Title: Bayesian image-on-image latent factor models for predicting task fMRI using task-free MRI
Authors: Timothy Johnson - University of Michigan (United States) [presenting]
Cui Guo - University of Michigan (United States)
Jian Kang - University of Michigan (United States)
Abstract: Our brains show different activity during task performance across many behavioral domains. Prevailing thought was that individual differences in brain response were attributed to two factors: differences in gross brain morphology and differences in task strategy and/or cognitive processes. However, recent research suggests that these differences can attributed to task-free MRI. That is, individual differences in task-evoked brain activity are inherent features of individual brains such that can be predicted from task-free MRI. Their model is simply a linear regression of the $z$-score maps on the task-free MRI features. They fit one regression model for each of 50 parcels of the cortex for each individual. We set out to build a more sophisticated statistical model and compare results. We propose an image-on-image Bayesian latent factor regression model. We model the task-evoked maps via basis functions. The low-dimensional representation of the basis parameters is obtained using a sparse latent factor model. Then we use a scalar-on-image regression model to link the latent factors with the task-free maps which are selected using a Bayesian variable selection procedure. Head-to-head comparison shows that our modelling strategy is statistically more efficient than the simple model originally proposed.