Self-supervised spectral-temporal denoising for dynamic deuterium metabolic imaging

Area of Science:

• Magnetic Resonance Imaging
• Metabolic Imaging
• Neuroimaging

Background:

• Dynamic deuterium metabolic imaging (DMI) provides time-resolved mapping of cerebral glucose metabolism.
• Low signal-to-noise ratio (SNR) in DMI limits voxel-wise metabolite quantification, particularly in early dynamic scans.
• Existing low-rank denoising methods struggle in very low-SNR and dynamic DMI scenarios.

Purpose of the Study:

• To develop and evaluate a pragmatic denoising pipeline for dynamic DMI/MRSI.
• To improve the robustness and stability of time-resolved metabolite quantification in low-SNR DMI.
• To enhance the reliability of dynamic metabolite mapping.

Main Methods:

• A denoising pipeline combining mild low-rank stabilization with self-supervised learning in the spectral-temporal (f×T) domain.
• Exploitation of metabolite-specific spectral structure and redundancy across repeated measurements.
• Assumptions of additive, zero-mean noise and approximate noise independence across acquisitions.

Main Results:

• The f×T domain proved effective for denoising with both spatially correlated and uncorrelated noise.
• The proposed pipeline enhanced the robustness of time-resolved metabolite estimates compared to a state-of-the-art low-rank baseline (tMPPCA).
• Significant improvements were observed for weak metabolites and early low-SNR repetitions, increasing LCModel fit stability.

Conclusions:

• The developed denoising pipeline improves dynamic DMI/MRSI robustness and metabolite quantification stability.
• This approach enables more reliable dynamic metabolite mapping, particularly in challenging low-SNR regimes.
• The method offers substantial gains for weak metabolites and early dynamic scans, advancing in vivo metabolic imaging.