BrainSTEAM: A Practical Pipeline for Connectome-based fMRI Analysis towards Subject Classification

Area of Science:

• Neuroscience
• Machine Learning
• Medical Imaging

Background:

• Functional brain networks reveal complex interactions crucial for understanding neural patterns and diagnosing disorders.
• Graph neural networks (GNNs) excel at analyzing structured network data but are prone to overfitting with limited neuroimaging datasets.
• Overfitting in GNNs hinders their ability to capture relevant neural patterns for predictive tasks.

Purpose of the Study:

• To introduce BrainSTEAM, an integrated framework designed to improve the analysis of functional brain networks using GNNs.
• To address the challenges of overfitting and limited data in neuroimaging analysis.
• To enhance the predictive power of GNNs for clinical applications.

Main Methods:

• BrainSTEAM utilizes a spatio-temporal module incorporating EdgeConv GNN, an autoencoder, and Mixup strategy.
• The framework dynamically segments ROI time-series signals into sequences to augment training data.
• EdgeConv GNN captures ROI connectivity, an autoencoder denoises data, and Mixup enhances training via data augmentation.

Main Results:

• BrainSTEAM was evaluated on the ABIDE (Autism) and HCP (Gender) datasets.
• The framework demonstrated superior performance and robustness compared to existing models.
• Results indicate BrainSTEAM's effectiveness in generalizing to diverse connectome-based fMRI studies.

Conclusions:

• BrainSTEAM offers a robust solution for analyzing functional brain networks, overcoming limitations of traditional GNNs in neuroimaging.
• The proposed framework shows significant potential for improving diagnostic accuracy in neurological and psychiatric disorders.
• BrainSTEAM's mechanisms are adaptable for broader applications in connectome-based fMRI analysis.