磁共振成像2025，Vol.16Issue(12)：7-13,7.DOI:10.12015/issn.1674-8034.2025.12.002
基于交叉注意力结合静态-动态图卷积技术在帕金森病脑静息态功能连接分析中的应用
Cross-attention fusion of static-dynamic graph convolutional networks for Parkinson's disease diagnosis
摘要
Abstract
Objective:Based on resting-state functional magnetic resonance imaging(rs-fMRI)data,the cross attention mechanism(CAM)combined with static-dynamic graph convolutional network(GCN)technology was utilized to evaluate the classification efficacy of this method for patients with Parkinson's disease(PD),and to explore potential imaging biomarkers,providing a new perspective for the clinical diagnosis and pathological mechanism analysis of PD.Materials and Methods:A total of 32 patients with PD were prospectively recruited from the outpatient department of the Affiliated Hospital of Qingdao University,and 30 healthy controls(HC),matched for gender,age and education years were recruited from the community health management center of the Affiliated Hospital of Qingdao University.Resting-state functional magnetic resonance imaging was collected from both groups of subjects.After image preprocessing,static-graph convolutional networks(static-GCN)and dynamic-graph convolutional networks(dynamic-GCN)were constructed for each subject based on the AAL atlas and GCN.Through multi-scale feature extraction and CAM,the complementary information of static-GCN and dynamic-GCN was fused.The performance was evaluated using the accuracy of five-fold cross-validation and the area under the receiver operating characteristic(ROC)curve(AUC).The attention weight coefficients obtained during the process were combined with statistical analysis to identify the abnormal brain regions and static-dynamic functional connections(static-dynamic FC)most related to PD.Two independent sample t-tests were used for inter-group comparisons,and Pearson correlation analysis was used to explore the correlation between the statistically significant static-dynamic FC and clinical scales.Results:The method based on CAM combined with static-dynamic graph convolution network has excellent classification performance(with an accuracy of 79.84%,a sensitivity of 80.47%,and a specificity of 78.47%).The ROC curve analysis results show that the AUC for diagnosing PD is 0.814(95%CI:0.727 to 0.902,P<0.001).Five PD high-weight brain regions were identified:the right supplementary motor area,the left posterior cingulate gyrus,the left postcentral gyrus,cerebellar Lobule Ⅵ,and vermis 10.At the same time,two most relevant static-dynamic FC were discovered.Compared with the HC group,the static-dynamic FC in the following two pairs of brain regions was significantly enhanced in the PD group(P<0.05):(1)the left posterior cingulate gyrus-cerebellar Lobule Ⅵ;(2)the right supplementary motor area-vermis 10-cerebellar Lobule Ⅵ/the left postcentral gyrus.Moreover,both of these enhanced static-dynamic FC were significantly positively correlated with the UPDRS-Ⅲ score(r=0.432,P=0.017;r=0.420,P=0.021).Conclusions:The method combining CAM with static-dynamic graph convolution networks has excellent diagnostic performance,and has discovered abnormal enhanced patterns of specific static-dynamic FC between the cerebellum and the cerebral cortex in patients with PD,providing a new basis for the objective imaging diagnosis of PD.关键词
帕金森病/磁共振成像/静态功能连接/动态功能连接/图卷积网络/交叉注意力机制Key words
Parkinson's disease/magnetic resonance imaging/static functional connectivity/dynamic functional connectivity/graph convolutional network/cross-attention mechanism分类
医药卫生引用本文复制引用
TANG Yueshan,ZHANG Xiaofei,LIU Xuejun,YU Mengmeng,CHEN Xue,REN Yande..基于交叉注意力结合静态-动态图卷积技术在帕金森病脑静息态功能连接分析中的应用[J].磁共振成像,2025,16(12):7-13,7.基金项目
Natural Science Foundation of China(No.82301662) （No.82301662）
Shandong Provincial Natural Science Foundation(No.ZR2024QH005).国家自然科学基金项目(编号:82301662) （No.ZR2024QH005）
山东省自然科学基金项目(编号:ZR2024QH005) （编号:ZR2024QH005）
