Abstract
aims: In recent years, the integration of multi-site neuroimaging datasets with deep learning methods has yielded substantial advancements in the realm of autism spectrum disorder (ASD) research. background: However, existing graph convolutional neural networks (GCNs) only aggregate neighbor information at a fixed scale and cannot adjust the scale parameters to aggregate the best information. objective: In this paper, a homogeneous graph wavelet neural network (H-GWNN) model is proposed to learn representations for graph classification in an end-to-end manner. method: Specifically, the heterogeneity of multi-site datasets is processed by the ComBat algorithm, which is based on location and scale effects. Image data and phenotypic data are fused to construct a population graph, and the most discriminative information on the graph is extracted by adjusting the scale in a graph wavelet neural network. result: Experiments on 1035 subjects from the autism dataset ABIDE show that the classification accuracy of H-GWNN on autism is 83.58%, which exceeds that of other related GCN models. conclusion: The findings demonstrate that the proposed model can not only process the heterogeneity problem, but also analyze node features at the optimal scale. It captures discriminative features that identify diseases and improves classification performance for brain disease diagnosis. other: N/A
Published Version
Join us for a 30 min session where you can share your feedback and ask us any queries you have