Generative AI Enhances Multimodal Neuroimaging Analysis with Graph Encoding.

Generative models are proving valuable for encoding complex neuroimaging data, enabling feature generation and reconstruction. However, optimizing the architectural frameworks, especially regarding encoding strategies and latent space processes, is critical for effectively studying the brain's structural and functional properties. A new multimodal generative framework has been designed to process structural gray matter volume (GMV) and static functional network connectivity (sFNC) features from MRI data. This framework systematically evaluates various encoding strategies, latent multimodal fusion techniques, and generative models like VAEs, transformers, GANs, and diffusion models. The findings indicate that architectures employing modality-aware graph encoding for functional connectivity into a lower-dimensional latent space significantly outperform vectorized encoders or direct data space approaches. The proposed multimodal graph VAE (gMMVAE) demonstrates superior performance across multiple metrics, including generation fidelity, reconstruction quality, efficiency, and latent space discriminability, highlighting its potential for robust multimodal neuroimaging analysis.