目的 针对孤独症多模态证据融合与定量化辨识的关键问题,本研究提出基于图卷积神经网络(GCN)的孤独症谱系障碍(ASD)诊断模型研究思路。方法 通过对来源于ABIDE的ASD儿童脑部fMRI数据进行整理和筛选,提取脑区功能连接矩阵作为图结构的邻接矩阵,并融合临床表型数据,构建了ASD多模态关联网络。通过网络特征比较分析,识别出了ASD与典型发育组的脑功能连接网络组间差异。进一步地构建一个端到端的GCN模型,并尝试引入注意力机制,提高模型决策的可解释性。结果 该模型在诊断性能指标优于传统机器学习方法(准确率=0.710,精确率=0.709,召回率=0.780,F1=0.743,曲线下面积=0.746)。背侧注意网络与边缘系统-颞极枢纽的功能连接减弱是模型做出判断的最主要依据。结论 以异质图为多模态数据整合的基本架构,本模型为ASD的潜在病理机制探索提供了新的方法学范例。
Objective To develop a quantitative model for autism spectrum disorder(ASD)integration multimodal evidences.Methods The fMRI dataset from ABIDE was used for extracting connectivity function network of ASD after data preprocessing.Difference between ASD and typical development about their brain connectivity function was evaluated with t-test.Integrating phenotypic data and fMRI dataset,an graph convolutional neural network (GCN)with attention module was estimated and compared against benchmark models about their efficiency and interpretability.Results The GCN model was evaluated outperformed other models with better accuracy indices.And regions from Dorsal Attention Network and Limbic-Temporal Pole were ranked as the highest weights for the differentiation in the model.Conclusions This study provided a novel paradigm for quantitative diagnosis and exploring pathogenesis of ASD.