Comparison of Convolutional Neural Networks Based Method and LCmodel on the Quantification of in Vivo Magnetic Resonance Spectroscopy

Abstract

Quantification of metabolites concentrations in institutional unit (IU) for between subject and long-term comparison is considered important strategy in the applications of magnetic resonance spectroscopy (MRS). The aim of this study is to investigate if metabolite concentrations quantified by convolutional neuronal network (CNN) based method associated with a proposed scaling procedure can reflect variations of the metabolite concentrations in institution unit (IU) at different brain regions with different signal-to-noise-ratio (SNR) and linewidth (LW). An error index based on standard error (SE) is proposed to indicate the confidence levels on the prediction for metabolites. In vivo MRS spectra were collected at 3 brain regions from 44 subjects at 3T system. Metabolite concentrations in IU quantified by LCModel and CNN from 44 subjects were compared. For in vivo spectra characterized under different spectral quality in terms of SNR and LW, line narrowing and noise free spectra were successfully exported by CNN. Concentrations of five metabolites quantified by CNN and LCModel are in similar range with statistically significant Pearson’s correlation coefficients (0.28~0.70). SE of the metabolites show positive correlation with Cramer-Rao lower bound (CRLB) (r=0.60) and with absolute CRLB (r=0.84). In conclusion, the CNN based method with the proposed scaling procedures can be used to quantify in vivo MRS spectra. The concentrations of five major metabolites were reported in IU, which are in the same range as those quantified using a routine MRS quantification procedures by LCModel. The SE can be used as error index indicating predicted uncertainties for metabolites with the information similar to the absolute CRLB.