Fitting algorithms and baseline correction influence the results of non‐invasive in vivo quantitation of 2‐hydroxyglutarate with ¹H‐MRS

Abstract

1H‐MRS enables non‐invasive detection of 2‐hydroxyglutarate (2‐HG), an oncometabolite accumulating in gliomas carrying mutations in the isocitrate dehydrogenase (IDH) genes. Reliable 2‐HG quantitation requires reproducible post‐processing, deployment of fitting algorithms and quantitation methods. We prospectively enrolled 38 patients with suspected or recently diagnosed gliomas (IDH mutated n = 26). The MRI protocol included a 1H single voxel PRESS sequence with volumes of usually 8 mL or more (20 × 20 × 20 mm3) at TE = 97 ms and 180° pulse spacing. Our aim was to evaluate the reliability of 2‐HG quantitation comparing two frequently used software tools and their respective options of baseline correction (jMRUI with the time domain methods AQSES and QUEST, and LCModel, which analyzes the frequency domain data). For AQSES, degrees of freedom for baseline correction constrains were varied. For LCModel, baseline correction was obtained with and without correction of the unknown background term (predefined macromolecules, lipids). Tissue concentrations were calculated based on the phantom replacement method. Quantitation of 2‐HG levels showed similar mean 2‐HG tissue concentrations for IDH mutated tumors (2.65mM, range 3.06–2.20) for all methods. Bland–Altman plots (difference plots) did not reveal a systematic bias (fixed bias) for any of the algorithms tested, and we were able to show a significant correlation regarding 2‐HG concentration at the same echo time with few statistical outliers (parametric correlation). However, evaluation of outliers suggested that in vivo quantitation of 2‐HG is affected not only by the fitting domain (time or frequency), but also by the baseline correction, which is a major contributing factor to the result of 2‐HG fitting. Clinical application of 2‐HG quantitation as a prognostic or predictive biomarker, particularly in multicenter trials, requires standardized use of fitting methods and baseline correction procedures.

Non-Invasive Prediction of IDH Mutation in Patients with Glioma WHO II/III/IV Based on F-18-FET PET-Guided In Vivo 1H-Magnetic Resonance Spectroscopy and Machine Learning

“…Longer echo times (significant increased sensitivity and specificity in TE 97 ms compared to TE 30-35 ms) [30], ultra-high-field MRI (7T) [31,32], 2D correlation spectroscopy (COSY) at 3T [33] and 2D L-COSY at 7T [34] were used to improve detection of the 2-HG signal. Different post-processing techniques have also been applied in various settings, mostly in a limited number of patients (e.g., 38 patients in Reference [35]). Interestingly, independent of 2-HG, peak alterations of other metabolites were frequently seen (e.g., Cho, Glu, Gly, glutathione, cystathionine [36][37][38]) and sometimes used to improve detection of IDH mutation status, for example, combination of 2-HG and Glu levels [39] The described techniques are frequently not suitable for clinical routine as they need extensive technical and personal capacities [24].…”

Early Noninvasive Metabolic Biomarkers of Mutant IDH Inhibition in Glioma

“…Our study benefited from the use of a high field 14.1 T animal scanner, which allowed us to distinguish the different metabolites and probe their modulation with treatment. In the clinical setting MR scanners are typically at the lower fields of 1.5 T or 3 T. However, as mentioned above, sequences specifically optimized to probe for 2-HG and Glu or GLX have been developed and can be used to monitor the metabolic changes associated with treatment [23,26,32,41,42]. Thus, our observations are also relevant to the clinical setting.…”

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

“…In order to understand the substantial variability introduced by the choice of analysis tool, the influence of modelling strategies and parameters on quantitative results needs to be better understood. Previous investigations have shown that, within a given LCM algorithm, metabolite estimates can be affected by the choice of baseline knot spacing 37,38 , the modelling of MM and lipids 37,39, and SNR and linewidth 40–43 . In this study, we focused on the comparison of each LCM with their default and commonly used parameters, and observed differences resulting both from the default parameters and from differences in the core algorithm.…”