Motivation: Dynamic deuterium (2H) MRSI is a promising technique for mapping metabolic fluxes. However, achieving adequate spatiotemporal resolution is challenging due to its inherently low SNR. Goal(s): This work compares four low-rank denoising strategies for dynamic 2H-MRSI: Spatiotemporal-based low-rank approximation (ST), global-local higher-order singular value decomposition (GLHOSVD), and Mixed and Stacked ST, which reshape the 5D array into a matrix before denoising. Approach: Each denoising method was tested on simulated and in vivo dynamic 2H-MRSI brain data. Results: All methods reduced noise variation in the metabolite maps. GLHOSVD and Mixed denoising mostly preserved regional physiological variations, whereas ST and Stacked ST substantially degraded them. Impact: Our work helps identify a denoising algorithm for dynamic 2H-MRSI that achieves high denoising performance while preserving regional variations. The achieved reductions in metabolite concentration uncertainty may enable high-resolution, whole-brain dynamic 2H-MRSI, to investigate oxidative and non-oxidative human brain metabolism.
Duguid et al. (Tue,) studied this question.