Q-MRS: Quantitative Magnetic Resonance Spectral Analysis Using Deep Learning

Area of Science:

• Neuroimaging
• Biophysics
• Computational Biology

Background:

• Quantifying magnetic resonance spectroscopy (MRS) data via linear combination modeling (LCM) is complex due to numerous spectral parameters.
• Conventional LCM methods often use soft constraints, potentially introducing bias.
• Existing deep learning (DL) methods may oversimplify spectral analysis, limiting practical use.

Purpose of the Study:

• To develop a DL framework, Q-MRS, for robust MRS data quantification.
• To combine the strengths of Convolutional Neural Networks (CNNs) and Transformers for improved spectral analysis.
• To evaluate Q-MRS performance against established methods without amplitude ratio constraints.

Main Methods:

• Developed Q-MRS, a DL framework utilizing a Convolutional vision Transformer (CvT).
• Trained the CvT model on simulated MRS spectra.
• Evaluated Q-MRS on in vivo 3T GABA-edited MEGA-PRESS data from healthy adults.

Main Results:

• The CvT model outperformed baseline CNN and Inception networks on simulated data.
• Q-MRS achieved comparable fit quality and concentration estimates to LCModel and Osprey on in vivo data.
• The method provided accurate quantification without requiring metabolite amplitude ratio constraints.

Conclusions:

• Q-MRS demonstrates a promising approach for MRS data analysis.
• The DL framework offers a robust alternative to conventional LCM methods.
• This advancement can enhance the utility and accuracy of MRS in clinical and research settings.