MVS-GCN: A prior brain structure learning-guided multi-view graph convolution network for autism spectrum disorder diagnosis
Guangqi Wen,Peng Cao,Huiwen Bao,Wenju Yang,Tong Zheng,Osmar Zaiane
DOI: https://doi.org/10.1016/j.compbiomed.2022.105239
IF: 7.7
2022-03-01
Computers in Biology and Medicine
Abstract:PURPOSE: Recently, functional brain networks (FBN) have been used for the classification of neurological disorders, such as Autism Spectrum Disorders (ASD). Neurological disorder diagnosis with FBN is a challenging task due to the high heterogeneity in subjects and the noise correlations in brain networks. Meanwhile, it is challenging for the existing deep learning models to provide interpretable insights into the brain network. We propose a machine learning approach for the classification of neurological disorders while providing an interpretable framework.METHOD: In this paper, we build upon graph neural network in order to learn effective representations for brain networks in an end-to-end fashion. Specifically, we present a prior brain structure learning-guided multi-view graph convolutional neural network (MVS-GCN), which collaborates the graph structure learning and multi-task graph embedding learning to improve the classification performance and identify the potential functional subnetworks.RESULTS: To demonstrate the effectiveness of our approach, we evaluate the performance of the proposed method on the Autism Brain Imaging Data Exchange (ABIDE) dataset and Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset. The experimental results indicate that our MVS-GCN can achieve enhanced performance compared with state-of-the-art methods. Notably, MVS-GCN achieves an average accuracy/AUC of 69.38%/69.01% on the ABIDE dataset. Moreover, the obtained results from our model show high consistency with the previous neuroimaging derived evidence of within and between-networks biomarkers for ASD. The discovered subnetworks are used as evidence for the proposed MVS-GCN model.CONCLUSION: The proposed MVS-GCN method performs a graph embedding learning from the multi-views graph embedding learning perspective while considering eliminating the heterogeneity in brain networks and enhancing the feature representation of functional subnetworks, which can capture the essential embeddings to improve the classification performance of brain disorder diagnosis. The code is available at https://github.com/GuangqiWen/MVS-GCN.
engineering, biomedical,computer science, interdisciplinary applications,mathematical & computational biology,biology