Identification of autism spectrum disorder using multiple functional connectivity-based graph convolutional network

• Published in: Medical & biological engineering & computing Vol. 62; no. 7; pp. 2133 - 2144
• Main Authors: Ma, Chaoran, Li, Wenjie, Ke, Sheng, Lv, Jidong, Zhou, Tiantong, Zou, Ling
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.07.2024; Springer Nature B.V
• Subjects: Artificial neural networks; Autism; Biomedical and Life Sciences; Biomedical Engineering and Bioengineering; Biomedicine; Brain; Classification; Computer Applications; data collection; Data exchange; Datasets; early diagnosis; Functional magnetic resonance imaging; Graph theory; Heterogeneity; Human Physiology; Imaging; Magnetic resonance imaging; magnetism; Medical imaging; Neural networks; Neuroimaging; Original Article; Radiology; Functional connectivity; ASD; Graph convolutional network; Brain network; Readout
• ISSN: 0140-0118; 1741-0444; 1741-0444
• DOI: 10.1007/s11517-024-03060-9

Presently, the combination of graph convolutional networks (GCN) with resting-state functional magnetic resonance imaging (rs-fMRI) data is a promising approach for early diagnosis of autism spectrum disorder (ASD). However, the prevalent approach involves exclusively full-brain functional connectivity data for disease classification using GCN, while overlooking the prior information related to the functional connectivity of brain subnetworks associated with ASD. Therefore, in this study, the multiple functional connectivity–based graph convolutional network (MFC-GCN) framework is proposed, using not only full brain functional connectivity data but also the established functional connectivity data from networks of key brain subnetworks associated with ASD, and the GCN is adopted to acquire complementary feature information for the final classification task. Given the heterogeneity within the Autism Brain Imaging Data Exchange (ABIDE) dataset, a novel External Attention Network Readout (EANReadout) is introduced. This design enables the exploration of potential subject associations, effectively addressing the dataset’s heterogeneity. Experiments were conducted on the ABIDE dataset using the proposed framework, involving 714 subjects, and the average accuracy of the framework was 70.31%. The experimental results show that the proposed EANReadout outperforms the best traditional readout layer and improves the average accuracy of the framework by 4.32%. Graphical Abstract