LDST-Net: A Spatiotemporal Framework Based on Learnable Dynamic Graph for Brain Disease Diagnosis

Using resting-state functional magnetic resonance imaging (rs-fMRI) signal to construct and analyze functional brain networks (FBNs) is an effective method for brain disease diagnosis. The model based on adaptively constructing FBNs is popular for its ability to integrate the construction of graph, feature extraction, and classification into a unified framework for optimization. However, there is a large amount of noise in the rsfMRI signal, which is likely to result in suboptimal FBNs modeling effects if it is not adaptively attenuated. In addition, most existing methods based on learnable graph ignore the dynamic of brain activity and the spatiotemporal characteristics of rs-fMRI signal. To address the above issues, we propose a spatiotemporal brain network analysis framework based on learnable dynamic graph (called LDST-Net). Specially, we use a novel graph encoder module to reduce the abnormal information in the signal data during the generation of dynamic FBNs. Furthermore, we combine Transformer and graph neural networks (GNNs) to obtain hierarchical spatial features, and use attention module to perceive temporal information. Experimental results on ABIDE dataset show that our model not only achieves satisfactory diagnostic performance, but also provides an effective method for biomarker discovery.