Model-informed unsupervised deep learning approaches to frequency and phase correction of MRS signals

Area of Science:

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

Background:

• Supervised deep learning (DL) for MRS frequency and phase correction (FPC) shows promise but requires labeled data, which is difficult to obtain.
• Unsupervised DL offers a potential solution to overcome data labeling challenges in MRS FPC.

Purpose of the Study:

• To investigate the feasibility and efficiency of unsupervised deep learning-based FPC for MRS data.
• To develop and evaluate novel DL-based FPC methods incorporating physics domain knowledge.

Main Methods:

• Two unsupervised DL methods were developed: DL-based Cr referencing and DL-based spectral registration.
• Networks were trained and validated using simulated, phantom, and in vivo MEGA-edited MRS data.
• Performance was compared against existing FPC methods using a novel evaluation metric, with assessments at varying SNR levels.

Main Results:

• Unsupervised DL methods demonstrated comparable FPC performance to existing methods, even with low SNR data.
• DL-based spectral registration achieved top performance on phantom data.
• Successful application to GABA-edited in vivo MRS data was shown, with significant potential for reduced computation time.

Conclusions:

• Physics-informed, unsupervised deep neural networks can efficiently perform FPC on large MRS datasets.
• These methods offer a faster and more accessible approach to MRS data correction.