BrainNet-GAN: Generative Adversarial Graph Convolutional Network for Functional Brain Network Synthesis from Routine...

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Graph Theory

Background:

• Functional brain networks (FBNs) derived from fMRI show clinical potential but fMRI is not routinely acquired.
• Generating FBNs from standard structural MRI is crucial for broader clinical adoption.

Purpose of the Study:

• To develop a model capable of generating FBNs from radiomics-based morphological brain networks (radMBNs) derived from T1-weighted images (T1WI).
• To validate the model's performance and consistency in generating individual-level brain networks.

Main Methods:

• Proposed the BrainNet-GAN model, integrating Multi-Channel Multi-Scale Adaptive (Multi^2Ada) generators and Local_to_Global discriminators.
• Utilized Graph Convolutional Networks (GCN) for multi-scale information aggregation and adaptive fusion within generators.
• Employed Multi-channel GCN and a feature selection module in discriminators, guided by a Multi-Angle Multi-Constraint (MAMC) loss function.

Main Results:

• BrainNet-GAN demonstrated promising performance in generating FBNs across two datasets with 2116 subjects.
• High consistency was observed between generated and target individual-level brain network visualizations.
• Consistent identification of top brain regions by graph-theory metrics between generated and target networks.

Conclusions:

• The BrainNet-GAN model effectively generates FBNs from radMBNs, overcoming limitations of fMRI acquisition.
• This approach facilitates the clinical application of FBN analysis by leveraging routine MRI data.
• The model shows potential for advancing neuroimaging research and clinical diagnostics.