LEARNING HIERARCHICAL-ORDER FUNCTIONAL CONNECTIVITY NETWORKS FOR MILD COGNITIVE IMPAIRMENT DIAGNOSIS

Functional connectivity network (FCN) extracted from resting-state fMRI has been widely used for brain disease diagnosis. However, previous FCN-based studies have at least two limitations: 1) The FCN construction procedure is often handcrafted and not optimized for diagnosis tasks; 2) The connectivity is limited to be pair-wise (low-order) and not to capture high-order collective interactions between groups of brain regions. Accordingly, we propose a unified framework to learn both low- and high-order diseased-related FCNs. First, an encoder is designed to extract disease-related features from fMRI signals, based on which disease-related low-order FCN (D-LOFCN, order k = 1) is built. Then, by correlating disease-related correlation profiles from D-LOFCN by the graph attention mechanism, we iteratively construct the disease-related high-order FCNs (D-HOFCNs) at k-order (k > 1). Finally, both D-LOFCN and D-HOFCNs are forwarded into corresponding GNNs for producing the diagnosis. The experiments demonstrate that our method has higher performance over other state-of-the-art methods on mild cognitive impairment diagnosis task.