A deep learning approach to multi-fiber parameter estimation and uncertainty quantification in diffusion MRI.

Area of Science:

• Neuroimaging
• Biophysics
• Computational Neuroscience

Background:

• Diffusion MRI (dMRI) is crucial for in vivo brain microstructure analysis.
• Current dMRI parameter inference faces challenges like variable dimensionality, low signal-to-noise, and non-linear models.
• Existing methods often use simplified, biologically implausible models to ensure stable estimation.

Purpose of the Study:

• To develop a novel sequential method for multi-fiber parameter inference in dMRI.
• To address the challenges of complex inverse problems in dMRI analysis.
• To enable scalable parameter estimation and uncertainty quantification.

Main Methods:

• A sequential approach decomposing the inference into subproblems.
• Utilizing deep neural networks tailored for specific structures and symmetries.
• Training neural networks via simulation for efficient parameter estimation.
• Amortized inference for scalable computation and uncertainty quantification.

Main Results:

• The novel method demonstrates advantages over standard alternatives on simulated and Human Connectome Project (HCP) data.
• Analysis under HCP-like acquisition shows high uncertainty in extracellular parallel diffusivity estimates.
• Intracellular volume fraction estimates exhibit relatively high precision.

Conclusions:

• The proposed deep learning-based sequential method offers a robust and scalable solution for multi-fiber dMRI parameter inference.
• The findings highlight the precision of intracellular volume fraction estimation and the uncertainty in extracellular diffusivity under specific acquisition conditions.
• This work advances the analysis of brain microstructure using diffusion MRI.