and
more » ... e analysis programs. This large-scale, in-vivo, multi-vendor study compares the levels of four major metabolite complexes estimated by three LCM algorithms (Osprey, LCModel, and Tarquin). 277 short-TE spectra from a recent multi-site study were pre-processed with the Osprey software. The resulting spectra were modelled with Osprey, Tarquin and LCModel, using the same vendor-specific basis sets. Levels of total N-acetylaspartate (tNAA), total choline (tCho), myo-inositol (mI), and glutamate+glutamine (Glx) were quantified with respect to total creatine (tCr). Mean spectra and models showed high agreement between all vendors and LCM algorithms. In contrast, the algorithms differed notably in their baseline estimates and mI models. Group means and CVs of the metabolite estimates agreed well for tNAA and tCho across vendors and algorithms. For mI and Glx, group means metabolite estimates and CVs agreed less well between algorithms, with mI systematically estimated lower by Tarquin. Across all datasets, the metabolite-mean CV was 10.4% for Osprey, 12.6% for LCModel, and 14.0% for Tarquin. The grand mean correlation coefficient for all pairs of LCM algorithms across all datasets and metabolites was R2= 0.39, indicating generally moderate agreement of individual metabolite estimates between algorithms. Stronger correlations were found for tNAA and tCho than for Glx and mI. Correlations between pairs of algorithms were comparably moderate (Tarquin-vs-LCModel: R2= 0.43; Os-prey-vs-LCModel: R2 = 0.38; Osprey-vs-Tarquin: R2= 0.37). There was a significant association between local baseline power and metabolite estimates (grand mean R2= 0.10; up to 0.62 for LCModel analysis of Glx in Siemens datasets). Metabolite estimates with stronger associations to the local baseline power (mI and Glx) showed higher variance, suggesting that baseline estimation has a stronger influence on those metabolites than on tNAA and tCho. While estimates of major metabolite complexes broadly agree between linear-combination model-ling algorithms at group level, correlations between algorithms are only weak to moderate, despite standardized pre-processing, a large sample of young, healthy and cooperative subjects, and high spectral quality. These findings raise concerns about the comparability of MRS studies, which typically use one algorithm and much smaller sample sizes. The study is the first benchmark comparison of different LCM algorithms using large datasets. Finally, it provides a standardized framework to interrogate factors with critical impact on LCM out-comes.
doi:10.1101/2020.06.05.136796 fatcat:lt4xny2hdjh3bbxn3pmd44dq6u