Bayesian Spatial Variable Selection Methods for Improved Detection of Neural Activation in fMRI

• Activity Number: 135 - Applications of Machine Learning Methods to Imaging Data Analysis
• Type: Contributed
• Date/Time: Monday, July 29, 2019 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistics in Imaging
• Author(s): Somak Dutta* and Ranjan Maitra
• Companies: Iowa State University and Iowa State University
• Keywords: Potts process; Swendsen-Wang algorithm; False belief; conditional maximum a posteriori probability; BOLD contrast

Abstract

Functional Magnetic Resonance Imaging (fMRI) has been widely adopted as a non-invasive technique for understanding spatial localization of neural associates of human cognitive and motor functions. In this talk we propose a whole-brain 3D analysis for identifying regions with high neural activities associated with external stimuli. We follow a hierarchical Bayesian route with a three states 3D Potts model on the latent activation classes and a spike-and-two-slabs prior on the activation signals. We incorporate the prior information on activation proportion in the Potts prior by suitably selecting the hyper parameters. We develop a fast matrix-free computational framework for selecting the activated voxels. Via extensive simulation studies we also demonstrate the robustness of our method near the critical temperature of the Potts prior model. We illustrate our methodology on a sport imagination experiment.