The Impact of Training Data on MRS Metabolite Quantification with Deep Learning

Motivation: We aim to improve metabolite quantification in MRS with deep learning by investigating the impact of training data quality on model performance. Goal(s): The goal is to investigate dataset factors, specifically the variability of metabolite models and noise realism, impacting the quantification performance of a deep learning architecture trained on varying datasets. Approach: We evaluate deep learning models for metabolite quantification on experimental phantom spectra that were trained on varying simulated datasets. Results: Results show that training datasets significantly impact quantification performance. Specifically, more realistic noise models yield improvements. Impact: This study underscores the importance of training data quality in deep learning for MRS. By demonstrating the impact of noise model realism, it provides insights for developing more accurate metabolite quantification models, potentially improving clinical diagnosis and monitoring neurological disorders.