Comparison of Low‐Rank Denoising Methods for Dynamic Deuterium MRSI at 7 T

Dynamic deuterium (²H)-MRSI enables mapping of metabolic fluxes in vivo, but its sensitivity is hampered by the low ²H gyromagnetic ratio and ²H-labelled metabolite concentrations. Low-rank denoising can enhance MRSI sensitivity by separating signal from noise. Several methods have been proposed, but the optimal approach for dynamic ²H-MRSI remains unclear. This work compares six low-rank denoising methods for dynamic ²H-MRSI: four variations of spatiospectral partial separability (PS)-Global PS, Local PS, Stacked PS and SPectral and temporal INtegration for SVD (SPIN-SVD)-as well as global-local higher order singular value decomposition (GL-HOSVD) and tensor Marchenko-Pastur principal component analysis (tMPPCA). Performance was evaluated using realistic ²H-MRSI brain simulations across a range of noise and B₀ inhomogeneity levels, including a lesion model featuring focal metabolic alteration, and in vivo data. For simulations at the in vivo SNR and B₀ level, the concentration root-mean-square errors (RMSE) of the fitted metabolite maps against gold standard maps were reduced by 29.3%, 24.4%, 20.2%, 33.4%, 23.5% and 21.9% following denoising with Global PS, Local PS, Stacked PS, SPIN-SVD, GL-HOSVD and tMPPCA, respectively. Across all simulated noise levels and all but the highest B₀ inhomogeneity level, SPIN-SVD achieved the lowest concentration RMSE and best preserved spatial distributions of ²H-labelled water (HDO) and lactate (Lac), including lesion-associated changes. While Global PS and Stacked PS reduced noise variance, they failed to preserve signal variations in simulated and in vivo data. SPIN-SVD, GL-HOSVD and tMPPCA reduced noise in vivo while maintaining spatial and temporal metabolite variations. GL-HOSVD and tMPPCA performed similarly, with tMPPCA preferred for its computational efficiency. Both SPIN-SVD and tMPPCA are suitable for denoising dynamic ²H-MRSI, with SPIN-SVD preferable for clinical applications owing to its simple implementation and superior preservation of local metabolic alterations, and tMPPCA better suited for absolute quantification of very low SNR metabolites.