Spline baseline model flexibility independently affects the accuracy and precision of in vivo proton magnetic resonance spectral fitting in a metabolite-specific manner not visually predicted by fit residuals

Abstract

In vivo proton magnetic resonance spectroscopy (1H-MRS) data often exhibit baselines or low-amplitude signal variations resulting from residual water, imperfectly suppressed lipids, low-amplitude metabolites not considered for fitting, and other features not represented in a basis set. While multitudinous approaches exist to model these baselines in 1H-MR spectral analysis, many continue to lack systematic validation against varied and realistic ground-truth standards. Here, we compare the accuracy (error mean) and precision (error standard deviation) of metabolite scaling estimates by linear combination modeling (LCM) spectral fitting accounting for spectral baselines via smoothed cubic splines at 50 different combinations of fixed knot interval and smoothing weight, either with or without additionally simulated Gaussian basis signals to separately model spectral macromolecules. Synthesized in-vivo-like metabolite brain spectra incorporating macromolecule signals measured using double-inversion-recovery-prepared sLASER (TE 20.1 ms; TR 2 s; TI1 920 ms; TI2 330 ms) at 3 T from single voxels in the frontal and occipital cortex of 10 healthy volunteers (five female; 23 ± 5 y.o.) provided both in vivo realism and a standard ground truth for error calculation. Optimal baseline flexibility differed both by definition of “optimum” as either accuracy or precision and by metabolite. Regardless of definition or metabolite, optimal models were not those yielding the smallest fit residuals. Optimized spline baseline definitions yielded high accuracies (lowest mean error −0.003 ± 2.1% for total N-acetyl aspartate and highest mean error 10.1 ± 19.2% for glutamate + glutamine within fits including macromolecule bases) as well as comparable precision for most metabolites to fits achieved in LCModel; inclusion of simulated macromolecules in baseline models improved maximum fit precision but not accuracy. Taken together, these data illustrate that optimized spline baseline model flexibility exhibits metabolite-specific relationships with 1H-MR spectral quantification accuracy or precision not readily predicted by visual inspection of associated fit residuals and not necessarily improved by adaptive relative to absolute constraints.

Abstract (Swedish)

In vivo proton magnetic resonance spectroscopy (1H-MRS) data often exhibit baselines or low-amplitude signal variations resulting from residual water, imperfectly suppressed lipids, low-amplitude metabolites not considered for fitting, and other features not represented in a basis set. While multitudinous approaches exist to model these baselines in 1H-MR spectral analysis, many continue to lack systematic validation against varied and realistic ground-truth standards. Here, we compare the accuracy (error mean) and precision (error standard deviation) of metabolite scaling estimates by linear combination modeling (LCM) spectral fitting accounting for spectral baselines via smoothed cubic splines at 50 different combinations of fixed knot interval and smoothing weight, either with or without additionally simulated Gaussian basis signals to separately model spectral macromolecules. Synthesized in-vivo-like metabolite brain spectra incorporating macromolecule signals measured using double-inversion-recovery-prepared sLASER (TE 20.1 ms; TR 2 s; TI1 920 ms; TI2 330 ms) at 3 T from single voxels in the frontal and occipital cortex of 10 healthy volunteers (five female; 23 ± 5 y.o.) provided both in vivo realism and a standard ground truth for error calculation. Optimal baseline flexibility differed both by definition of “optimum” as either accuracy or precision and by metabolite. Regardless of definition or metabolite, optimal models were not those yielding the smallest fit residuals. Optimized spline baseline definitions yielded high accuracies (lowest mean error −0.003 ± 2.1% for total N-acetyl aspartate and highest mean error 10.1 ± 19.2% for glutamate + glutamine within fits including macromolecule bases) as well as comparable precision for most metabolites to fits achieved in LCModel; inclusion of simulated macromolecules in baseline models improved maximum fit precision but not accuracy. Taken together, these data illustrate that optimized spline baseline model flexibility exhibits metabolite-specific relationships with 1H-MR spectral quantification accuracy or precision not readily predicted by visual inspection of associated fit residuals and not necessarily improved by adaptive relative to absolute constraints.

• author: Swanberg, Kelley M. ; Gajdošík, Martin ; Landheer, Karl ; Treacy, Michael S and Juchem, Christoph
• publishing date: 2025-03-04
• type: Contribution to journal
• publication status: published
• subject: Medical Imaging
• in: NMR in Biomedicine
• volume: 38
• issue: 4
• article number: e70010
• publisher: John Wiley & Sons Inc.
• external identifiers: pmid:40040376; scopus:86000111305
• ISSN: 0952-3480
• DOI: 10.1002/nbm.70010
• language: English
• LU publication?: no
• id: 463d9070-e79d-406b-986f-eef7808f8cbe
• alternative location: https://analyticalsciencejournals.onlinelibrary.wiley.com/doi/full/10.1002/nbm.70010
• date added to LUP: 2025-03-07 10:44:44
• date last changed: 2025-04-23 04:01:04