Scalable Bayesian Image-on-Scalar Regression for Population-Scale Neuroimaging Data Analysis

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Statistical Modeling

Background:

• Bayesian Image-on-Scalar Regression (ISR) offers flexible, uncertainty-aware neuroimaging analysis.
• Large-scale datasets (e.g., UK Biobank) present computational challenges for ISR, particularly with subject-specific brain masks.

Purpose of the Study:

• To propose a novel Bayesian ISR model that efficiently scales to large datasets.
• To accommodate subject-specific brain masks in neuroimaging analyses.
• To improve computational efficiency and statistical power for neuroimaging studies.

Main Methods:

• Developed a Bayesian ISR model using Gaussian process priors with salience area indicators.
• Implemented a scalable posterior computation algorithm with stochastic gradient Langevin dynamics and memory mapping.
• Achieved linear scaling with subsample size and constrained memory usage to batch size.

Main Results:

• Demonstrated a 4- to 11-fold speed increase on UK Biobank task fMRI data (38,639 subjects).
• Showcased an 8-18% enhancement in statistical power compared to traditional Gibbs sampling.
• Identified a subregion of the amygdala with a ~58% decrease in emotion-related activation between ages 50-60.

Conclusions:

• The novel Bayesian ISR model provides an efficient and scalable solution for large-scale neuroimaging analysis.
• The method successfully handles subject-specific brain masks and enhances statistical inference.
• The findings highlight age-related changes in emotion processing within the amygdala.