Brain Functional Connectivity Analysis Based On Graph Convolutional Network

The early diagnosis of neurological disorders is a challenging task.Based on brain connectivity networks,more and more studies have used deep learning methods for disease classification in recent years.However,existing models still suffer from insufficient interpretability due to the black-box representation.Features extracted from them often can not correspond to real brain regions and therefore can not be useful biomarkers.Given that Graph Convolutional Network(GCN)has demonstrated superiority in learning discriminative representations of brain connectivity networks,in this paper,we propose two interpretable models that can construct brain connectivity networks from functional magnetic resonance imaging(fMRI)data for disease classification and biomarker extraction.(1)An invertible dynamic graph convolutional network(IDGCN)model which can extract brain connections that are important for disease classification as biomarkers.The model incorporates the functional,spatial,and temporal information of the brain connectivity networks and uses the invertible network to select interpretable biomarkers by reconstructing the extracted disease-related features back to the original connectivity graph.In addition,a random forest is adopted for feature pre-screening in order to reduce the redundancy of the data,which improves classification performance.Experimental results on the multi-center Autism Brain Imaging Data Exchange(ABIDE)dataset and Parkinson’s Disease(PD)dataset prove that our IDGCN model achieves superior disease classification performance and can extract meaningful biomarkers.These findings are of great potential in studying brain-related disorders and auxiliary diagnosis.(2)A graph convolutional network model based on self-adaptive fusion(SAfusion)which can obtain hierarchical and meaningful fused brain regions as biomarkers through the pooling layer.In this model,a pooling layer with the advantages of the selection method and fusion method is added between the convolution layers of IDGCN,in which the fusion matrix is designed based on biomedical knowledge.Therefore,the output nodes of the pooling layer not only retain the segmented brain regions that are important for the classification of brain-related disorders but also integrate the remained brain regions into hierarchical nodes of biological significance.Experimental results on ABIDE dataset and PD dataset show that SAfusion shows better performance than the most advanced graph pooling method in disease classification.The extracted hierarchical brain areas with symmetry have clearer biological significance and have never been found in other models.