SFPGCL: Specificity-preserving federated population graph contrastive learning for multi-site ASD identification...

Area of Science:

• Neuroscience
• Medical Imaging
• Machine Learning

Background:

• Autism Spectrum Disorder (ASD) diagnosis is challenged by limited sample sizes and geographic bias in single-site neuroimaging studies.
• Centralizing multi-site resting-state functional magnetic resonance imaging (rs-fMRI) data for Autism Spectrum Disorder (ASD) research faces privacy, security, and storage hurdles.
• Existing federated learning (FL) and Graph Neural Network (GNN) methods for rs-fMRI analysis often overlook multi-site data heterogeneity and inter-subject relationships.

Purpose of the Study:

• To develop a novel framework, Specificity-Preserving Federated Population Graph Contrastive Learning (SFPGCL), for multi-site rs-fMRI analysis and Autism Spectrum Disorder (ASD) identification.
• To address data heterogeneity and preserve site-specific characteristics in federated learning settings for neuroimaging data.
• To integrate non-imaging demographic information with functional connectivity for improved diagnostic performance in ASD.

Main Methods:

• Proposed a federated learning framework (SFPGCL) with a server and multiple clients, featuring a shared branch for invariant knowledge sharing and a personalized branch for site-specific information.
• Employed a spatio-temporal attention Graph Neural Network (GNN) in the shared branch for learning site-invariant temporal dynamics and contrastive learning to mitigate data heterogeneity.
• Utilized population graph structure in the personalized branch to integrate demographic data and functional network connectivity, creating site-invariant representations for classification.

Main Results:

• The SFPGCL framework achieved 80.0% accuracy and 79.7% Area Under the Curve (AUC) for Autism Spectrum Disorder (ASD) identification on the Autism Brain Imaging Data Exchange (ABIDE) dataset.
• Demonstrated superior performance compared to several state-of-the-art approaches in multi-site rs-fMRI analysis for ASD diagnosis.
• Effectively mitigated data heterogeneity and preserved crucial site-specific characteristics during federated model training.

Conclusions:

• The SFPGCL framework offers a robust and effective solution for multi-site Autism Spectrum Disorder (ASD) identification using rs-fMRI data.
• Federated learning combined with Graph Neural Networks and contrastive learning can successfully address challenges of data heterogeneity and privacy in large-scale neuroimaging studies.
• The proposed method highlights the importance of integrating both shared and site-specific information for accurate and generalizable diagnostic models in neurodevelopmental disorders.