A Novel Method to Preserve Scale-Free Property for the Inference of Dynamic Effective Connectivity Networks From fMRI

Abstract

The study of dynamic effective connectivity (dEC) has demonstrated its importance in contemporary neuroscience research. Though it is now widely understood that brain networks have scale-free property, it has been seldom considered by conventional dEC inferring methods. In this work, we pro-pose a new method to employ a group-wise penalty together with spatial sparsity and temporal smoothness regularizations (GSTR) to preserve scale-free property for the inference of dEC networks from functional magnetic resonance imaging (fMRI). It employs a time-varying vector autoregressive mod-el to encode the network adjacency matrices. The proposed group-wise regularization can preserve the connectivities of potential hubs by grouping them together. We further propose an effective algorithm based on the augmented Lagrangian multiplier to deal with multiple regularizations problem. The efficacy of the GSTR method is validated using a variety of synthetic datasets and an open-source block design visual task-related experiment.