A time-frequency-coupled deep learning approach for enhancing MRS quantification

Motivation: Current deep learning MRS analysis methods rely solely on time-domain or frequency-domain signals, without utilizing the combined information from both domains. Goal(s): Our objective was to develop a robust and effective time-frequency-coupled deep learning approach for metabolite quantification in MRS. Approach: We proposed a time-frequency-coupled deep learning model using Transformer and CNN(1D) to process time and frequency signals, respectively, for joint metabolite parameter prediction. The model was trained with a physics-informed decoder to reconstruct the signal, enabling a self-supervised learning framework. Results: The proposed framework exhibited lower error (MAE and MAPE) compared to single-domain models and higher robustness across various noise levels. Impact: The time-frequency-coupled deep learning model significantly enhances metabolite quantification accuracy and robustness in MRS , making it a more reliable tool for metabolic analysis, especially under varying noise conditions, compared to conventional single-domain approaches.