Robust and Fast Markov Chain Monte Carlo Sampling of Diffusion MRI Microstructure Models.

Area of Science:

• Neuroimaging
• Biophysics
• Computational Neuroscience

Background:

• Diffusion MRI (dMRI) analysis benefits from biophysical multi-compartment models over conventional Diffusion Tensor Imaging (DTI) for enhanced specificity to cellular microstructure.
• Parameter estimation in these models typically uses Maximum Likelihood Estimation (MLE) or Markov Chain Monte Carlo (MCMC) sampling.
• MCMC provides posterior distributions of model parameters, offering insights into parameter uncertainty and correlations, unlike MLE's point estimates.

Purpose of the Study:

• To investigate the performance of MCMC algorithm variations across popular dMRI microstructure models.
• To determine if a single, efficient, and robust MCMC variation can be applied to multiple models.
• To establish guidelines for optimizing MCMC sampling in dMRI microstructure modeling.

Main Methods:

• Utilized an efficient GPU-based implementation to overcome computational time constraints of MCMC sampling.
• Evaluated various adaptive MCMC algorithms, burn-in strategies, initialization methods, and thinning parameters.
• Applied Effective Sample Size theory to establish general targets for sample numbers in dMRI multi-compartment models.

Main Results:

• GPU implementation significantly reduced run times, making MCMC sampling feasible for multi-compartment models.
• Adaptive Metropolis methods, specifically Adaptive Metropolis-Within-Gibbs (AMWG), demonstrated improved MCMC performance.
• Initializing with an MLE estimate requires 100-200 samples for burn-in; thinning is generally not advised.

Conclusions:

• Adaptive Metropolis methods enhance MCMC performance for dMRI microstructure modeling, with AMWG recommended.
• Optimized MCMC protocols include MLE initialization and a burn-in of 100-200 samples.
• A multivariate Effective Sample Size of 2,200 is recommended as a general target for characterizing parameter distributions across models and datasets.