Modular Graph Encoding and Hierarchical Readout for Functional Brain Network Based eMCI Diagnosis

Abstract

The functional brain network, estimated from functional magnetic resonance imaging (fMRI), have been widely used to capture subtle brain function abnormality and perform diagnosis of brain diseases, such as early mild cognitive impairment (eMCI), i.e., with Graph Convolutional Network (GCN). However, there are at least two issues with GCN-based diagnosis methods, i.e., (1) over-smoothed representation of nodal features after using general convolutional kernels, and (2) simple blind readout of graph features without considering hierarchical organizations of brain functions. To address these two issues, we propose a GCN-based architecture (HFBN-GCN), based on the hierarchical functional brain network (defined with priors from brain atlases). Specifically, first, we design a “topology-focused brain encoder” to enhance nodal features by using (1) one branch of GCNs to focus on limited message passing among functional modules of each hierarchical level for alleviating over-smoothing issue and (2) another branch of GCNs to processes whole brain network for retaining original communication of information. Second, we design a “hierarchical brain readout” to utilize pre-defined hierarchical information to guide the coarse-to-fine readout process. We evaluate our proposed HFBN-GCN on the ADNI dataset with 910 fMRI data. Our proposed method achieves 73.4% accuracy (with 77.1% sensitivity and 71.1% specificity) in eMCI diagnosis, where both proposed strategies help boost performance compared to simply-stacked GCNs. In addition, our method suggests the dorsal attention network, saliency network and default mode network as the most crucial functional sub-networks for eMCI identifications. Our method thus is potentially beneficial for both clinical applications and neurological studies.