Bayesian Lesion Estimation with a Structured Spike-and-Slab Prior.

Area of Science:

• Neuroimaging
• Statistical Modeling
• Computational Biology

Background:

• White matter lesions on T2-weighted MRI scans indicate neural demyelination and brain damage.
• Population-level modeling of these lesions is crucial for understanding aging and inflammatory diseases.

Purpose of the Study:

• To introduce a scalable hierarchical Bayesian spatial model (BLESS) for analyzing binary neuroimaging data.
• To enable accurate modeling of brain lesions at a population level.

Main Methods:

• Developed BLESS, a Bayesian spatial model using spike-and-slab priors for binary responses.
• Employed mean-field variational inference with dynamic posterior exploration for scalability.
• Incorporated Bayesian bootstrap for accurate posterior variance estimation and cluster-based statistics.

Main Results:

• The BLESS model scales effectively to large sample sizes (e.g., 40,000 subjects).
• Achieved accurate uncertainty quantification and generated novel cluster size imaging statistics.
• Validated through simulations and application to UK Biobank data.

Conclusions:

• BLESS provides a robust and scalable framework for analyzing population-level brain lesion data.
• The model facilitates advanced imaging statistics for enhanced understanding of neurological conditions.
• Demonstrated utility in large-scale studies like the UK Biobank.