Unified Brain Network with Functional and Structural Data

Abstract

Brain network analysis has been proved as an effective technology for brain disease diagnosis. For improving diagnosis performance, some efforts have been made to merge functional network and structural network. Whether using single modal or multi-modal data, the construction of brain network plays an important role in the whole diagnosis system. However, the existing multi-modal brain network analysis methods usually construct functional network and structural network separately, in which the complementary information of different modalities is difficult to embed. In this paper, a unified brain network construction algorithm that is jointly learned from both functional and structural data is proposed, and it is applied to epilepsy diagnosis. First, we built a correlation model among all brain regions with functional data by low-rank representation, and simultaneously embed the local manifold with structural data into this model. The constructed network then captures the global brain region correlation by the low-rank constraint and preserves the local structural information by manifold learning. Second, we adaptively estimate the importance of different brain regions by PageRank algorithm. Finally, we utilize a multi-kernel method to fuse the connectivity and node information from the constructed unified network for classification. The proposed method (UBNfs) is evaluated on multi-modal epilepsy dataset, and the experimental results show that our method is effective and can achieve a promising performance in diagnosis of epilepsy.