Bayesian Spatial Blind Source Separation via the Thresholded Gaussian Process.

Area of Science:

• Neuroimaging
• Biomedical Signal Processing
• Statistical Modeling

Background:

• Existing blind source separation (BSS) methods often neglect spatial dependence and sparsity in high-dimensional neuroimaging data.
• This limitation hinders accurate analysis of complex brain signals.

Purpose of the Study:

• To develop a novel Bayesian spatial blind source separation (BSP-BSS) approach tailored for neuroimaging data.
• To address the limitations of existing methods by incorporating spatial dependence and signal sparsity.

Main Methods:

• Proposed a Bayesian nonparametric prior model using thresholded Gaussian processes for sparse, piece-wise smooth latent source signals.
• Utilized von Mises-Fisher (vMF) priors for mixing coefficients.
• Demonstrated theoretical properties including posterior consistency and source number selection consistency.

Main Results:

• BSP-BSS demonstrated superior performance in separating latent brain networks compared to existing methods.
• The approach effectively detected activated brain regions within latent sources.
• Validated through extensive simulations and analysis of resting-state fMRI data from the ABIDE study.

Conclusions:

• The proposed BSP-BSS method offers a robust framework for analyzing spatially dependent neuroimaging data.
• It significantly improves the separation of latent sources and detection of brain activation.
• BSP-BSS provides a valuable tool for advancing neuroimaging data analysis and understanding brain function.