The Use of Generative Adversarial Network and Graph Convolution Network for Neuroimaging-Based Diagnostic...

Area of Science:

• Neuroscience
• Machine Learning
• Medical Imaging

Background:

• Functional connectivity (FC) from resting-state fMRI is crucial for brain disease classification.
• Traditional machine learning requires specialized feature selection for FC data.
• Deep learning models like CNNs excel with grid data but often overlook graph structures in brain networks.

Purpose of the Study:

• To evaluate the generalization of Graph Convolutional Networks (GCN) and Generative Adversarial Networks (GAN) for brain disease classification.
• To assess the effectiveness of GAN-based data augmentation in improving diagnostic accuracy across diverse datasets.
• To compare the performance of GCN and GAN against established machine learning models and BrainNetCNN.

Main Methods:

• Utilized GCN to analyze brain network structures and extract informative features.
• Employed GAN for data augmentation to address limited training samples and enhance model generalization.
• Applied and validated models on multiple public datasets (ADHD, ABIDE-II, ADNI) and an in-house PTSD dataset.

Main Results:

• GAN data augmentation significantly improved diagnostic accuracy across ADHD, ABIDE-II, and ADNI datasets.
• GAN achieved accuracy increases in ADHD (67.74% to 73.96%), ABIDE-II (70.36% to 77.40%), and ADNI (52.84% to 88.56%).
• GCN demonstrated strong performance, achieving high accuracy on ADHD and ABIDE-II datasets, and both GAN and GCN reached 97.76% accuracy on the PTSD dataset.

Conclusions:

• GAN and GCN show significant potential for improving brain disease prediction and diagnosis.
• GAN's data augmentation capability is particularly effective in enhancing classification accuracy with limited data.
• Further research and development of these graph-based deep learning methods are warranted for clinical applications.