The impact of training data on MRS metabolite quantification with deep learning

We investigate the impact of training data quality on deep learning models for metabolite quantification in MRS. Our focus is on two key aspects of simulated training datasets: the variability of metabolite models and the realism of the noise model. Our results demonstrate that the training dataset, particularly the choice of noise model, impacts quantification performance, highlighting the importance of realistic simulations. Our results are limited to a few cases but clearly indicate potential for investigating the training datasets. Focusing on improving the realism of simulations or obtaining large real datasets may yield substantial improvements in quantifying metabolites in MRS spectra.