基于多模态图神经网络的自闭症识别方法
A Multi-modal Graph Random Attention Network for Autism Diagnosis
李文源 1王学研 1俞凌雲 1孙淑慧 1黄惠芳1
作者信息
- 1. 北京交通大学计算机与信息技术学院，北京，100044
- 折叠
摘要
自闭症（Autism Spectrum Disorder，ASD）是一种常见的精神发育障碍性疾病，其患病率正在逐年增加，严重影响患者的正常生活。磁共振成像技术（Magnetic resonance imaging，MRI）能为ASD识别提供客观的生物标志物，在一定程度上弥补了临床诊断方法主观性较强的不足，其中结构磁共振成像（Structure MRI，sMRI）和功能磁共振成像（Resting-state functional MRI，rs-fMRI）分别提供大脑结构和功能信息，已被广泛用于ASD识别研究。然而，现有完全监督的多模态融合方法往往利用双模态，忽视了更多模态间的复杂关联关系，不能模拟人群中被试之间的关联，也不能较好地利用人口统计学信息。此外，这些方法通常忽略了不同模态数据中存在的噪声，导致诊断性能不佳。为了解决这些限制，我们提出了一种新颖的用于ASD识别的多模态图随机注意力（Multi-modal Graph Random Attention，MGRA）网络。MGRA从研究被试间关联的角度出发，充分利用了人口统计学信息和图神经网络强大的建模能力。具体来说，我们首先构建了基于信息瓶颈的成对关联注意力模块（Pairwise Association Graph Attention-Information Bottleneck, PAGA-IB），根据非成像信息自适应地学习被试间的关联，滤除图结构中的噪声和冗余信息。然后设计了用于模态内和模态间融合的图学习模块，利用PAGA-IB和图卷积获得模态的深层特征表示，实现模态内和模态间特征的充分交互，提高多模态特征的表征能力。此外，在模态间图学习模块之前构建融合门以接收来自不同模态的信息，实现对sMRI与rs-fMRI 特征的初步融合，以便图学习模块实现更好地多模态特征融合。最后，我们在ABIDE-I数据集上评估了所提出的方法。实验结果表明，该方法相较于其它方法在ASD识别性能上有较大的提升。
Abstract
Autism spectrum disorder (ASD) is a common neurodevelopmental disorder whose prevalence has been increasing year by year, severely affecting patients\' daily lives. Magnetic resonance imaging (MRI) can provide objective biological biomarkers for ASD identification, thereby alleviating the strong subjectivity of clinical diagnostic methods to some extent. Among them, structural MRI (sMRI) and resting-state functional MRI (rs-fMRI) provide complementary structural and functional information of the brain and have been widely used in ASD identification studies. However, existing fully supervised multimodal fusion methods usually exploit only dual modalities, ignoring the complex relationships among multiple modalities. They also fail to model the relationships among subjects in the population and cannot effectively leverage demographic information. Moreover, these methods often overlook the noise present in different modalities, resulting in suboptimal diagnostic performance.To address these limitations, we propose a novel Multimodal Graph Random Attention (MGRA) network for ASD identification. From the perspective of modeling inter-subject relationships, MGRA fully exploits demographic information and the powerful modeling capability of graph neural networks. Specifically, we first construct a pairwise association-guided attention module based on the information bottleneck principle (PAGA-IB), which adaptively learns inter-subject associations from non-imaging information and filters out noise and redundant information in the graph structure. We then design graph learning modules for intra-modal and inter-modal fusion, which leverage PAGA-IB and graph convolution to obtain deep feature representations of each modality, enabling sufficient interaction within and across modalities and enhancing the representational capacity of multimodal features. In addition, a fusion gate is introduced before the inter-modal graph learning module to receive information from different modalities and achieve preliminary fusion of sMRI and rs-fMRI features, thereby facilitating more effective multimodal feature integration in the subsequent graph learning module. Finally, we evaluate the proposed method on the ABIDE-I dataset. Experimental results demonstrate that the proposed approach achieves significant performance improvements over existing methods in ASD identification.关键词
自闭症/ 磁共振成像/ 多模态融合/ 图神经网络。Key words
Autism spectrum disorder/magnetic resonance imaging/multimodal fusion/graph neural networks.引用本文复制引用
李文源,王学研,俞凌雲,孙淑慧,黄惠芳.基于多模态图神经网络的自闭症识别方法[EB/OL].(2026-04-22)[2026-04-24].http://www.paper.edu.cn/releasepaper/content/202604-161.学科分类
神经病学、精神病学/医学研究方法
评论