WAND: Wavelet Analysis-Based Neural Decomposition of MRS Signals for Artifact Removal.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Spectroscopy

Background:

• Magnetic Resonance Spectroscopy (MRS) quantification is hindered by low signal-to-noise ratio (SNR), overlapping signals, and artifacts.
• Unparameterized baseline effects are a significant challenge, obscuring low-concentration metabolites and reducing MRS reliability.

Purpose of the Study:

• To introduce Wavelet Analysis-based Neural Decomposition (WAND), a novel data-driven method for decomposing MRS signals.
• To improve the accuracy of metabolite quantification in MRS by effectively separating signals, baseline, and artifacts.

Main Methods:

• WAND utilizes the continuous wavelet transform to enhance component separability in the wavelet domain.
• A U-Net neural network predicts masks for wavelet coefficients, isolating metabolite signals, baseline, and artifacts.
• An artifact mask is generated by inverting known signal masks, allowing for the removal of unpredictable artifacts.

Main Results:

• Numerical evaluations with simulated spectra demonstrate WAND's accurate signal decomposition capabilities.
• WAND significantly improves quantification accuracy when used with linear combination model fitting by effectively removing artifacts.
• The method's robustness is validated using data from the 2016 MRS Fitting Challenge and in vivo experiments.

Conclusions:

• WAND offers a robust and effective solution for decomposing complex MRS signals.
• The method enhances metabolite quantification accuracy by addressing baseline and artifact challenges.
• WAND shows promise for improving the reliability and applicability of MRS in various research and clinical settings.