A Manifold Learning-based Approach for Denoising in Deuterium Metabolic Imaging

Motivation: Conventional deuterium metabolic imaging (DMI) suffers from a low signal-to-noise ratio (SNR), which imposes challenges such as long acquisition time, low spatial resolution, or poor accuracy in metabolite concentration estimation. Goal(s): Our goal is to improve the SNR of the DMI signals and thus enhance the estimation of metabolite concentration. Approach: A manifold learning-based method, Linear Tangent Space Alignment (LTSA) model, was proposed to denoise the DMI signal. Results: Results from theoretical calculation, numerical simulation and in vivo study showed that the proposed method provided a reduction in noise, and in the variance of the metabolite concentration estimates. Impact: The LTSA model reduces the noise in DMI signal and thus improves the estimation of metabolite concentration. This improvement prospectively allows DMI with high spatial-temporal resolution, which can assist tumor diagnosis and treatment response assessment in clinical settings.