SEMI-SUPERVISED LEARNING OF BRAIN FUNCTIONAL NETWORKS

Identification of subject-specific brain functional networks of interest is of great importance in fMRI based brain network analysis. In this study, a novel method is proposed to identify subject-specific brain functional networks using a graph theory based semi-supervised learning technique by incorporating not only prior information of the network to be identified as similarly used in seed region based correlation analysis (SCA) but also background information, which leads to robust performance for fMRI data with low signal noise ratio (SNR). Comparison experiments on both simulated and real fMRI data demonstrate that our method is more robust and accurate for identification of known brain functional networks than SCA, blind independent component analysis (ICA), and clustering based methods including Ncut and Kmeans.