MAGE: Automatic diagnosis of autism spectrum disorders using multi-atlas graph convolutional networks and ensemble learning

Currently, it is still a great challenge in clinical practice to accurately diagnose autism spectrum disorder (ASD). To address this challenge, in this study we propose a method for automatic diagnosis of ASD based on multi-atlas graph convolutional networks and ensemble learning. Firstly, we extract multiple feature representations based on functional connectivity (FC) of different brain atlases from fMRI data of each subject. Then, to obtain the features that are more helpful for ASD automatic diagnosis, we propose a multi-atlas graph convolutional network method (MAGCN). Finally, to combine different feature representations, we propose a stacking ensemble learning method to perform the final ASD automatic diagnostic task. Our proposed method is evaluated on 949 subjects (including 419 subjects with ASD and 530 subjects with typical control (TC)) from the Autism Brain Imaging Data Exchange (ABIDE). Experimental results show that our proposed method achieves an accuracy of 75.86% and an area under the receiver operating characteristic curve (AUC) of 0.8314 for automatic diagnosis of ASD. In addition, compared with some methods published in recent years, our proposed method obtains the best performance of ASD diagnosis. Overall, our proposed method is effective and promising for automatic diagnosis of ASD in clinical practice. (c) 2021 Elsevier B.V. All rights reserved.