High-Resolution Deuterium Metabolic Spectroscopy and Imaging with Self-Supervised Deep Denoising

Motivation: Denoising is a crucial step in deuterium metabolic resonance spectroscopy/imaging (DMRS/DMI ). However, existing supervised learning-based methods require large paired data, which are challenging to obtain in low-sensitivity DMRS/DMI. Goal(s): Develop a self-supervised deep denoising method that leverages the data acquisition characteristics of DMRS/DMI, eliminating the need for paired data. Approach: Our method leverages the observation that when target data for training is corrupted samples from the same scene, the neural network can converge to the mean estimator, enabling DMRS/DMI denoising via self-supervised learning. Results: This method significantly enhances the sensitivity of DMRS/DMI, enabling high spatiotemporal resolution DMRS/DMI. Impact: The constructed self-supervised deep denoising method significantly enhances SNR, enabling high spatiotemporal resolution DMRS/DMI.