Fully Bayesian spectral methods for imaging data.

Area of Science:

• Neuroimaging
• Statistical modeling
• Computational neuroscience

Background:

• Medical imaging generates vast, spatially correlated data, posing analytical challenges.
• Existing methods often use approximations or block analyses, limiting scalability and accuracy.
• Accounting for spatial dependence is crucial for precise statistical inference in neuroimaging.

Purpose of the Study:

• To develop a scalable method for analyzing large, multi-subject medical imaging data with spatial correlation.
• To incorporate nonstationarity, functional connectivity, and local signals into the analysis.
• To investigate the association between cortical thickness and Alzheimer's disease.

Main Methods:

• Utilized spectral methods combined with Markov Chain Monte Carlo (MCMC) sampling.
• Developed a novel approach to handle nonstationarity and long-range functional connectivity.
• Applied the method to large-scale, multi-subject neuroimaging datasets.

Main Results:

• Simulated data demonstrated improved precision and valid statistical inference when accounting for spatial dependence.
• The method successfully analyzed large datasets without prohibitive computational cost.
• Identified specific cortical regions with significantly reduced thickness in Alzheimer's disease patients compared to controls.

Conclusions:

• The proposed spectral and MCMC method offers an efficient and accurate approach for analyzing complex medical imaging data.
• This technique enhances statistical power for detecting subtle neuroanatomical changes.
• Findings provide insights into the structural brain changes associated with Alzheimer's disease.