LCGNet: Local Sequential Feature Coupling Global Representation Learning for Functional Connectivity Network Analysis...

Area of Science:

• Neuroimaging
• Machine Learning
• Brain Network Analysis

Background:

• Functional connectivity networks (FCNs) from resting-state fMRI are crucial for understanding brain diseases like Alzheimer's (AD) and ADHD.
• Convolutional neural networks (CNNs) excel at local feature extraction but struggle with global temporal patterns in FCNs.
• Transformers capture global temporal features effectively but often miss local network characteristics.

Purpose of the Study:

• To propose a novel network structure, LCGNet, that integrates CNNs and transformers for enhanced FCN representation learning.
• To leverage both local and global topological information within brain networks for improved disease classification.
• To address the limitations of existing methods in capturing comprehensive FCN features.

Main Methods:

• Construct dynamic FCNs using an overlapped sliding window approach from rs-fMRI data.
• Employ a dual backbone architecture combining CNNs and transformers within sequential components (edge-to-vertex, vertex-to-network, network-to-temporality layers).
• Develop LCGNet to couple local sequential features with global representations for comprehensive analysis.

Main Results:

• LCGNet demonstrated superior performance in classifying brain diseases compared to existing methods on real-world datasets.
• The integrated approach effectively captures both local network properties and global temporal dynamics of FCNs.
• Experimental validation on ADNI and ADHD-200 datasets confirmed the efficacy of LCGNet.

Conclusions:

• LCGNet offers a powerful new approach for brain disease classification by effectively integrating local and global FCN features.
• The proposed architecture advances the application of deep learning in neuroimaging analysis.
• This method holds promise for improving diagnostic accuracy and understanding of neurological disorders.