A Multi-Site Anti-Interference Neural Network for ASD Classification

Abstract

Autism spectrum disorder (ASD) is a complex neurodevelopmental disorder that can reduce quality of life and burden families. However, there is a lack of objectivity in clinical diagnosis, so it is very important to develop a method for early and accurate diagnosis. Multi-site data increases sample size and statistical power, which is convenient for training deep learning models. However, heterogeneity between sites will affect ASD recognition. To solve this problem, we propose a multi-site anti-interference neural network for ASD classification. The resting state brain functional image data provided by the multi-site is used to train the ASD classification model. The model consists of three modules. First, the site feature extraction module is used to quantify the inter-site heterogeneity, in which the autoencoder is used to reduce the feature dimension. Secondly, the presentation learning module is used to extract classification features. Finally, the anti-interference classification module uses the output of the first two modules as labels and inputs for multi-task adversarial training to complete the representation learning that is not affected by the confounding of sites, so as to realize the adaptive anti-interference ASD classification. The results show that the average accuracy of ten-fold cross validation is 75.56%, which is better than the existing studies. The innovation of our proposed method lies in the problem that the traditional single-task deep learning ASD classification model will be affected by the heterogeneity of multi-site data and interfere with the classification. Our method eliminates the influence of multi-site factors on feature extraction through multi-task adversarial training, so that the model can better adapt to the heterogeneity of multi-site data. Meanwhile, large-scale 1DconV is introduced to extract features of brain functional network, which provides support for the interpretability of the model. This method is expected to take advantage of multiple sites and provide reference for early diagnosis and treatment of ASD.

Application of Variational Graph Autoencoder in Traction Control of Energy-Saving Driving for High-Speed Train

“…The VGAE developed from the variational auto-encoder (VAE) is a graph neural network that uses latent variable learning to approximate the Gaussian distribution and can effectively solve the link prediction problem [11][12][13]. Nowadays, an algorithm based on graph neural networks is mainly applied in the field of transportation to solve the problem of traffic flow prediction in urban traffic networks and is rarely applied to issues related to high-speed trains [14][15][16].…”