Graph-Regularized Manifold-Aware Conditional Wasserstein GAN for Brain Functional Connectivity Generation.

Area of Science:

• Neuroimaging
• Machine Learning
• Computational Neuroscience

Background:

• Brain functional connectivity (FC) data, often represented as covariance or correlation matrices, possess a unique Symmetry-Positive Definite (SPD) structure residing on a Riemannian manifold.
• Standard Generative Adversarial Networks (GANs) often fail to capture this inherent SPD structure, leading to unrealistic FC data generation and neglecting the inter-relatedness of network edges.

Purpose of the Study:

• To develop a novel GAN model capable of generating realistic manifold-valued FC data while preserving the intrinsic SPD structure.
• To improve the generation of class-conditional FC data for applications like distinguishing between healthy controls and patients with brain disorders.

Main Methods:

• Proposed a graph-regularized manifold-aware conditional Wasserstein GAN (GR-SPD-GAN) that optimizes a generalized Wasserstein distance on the SPD manifold.
• Incorporated population graph-based regularization to maintain inter-subject similarity and stabilize training, avoiding mode collapse.
• Conditioned the GAN on class labels to generate specific types of brain network data.

Main Results:

• The GR-SPD-GAN successfully generated more realistic fMRI-based FC samples compared to state-of-the-art GANs when evaluated on major depressive disorder (MDD) data.
• FC data augmentation using GR-SPD-GAN significantly improved classification accuracy for MDD identification, outperforming other GAN-based augmentation methods.

Conclusions:

• The proposed GR-SPD-GAN is effective for generating high-fidelity SPD-valued FC data, preserving global network structure and inter-subject relationships.
• This approach offers a powerful tool for FC data augmentation, enhancing the performance of diagnostic classification models in neuroscience research.