Enhancing fMRS Data Quality and GABA Quantification Using SpectroViT: A Deep Learning Reconstruction Approach

GABA-edited MEGA PRESS acquisitions are time-consuming and rely on repeated acquisitions (transients) to compose good-quality spectroscopic data. Acquiring Functional Magnetic Resonance Spectroscopy (fMRS) is even more time-consuming and the number of transients must be reduced in order to be practicable. A small number of transients reduces the quality of the spectra, reducing the accuracy of GABA quantification. This article presents the use of a reconstruction Deep Learning (DL) model, SpectroViT, as a way of improving fMRS data quality, allowing a more robust and trustful GABA quantification. Experiments showed that reconstructed spectra presented more consistency in measuring the variation of the GABA metabolite during a visual stimulus of 4Hz, 8Hz, and 16Hz when compared to the acquired spectra (without reconstruction) and that the model was able to generalize to different datasets.