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

Dynamic deuterium metabolic imaging (DMI) enables time-resolved mapping of cerebral glucose metabolism in vivo, yet its intrinsically low SNR often renders voxel-wise metabolite quantification unstable-particularly at early repetitions. Low-rank denoising is widely used in MR spectroscopic imaging (MRSI)/DMI to improve robustness, but very low-SNR regimes and dynamic studies remain challenging. Self-supervised learning is promising for DMI/MRSI denoising, but its performance depends strongly on the representation domain and the noise assumptions underlying the training objective Here, we study a pragmatic denoising pipeline for dynamic DMI/MRSI that combines a mild low-rank stabilization with self-supervised learning in the spectral-temporal (f×T) domain. Exploiting metabolite-specific spectral structure and redundancy across repeated measurements, our approach relies on two mild assumptions: (i) additive, approximately zero-mean noise and (ii) approximate noise independence across repeated acquisitions. These conditions are expected to be satisfied across essentially any reconstruction and preprocessing pipeline. Our results indicate that the f×T domain is effective both with spatially correlated and approximately uncorrelated noise. We evaluate the approach in simulations and in vivo dynamic DMI data from healthy volunteers (n=6) and a brain tumor patient. The proposed pipeline improves the robustness of time-resolved metabolite estimates and increases LCModel fit stability relative to a state-of-the-art low-rank baseline (tMPPCA), with the largest gains for weak metabolites and early low-SNR repetitions. Together, this enables more reliable dynamic metabolite mapping in low-SNR regimes.