Comparison of MRI-derived vs. traditional estimations of fatty acid composition from MR spectroscopy signals.

ABSTRACT

INTRODUCTION: The composition of fatty acids in the body is gaining increasing interest, and can be followed up noninvasively by quantitative magnetic resonance spectroscopy (MRS). However, current MRS quantification methods have been shown to provide different quantitative results in terms of lipid signals, with possible varying outcomes for a given biological examination. Quantitative magnetic resonance imaging using multigradient echo sequence (MGE-MRI) has recently been added to MRS approaches. In contrast, these methods fit the undersampled magnetic resonance temporal signal with a simplified model function (expressing the triglyceride [TG] spectrum with only three TG parameters), specific implementations and prior knowledge. In this study, an adaptation of an MGE-MRI method to MRS lipid quantification is proposed. METHODS: Several versions of the method - with time data fully or undersampled, including or excluding the spectral peak T2 knowledge in the fitting - were compared theoretically and on Monte Carlo studies with a time-domain, peak-fitting approach. Robustness, repeatability and accuracy were also inspected on in vitro oil acquisitions and test-retest in vivo subcutaneous adipose tissue acquisitions, adding results from the reference LCModel method. RESULTS: On simulations, the proposed method provided TG parameter estimates with the smallest variability, but with a possible bias, which was mitigated by fitting on undersampled data and considering peak T2 values. For in vitro measurements, estimates for all approaches were correlated with theoretical values and the best concordance was found for the usual MRS method (LCModel and peak fitting). Limited in vivo test-retest variability was found (4.1% for PUFAindx, 0.6% for MUFAindx and 3.6% for SFAindx), as for LCModel (7.6% for PUFAindx, 7.8% for MUFAindx and 3.0% for SFAindx). CONCLUSION: This study shows that fitting the three TG parameters directly on MRS data is one valuable solution to circumvent the poor conditioning of the MRS quantification problem.