Seeking Ground Truth for GABA Quantification by Edited Magnetic Resonance Spectroscopy:Comparative Analysis of TARQUIN, LCModel, JMRUI and GANNET

Purpose: Many tools exist for quantifying magnetic resonance spectroscopy (MRS) data. Literature comparing them is sparse but indicates potential methodological bias. We benchmark MRS analysis tools to elucidate this. Methods: Four series of phantom experiments, including both solutions and tissue-mimicking gels, with constant concentrations of NAA, Creatine, Glutamine and Glutamate, and iteratively increased concentrations of GABA are performed. MEGA-PRESS spectra are acquired and quantified with several state-of-the-art MRS analysis tools (LCModel, TARQUIN, JMRUI, GANNET) and in-house code (LWFIT). GABA-to-NAA ratios for reported metabolite amplitudes are compared to the ground truth of known concentration ratios. Overall estimation accuracy is assessed by linear fits of reported vs. actual ratios and coefficients of determination. Simulations further elucidate the experimental results. Results: Significant differences in reported GABA-to-NAA amplitude ratios are observed. TARQUIN consistently overestimates, while most tools underestimate ratios to varying degrees compared to the ground truth. Underestimationdue to reduced editing efficiency is predicted by simulations. LCModel performs comparatively well for well-resolved solution spectra but struggles for intentionally miscalibrated spectra and gel spectra mimicking in-vivo conditions. GANNET shows better consistency and robustness to calibration errors but greater underestimation related to how the 3 ppm GABA peak is fitted. Surprisingly, simple peak integration with minimal pre-processing yields the most consistent and accurate results compared to the ground truth. Conclusions: A methodological dependence is observed not only in the quantification results for individual spectra, but in GABA-to-NAA gradients across experimental series, systematic offsets and coefficients of determination.