- Q-MRS uses a Convolutional vision Transformer combining CNN and Transformer architectures, trained on a large simulated spectral dataset.
- CvT outperformed a simple CNN and an Inception network on simulated data and matched LCModel and Osprey on 3T GABA-edited in vivo fits.
- Q-MRS yields comparable concentration estimates without imposing metabolite amplitude ratio constraints, reducing bias and improving fitting stability.
NMR Biomed. 2026 May;39(5):e70279. doi: 10.1002/nbm.70279.
ABSTRACT
Quantification of magnetic resonance spectroscopy (MRS) data using linear combination modeling (LCM) is challenging, partly due to the large number of spectral parameters to be estimated. Popular conventional LCM approaches often place soft constraints on signal amplitude ratios to improve fitting stability, at the cost of introducing bias. Meanwhile, existing deep learning (DL) methods tend to oversimplify the problem by omitting important spectral parameters, limiting their real-world utility. With these considerations in mind, we developed Q-MRS, a DL framework based on a Convolutional vision Transformer (CvT) that combines the strengths of a convolutional neural network (CNN) and a Transformer. The model was trained on a large dataset of simulated spectra and evaluated on high-quality 3T GABA-edited MEGA-PRESS data acquired from health adults in the medial parietal lobe. On simulated data with known ground-truth metabolite levels, the CvT outperformed two baseline models, a simple CNN and an Inception network. When applied to the in vivo data, Q-MRS produced fit quality and concentration estimates comparable with those of two established LCM methods, LCModel and Osprey, without imposing constraints on metabolite amplitude ratios. These results suggest that the proposed method is a promising approach for MRS analysis.
PMID:41948854 | DOI:10.1002/nbm.70279
Share Evidence Blueprint
Save to Google Notes

Search Google Scholar
⭐ My Revision List

