Likelihood-free posterior estimation and uncertainty quantification for diffusion MRI models.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Biophysics

Background:

• Diffusion magnetic resonance imaging (dMRI) is crucial for estimating brain tissue microstructure and white matter connectivity (tractography).
• Accurate model parameter estimation is vital for inferring biophysical tissue properties and fiber orientations.
• Current dMRI models often lack uncertainty quantification for parameter estimates, impacting downstream measure reliability.

Purpose of the Study:

• To develop a deep learning algorithm for identifying crossing fibers in each voxel.
• To propose a robust likelihood-free deep learning method for estimating multi-fiber dMRI model parameters and their posterior distributions.
• To quantify the uncertainty in model parameters and derived dMRI measures.

Main Methods:

• A novel deep learning algorithm was designed to determine the number of crossing fibers per voxel.
• A likelihood-free deep learning approach was employed to estimate multi-fiber model parameters and their full posterior distributions.
• Synthetic and in-vivo data were used for quantitative validation across various noise levels and test samples.

Main Results:

• The proposed method accurately estimates the number of crossing fibers, fiber orientation, and tensor eigenvalues with lower error than existing methods.
• The approach provides a full posterior distribution for model parameters, enabling robust uncertainty quantification.
• The deep learning method is computationally efficient, requiring significantly less time than traditional nonlinear fitting techniques.

Conclusions:

• The developed deep learning methodology offers a robust and computationally fast solution for dMRI model parameter estimation and uncertainty quantification.
• This approach improves the accuracy of derived dMRI microstructural measures by accounting for estimation uncertainty.
• The generalizable methodology can be applied to various dMRI models for enhanced neuroimaging analysis.