The present study aims to develop a deep learning (DL) model to quantify metabolites. To apply DL to metabolite
quantification using 1H-MRS data, Convolutional autoencoder (CAE) were designed to extract line‐narrowed,
baseline‐removed, and noise-free metabolite spectra for each metabolite. Fifty thousand simulation data
were generated by varying the SNR (4-12), linewidth (6-22 Hz), phase shift (± 5°), and frequency shift (± 5 Hz)
on phantom spectra. The data were divided into 45,000 simulation data for training and 5,000 test data, and the
mean absolute percent errors (MAPEs) were used to evaluate the performance of the CAE. The average MAPE
of the metabolites was 13.64 ± 11.38 %. Fourteen metabolites were within the reported concentration ranges.
These findings showed that the proposed method had similar or improved performance than conventional
methods. The proposed method using DL was the recent and up-to-date quantification one and has clinically
potential applicability.