SIMBA: Scalable Image Modeling using a Bayesian Approach, A Consistent Framework for Including Spatial Dependencies...

Area of Science:

• Neuroimaging
• Computational Statistics
• Machine Learning

Background:

• Traditional massive univariate approaches in fMRI analysis often lead to information waste due to ignoring spatial dependencies.
• Bayesian spatial modeling offers a flexible framework for whole-brain fMRI analysis by incorporating these spatial dependencies.

Purpose of the Study:

• Introduce SIMBA (Scalable Image Modeling using a Bayesian Approach) for efficient group-level fMRI analysis.
• Address computational challenges in Bayesian Gaussian process (GP) inference for high-dimensional neuroimaging data.

Main Methods:

• Employ a low-rank kernel approximation for GP inference, enabling projection into a reduced-dimensional subspace.
• Develop efficient Python algorithms for fully Bayesian inference using Gibbs sampling (minutes) and mean-field variational inference (seconds).

Main Results:

• SIMBA outperforms competing methods in estimation accuracy, activation detection sensitivity, and uncertainty quantification, particularly in low signal-to-noise ratio settings.
• Demonstrate SIMBA's scalability and interpretability in large-scale fMRI analyses (NARPS and ABCD studies).

Conclusions:

• SIMBA provides a computationally efficient and statistically robust Bayesian framework for whole-brain fMRI analysis.
• The method effectively captures smooth spatial association patterns, offering improved performance over traditional approaches.