Bayesian posterior estimation for MR spectral quantification using diffusion probabilistic model

Motivation: Conventional MRS parameter estimation methods lack efficient ways to quantify uncertainties with prior knowledge, which is essential for reliable clinical/research applications. Bayesian posterior distribution allows for uncertainty estimation. However, conventional Bayesian inference techniques are slow. Goal(s): Develop a fast, accurate method to approximate MRS parameter posterior distributions, addressing the time and computational limitations of standard Markov Chain Monte Carlo (MCMC) methods. Approach: Train an improved Denoising Diffusion Probabilistic Model (iDDPM) on explicit prior to infer posterior distributions of MRS parameters, leveraging deep learning fast inference speed. Results: The proposed method provided accurate, and precise posterior distribution estimates, over 23x faster than MCMC. Impact: This study enables fast and reliable MRS parameters value and uncertainty assessment, significantly accelerating conventional methods. Patient diagnosis using this approach could benefit from precise metabolite assessment with informed margin of error, potentially broadening MRS clinical utility.