SFPGCL: Specificity-preserving federated population graph contrastive learning for multi-site ASD identification using rs-fMRI data

Abstract

Autism spectrum disorder (ASD) is a severe neurodevelopmental disorder that affects people’s social communication and daily routine. Most existing imaging studies on ASD use single site resting-state functional magnetic resonance imaging (rs-fMRI) data, which may suffer from limited samples and geographic bias. Improving the generalization ability of the diagnostic models often necessitates a large-scale dataset from multiple imaging sites. However, centralizing multi-site data generally faces inherent challenges related to privacy, security, and storage burden. Federated learning (FL) can address these issues by enabling collaborative model training without centralizing data. Nevertheless, multi-site rs-fMRI data introduces site variations, causing unfavorable data heterogeneity. This negatively impacts biomarker identification and diagnostic decision. Moreover, previous FL approaches for fMRI analysis often ignore site-specific demographic information, such as age, gender, and full intelligence quotient (FIQ), providing useful information as non-imaging features. On the other hand, Graph Neural Networks (GNNs) are gaining popularity in fMRI representation learning due to their powerful graph representation capabilities. However, existing methods often focus on extracting subject-specific connectivity patterns and overlook inter-subject relationships in brain functional topology. In this study, we propose a specificity-preserving federated population graph contrastive learning (SFPGCL) framework for rs-fMRI analysis and multi-site ASD identification, including a server and multiple clients/sites for federated model aggregation. At each client, our model consists of a shared branch and a personalized branch, where parameters of the shared branch are sent to the sever, while those of the personalized branch remain local. This setup facilitates invariant knowledge sharing among sites and also helps preserve site specificity. In the shared branch, we employ a spatio-temporal attention graph neural network to learn temporal dynamics in fMRI data invariant to each site, and introduce a model-contrastive learning method to mitigate client data heterogeneity. In the personalized branch, we utilize population graph structure to fully integrate demographic information and functional network connectivity to preserve site-specific characteristics. Then, a site-invariant population graph is built to derive site-invariant representations based on the dynamic representations acquired from the shared branch. Finally, representations generated by the two branches are fused for classification. Experimental results on Autism Brain Imaging Data Exchange (ABIDE) show that SFPGCL achieves 80.0 % accuracy and 79.7 % AUC for ASD identification, which outperforms several other state-of-the art approaches.