A Deep Biclustering Framework for Brain Network Analysis

Abstract

Brain functional connectivity (FC) analysis has emerged as a compelling quest to understand human brain dynamics and clarify disorder-related aberrations. Typically, FC can be portrayed as a graph of brain components (nodes) and their functional links (edges) known as the brain network (BN). The brain operates as a modular unit, with different regions forming semantically cohesive submodules to execute essential neuronal processing. Identifying these granules can provide insights into the underlying neurobiological mechanisms. Consequently, substantial research efforts have been directed toward clustering the constituents of brain networks. However, the inherent subject heterogeneity in the biological population and the wide spectrum of brain disease manifestation significantly impede cluster generalization. Thus, it often delivers a suboptimal solution and misses insightful nuances of neural systems. Therefore, we propose a deep neural network (DNN) framework for a more granular subgrouping of brain networks by simultaneously stratifying subjects and feature dimensions. The framework adapts discrete learning of BN edges and jointly optimizes instance and feature assignment probability distributions for a novel bicluster retrieval. Extensive experiments on multiple neuroimaging datasets show our model outperforms state-of-the-art biclustering methods. In addition, the extracted biclusters render more modular and semantically meaningful communities in the brain network highlighting significant neuroscientific relevance.