基于Transformer和图卷积神经网络的自闭症谱系障碍分类  

Classification of Autism Spectrum Disorder Based on Transformer and Graph Convolutional Neural Networks

作　　者：阿日茜 彭博 戴亚康[2] 庞春颖[1] Arixi;PENG Bo;DAI Yakang;PANG Chunying(School of Life Science and Technology,Changchun University of Science and Technology,Changchun 130022;Suzhou Institute of Biomedical Engineering Technology,Chinese Academy of Sciences,Suzhou 215163)

机构地区：[1]长春理工大学生命科学技术学院,长春130022 [2]中国科学院苏州生物医学工程技术研究所,苏州215163

出　　处：《长春理工大学学报（自然科学版）》2024年第6期107-115,共9页Journal of Changchun University of Science and Technology(Natural Science Edition)

基　　金：国家自然科学基金资助项目(62301557);江苏省重点研发计划项目(BE2022842,BE2022049)。

摘　　要：自闭症谱系障碍(ASD)是一种神经发育障碍,其特征包括社交、沟通和行为方面的严重缺陷。早期准确的诊断和干预对改善患者预后至关重要,然而由于其症状的多样性和复杂性,早期诊断常具有挑战性。近年来,结合功能性磁共振成像(fMRI)等神经影像技术和深度学习算法,为ASD的诊断提供了更准确的手段。脑功能连接网络数据中存在着复杂的局部和全局连接模式,传统的图卷积网络(GCN)更擅长捕捉局部连接模式,而Transformer模型则更适合捕捉全局依赖关系。因此,将这两种模型有效地融合起来以综合利用局部和全局信息对于提高ASD分类的性能至关重要。通过功能性磁共振成像等技术建立脑网络,结合GCN和Transformer模型,并利用数据增强技术,提高了对ASD分类的建模效果。使用联合模型和数据增强技术,在测试集上取得了87.10的AUC,显著高于单独使用GCN和Transformer模型的性能。这一研究不仅为ASD的诊断提供了一种高效手段,也有望为深入理解其潜在的神经机制提供重要启示。Autism Spectrum Disorder(ASD)is a neurodevelopmental disorder characterized by severe deficits in social interaction,communication,and behavior.Early accurate diagnosis and intervention are crucial for improving patient outcomes,yet early diagnosis is often challenging due to the diversity and complexity of symptoms.In recent years,the combination of neuroimaging techniques such as functional magnetic resonance imaging(fMRI)and deep learning algorithms has provided more accurate means for diagnosing ASD.Functional connectivity network data exhibits complex patterns of both local and global connections,with traditional Graph Convolutional Networks(GCNs)being better at capturing local connectivity patterns,while Transformer models are more suitable for capturing global dependencies.Therefore,effectively integrating these two models to leverage both local and global information is essential for improving the performance of ASD classification.By establishing brain networks using techniques like functional MRI,combined with GCN and Transformer models,and utilizing data augmentation techniques,we have enhanced the modeling effectiveness for ASD classification.Using the joint model and data augmentation techniques,we achieved an AUC of 87.10 on the test set,significantly higher than the performance of using GCN and Transformer models alone.This study not only provides an efficient approach for the diagnosis of ASD but also holds promise for offering important insights into its underlying neural mechanisms.

关 键 词：自闭症谱系障碍 功能性磁共振成像 脑功能连接网络 深度学习 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]