Classification of MCI Brain Network Based on Orthogonal Minimum Spanning Tree

Abstract

Brain network based on resting-state functional magnetic resonance (rs-fMRI) is the most popular method for brain disease diagnosis, which is expected to provide accurate and effective biomarkers. The original fully connected resting-state networks (RSNs) are too dense and must be filtered to get the real network model. In this study, orthogonal minimum spanning trees (OMSTs) was used to filter the connection matrix of 49 age-matched healthy controls (HC) and 50 patients with mild cognitive impairment (MCI). At the same time, we also used global cost efficiency (GCE) algorithm to filter brain network for comparison with OMSTs. We calculated the topological metrices of brain network. Fisher score was used to select features, and the optimal feature subset was used to construct SVM classifier. The classification accuracy of OMSTs was 87%, while that of GCE algorithm was 81%. The experimental results show that the classification accuracy is greatly improved by using OMSTs, which is an effective brain network filtering method.