Magnetic resonance spectroscopic imaging (MRSI) enables non-invasive quantification of metabolite concentrations in vivo. Proton (1H) MRSI is relatively sensitive, but has difficulties discriminating metabolites with similar chemical shifts. In contrast, phosphorus (31P) MRSI shows larger chemical shift dispersion, but is less sensitive. Both techniques provide mutual information that can be exploited by exploring their correlations. In this work, a novel framework for such an exploration was developed with the initial aim to advance 1H MRS (I) quantification using 31P MRS (I) as ground truth, employing sequential multiparametric MR protocols and machine learning analysis. This framework was applied to create models that robustly discriminate glycerophosphocholine (GPC) and phosphocholine (PC) in 1H spectra for the cases of synthetic spectra (nₜrained = 160, 000), acquired spectra of model solutions at B0 = 9. 4T (nₜrained = 5, 760), and MRSI from healthy volunteers at B0 = 7T (nₜrained = 1, 521). In all cases, the created interpretable, non-linear gradient boosting model outperformed the state-of-the-art method LCModel, e. g. , for synthetic spectra, with a mean absolute percentage error about 7 times lower. Feature analysis identified not only spectral features around the total choline resonance as important for the GPC/PC discrimination, but in vivo also spectral features from other metabolites, like glutamine/glutamate, not known before. In conclusion, the framework proved to enable novel applications, and might ultimately pave the way for approaches overcoming limitations of individual in vivo MRSI methods by combining their strengths.
Justyna Platek (2026) studied this question.