Variational inference with ARD prior for NIRS diffuse optical tomography.

Area of Science:

• Biomedical Engineering
• Neuroimaging
• Computational Science

Background:

• Diffuse optical tomography (DOT) is crucial for reconstructing 3-D brain activity images using near-infrared spectroscopy (NIRS).
• DOT is inherently an ill-posed inverse problem, presenting significant reconstruction challenges.
• Existing methods often exhibit high sensitivity to hyperparameter selection.

Purpose of the Study:

• To present an advanced NIRS-DOT method utilizing a hierarchical Bayesian approach.
• To incorporate automatic relevance determination priors and variational Bayes techniques for improved estimation.
• To reduce the dependency of sparse estimation on hyperparameters.

Main Methods:

• Implementation of a hierarchical Bayesian framework for NIRS-DOT.
• Application of automatic relevance determination (ARD) prior for feature selection.
• Utilization of variational Bayes (VB) techniques for efficient inference.

Main Results:

• The proposed method demonstrates reduced dependency on hyperparameters compared to conventional approaches.
• Numerical experiments revealed distinct regions of sparse and dense solutions based on hyperparameter settings.
• Theoretical analysis supported the observed experimental results in simplified scenarios.

Conclusions:

• The developed hierarchical Bayesian NIRS-DOT method offers a more robust approach to brain imaging.
• The findings suggest improved stability and reliability in reconstructing brain activity.
• This method advances the field of neuroimaging by addressing key challenges in DOT reconstruction.