Multi-modal Brain Network Fusion Based on Random Walk-Grassmann Model

Abstract

Brain network plays an important role in the diagnosis of many brain diseases. At present, some related studies are based on the structural or functional connection group of human brain, while others consider the related properties of structural and functional brain networks at the same time. Aiming at the problems of how to dynamically collect richer node interaction information and how to learn more effectively from small samples in the research of brain network fusion, we propose a Random Walk-Grassmann (RW-GM) model to effectively fuse them. Firstly, we obtain the structural connection matrix and the temporal characteristic matrix of the brain from the multi-modal data of each subject. Then, we use random walk algorithm and Grassmann pooling method to integrate the two matrices, in order to integrate the structural connection and the temporal characteristics of the brain, so as to obtain more abundant brain connection information. In order to better carry out small sample learning, we use recursive feature elimination method for feature selection, and put the selected features into support vector machine to get the final classification result. We have carried out four binary classification experiments on ADNI data set, and the classification accuracy is better than that of traditional brain network classification methods.