Constructing Multi-scale Connectome Atlas by Learning Graph Laplacian of Common Network

Abstract

Recent development of neuroimaging and network science allow us to visualize and characterize the whole brain connectivity map in vivo. As the importance of volumetric image atlas, a common brain connectivity map (called connectome atlas) across individuals can offer a new window to understand the neurobiological underpinning of cognition and behavior that are related to brain development or neuro-disorders. However, a major obstacle to the application of classic atlas construction methods in the setting of brain network is that the region-to-region connectivity, often encoded in a graph, does not exactly comply with the Euclidean space that has widely been used for the regular data structure such as grid. To address this challenge, we first turn the brain network (encoded in the graph) into a set of graph signals in the Euclidean space via the diffusion mapping technique. Furthermore, we cast the construction of connectome atlas into a learning-based graph inference model that simultaneously (1) aligns all individual graph signals to a common space spanned by the graph spectrum bases of the latent common network, and (2) learns graph Laplacian of the common network that is in consensus with all aligned graph signals. We have evaluated our novel connectome atlas method with the comparison to the counterpart non-learning based methods in analyzing the brain networks for both neurodevelopmental and neurodegenerative diseases, where our proposed learning-based method shows more reasonable results in terms of accuracy and replicability.