Fast, fully Bayesian spatiotemporal inference for fMRI data.

Area of Science:

• Neuroimaging
• Statistical Modeling
• Computational Neuroscience

Background:

• Functional magnetic resonance imaging (fMRI) data analysis presents computational challenges due to large datasets and complex dependencies.
• Oversimplified models are often used in fMRI analysis, potentially compromising inference quality.
• Efficient and accurate statistical methods are needed for robust detection of brain activation.

Purpose of the Study:

• To develop a computationally efficient spatial Bayesian variable selection method for fMRI data.
• To improve the quality of inference in detecting blood oxygenation level dependent (BOLD) activation.
• To address the computational burden associated with analyzing large fMRI datasets.

Main Methods:

• A spatial Bayesian variable selection approach is proposed, partitioning the brain into 3D parcels for parallel processing.
• Voxel-level activation is modeled using regressions with latent indicators for zero or non-zero changes.
• A sparse spatial generalized linear mixed model (SGLMM) is employed to capture spatial dependencies among indicator variables.

Main Results:

• The proposed parcellation scheme demonstrates robust performance across various realistic simulation scenarios.
• The sparse SGLMM offers significantly more efficient computation compared to traditional spatial models in fMRI.
• Edge effects between parcels were effectively mitigated, ensuring accurate boundary analysis.

Conclusions:

• The developed method provides an efficient and accurate approach for detecting BOLD activation in fMRI data.
• The spatial Bayesian variable selection and parcellation strategy enhance computational feasibility without sacrificing inference quality.
• The methodology was successfully applied to real-world task-based fMRI experimental data.