Guiding fusion of dynamic functional and effective connectivity in spatio-temporal graph neural network for brain disorder classification

Dongdong Chen, Mengjun Liu, Sheng Wang, Zheren Li, Lu Bai, Qian Wang, Dinggang Shen, Lichi Zhang

Abstract: Brain connectivity pattern describes complex information in the brain network, which is widely applied for understanding brain connectome and diagnosing neurological diseases. Researchers have investigated brain network analysis using functional MRI (fMRI) from broadly two different perspectives: functional connectivity (FC), which relies on statistical independence and is typically evaluated using correlations, and effective connectivity (EC), which is based on directional causal influences among brain regions. Thus, the fusion of FC and EC can further extract more comprehensive information for characterizing brain abnormalities. However, most current brain network analysis methods focus on either FC or EC. To address this problem, we propose a novel Spatio-Temporal Graph Neural Network named as DCSTN to fuse dynamic functional and effective connectivity networks in the feature space, and construct brain network embeddings in a comprehensive manner. First, we introduce dynamic FCs and ECs to simultaneously model the brain network from different perspectives. Then, we employ message passing-based spatial graph convolution to analyze spatial characteristics of the brain network in discrete time segments. Finally, we introduce a novel fusion module based on a cross-attention mechanism, which incorporates continuous temporal information to extract brain embeddings for the purpose of identifying brain disorders. The whole framework utilizes the causal linkage of dynamic ECs during time evolution to guide the fusion of discrete FC networks. Qualitative and quantitative experimental results from public datasets validate the effectiveness of fusing FC and EC, and the proposed DCSTN outperforms state-of-the-art methods in different types of brain disorder classification.