Abstract
PURPOSE: Using artificial intelligence neural networks to generate a representation that maps the input directly to neurochemical concentrations and metabolite-level average transverse relaxation times (T(2)). METHODS: The proposed model used time-domain JPRESS data as input and was trained to be invariant to phase shifts, frequency offsets, and lineshape variations, using computer-synthesized data without prior knowledge of in vivo metabolite concentration distributions. TE-specific representations were generated using a combination of WaveNet and gated recurrent units (GRUs) and integrated into a unified JPRESS representation. RESULTS: By focusing solely on target metabolite signals, the model effectively filtered out background signals, including spectral artifacts and unregistered metabolites. The predicted concentrations and metabolite-level average T(2) values were consistent with those reported in the literature. The model demonstrated robustness to phase shifts, frequency offsets, and line broadening. Additionally, it was capable of detecting low-concentration neurochemicals, such as gamma-aminobutyric acid (GABA), without spectral editing. CONCLUSION: This study demonstrates that deep learning can be used for automatically quantifying both metabolite concentrations and transverse relaxation times with high practical viability.