Frequency and phase correction of J-difference edited MR spectra using deep learning.

Area of Science:

• Magnetic Resonance Spectroscopy (MRS)
• Artificial Intelligence in Medical Imaging

Background:

• MEGA-edited MRS data often requires frequency-and-phase correction (FPC) for accurate analysis.
• Conventional FPC methods like spectral registration (SR) can be time-consuming and may fail with significant spectral distortions.

Purpose of the Study:

• To evaluate the efficacy of a deep learning (DL)-based approach for FPC of MEGA-edited MRS data.
• To compare the performance of DL-FPC against conventional spectral registration (SR) and model-based SR (mSR).

Main Methods:

• Two neural networks were developed for frequency and phase correction, trained on simulated MEGA-edited MRS data.
• The DL-FPC method was validated using simulated data and tested on in vivo datasets, with and without artificial offsets.
• Performance was compared to SR and mSR in terms of accuracy and computation time.

Main Results:

• DL-based FPC achieved high accuracy on simulated data (within 0.03 Hz frequency, 0.4° phase).
• DL-FPC performed comparably to SR on unmanipulated in vivo data but was more robust to larger artificial offsets.
• DL-FPC and mSR demonstrated significantly shorter computation times than SR for distorted spectra.

Conclusions:

• Deep learning provides a viable and efficient method for frequency-and-phase correction in MEGA-edited MRS data.
• DL-FPC represents a promising proof of principle for advanced preprocessing of MRS data.