Interpretable Siamese Graph Contrastive Networks for Brain Disease Diagnosis
In recent years,many studies have used rs-fMRI to construct dynamic functional connectivity network and applied it to the diagnosis of brain diseases such as epilepsy and schizophrenia.The results show that graph convolution network(GCN)can effectively preserve graph structure and extract higher-order features of brain network.However,the existing models using GCN model to diagnose brain diseases usually do not update the graph structure during the training process,so the model effect is often constrained by the quality of the input graph.In addition,the lack of explainability limits the wide application of GCN model in disease diagnosis.In this paper,we propose an interpretable Siamese graph contrastive network to extract spatio-temporal information from the brain network.Specifically,the interpretable module dynamically adjusts the graph structure of the input model by combining the correlation information between brain regions through graph generation network.The graph contrast module learns global and local information of brain network to obtain discriminative feature embedding.Siamese network module takes paired subject graph as model input to alleviate the problem of small data subject size.All modules in our method are optimized together to extract temporal and spatial features of brain networks at the individual level for reliable disease diagnosis.Experimental results show that the proposed method achieves promising diagnostic performance on epilepsy datasets and provides explainable results.
万方数据
维普
