MegaNet: Physical Intelligent Metabolite Quantification from Edited Magnetic Resonance Spectra with Synthetic Data Learning

Motivation: To accurately quantify Γ-Aminobutyric acid (GABA) and other metabolites useful for diagnosis of neurodegenerative diseases from edited magnetic resonance spectra. Goal(s): Our goal is to illustrate the accuracy and robustness of our proposed neural network, MegaNet. Approach: The analyses of Bland-Altman were performed on the quantification results between MegaNet and LCModel, and between the quantification results of MegaNet on data of low quality and ordinary quality. Results: MegaNet has good consistency with LCModel and shows robustness to noise-corrupted data. Impact: It shows that synthetic data learning with physical prior knowledge is probably a reliable method to address the problem that AI training of magnetic resonance spectroscopy (MRS) lacks abundant high-quality data.