Graph Autoencoders for Embedding Learning in Brain Networks and Major Depressive Disorder Identification

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Brain functional connectivity (FC) networks derived from fMRI show alterations in neuropsychiatric disorders.
• Traditional deep learning methods often overlook the topological information of brain networks, potentially limiting diagnostic performance.
• Major Depressive Disorder (MDD) diagnosis can benefit from advanced neuroimaging analysis techniques.

Purpose of the Study:

• To propose a novel graph deep learning framework for classifying brain networks in MDD.
• To leverage non-Euclidean information and graph structure for improved brain disorder identification.
• To develop a method that embeds topological and content features of fMRI networks into low-dimensional representations.

Main Methods:

• Utilized a graph autoencoder (GAE) architecture based on graph convolutional networks (GCNs).
• Employed the Ledoit-Wolf (LDW) shrinkage method for efficient estimation of high-dimensional FC metrics from fMRI data.
• Integrated graph embeddings as input features for a deep fully-connected neural network (FCNN) for MDD vs. healthy control (HC) classification.

Main Results:

• The proposed GAE-FCNN framework achieved superior performance compared to state-of-the-art methods in brain connectome classification.
• Highest accuracy was obtained when using LDW-FC edges as node features within the GAE.
• Learned graph embeddings revealed significant differences in brain network topology between MDD patients and HCs.

Conclusions:

• The graph deep learning framework effectively captures discriminative information from brain network topology for diagnosing MDD.
• This approach demonstrates the potential of graph embeddings in neuroimaging for identifying brain disorders.
• The findings highlight the importance of considering network topology in connectome-based machine learning analyses.