Modality-Aware Graph Encoding Outperforms Vectorized Encoders for Neuroimaging

A new multimodal generative framework evaluates VAEs, transformers, GANs, and diffusion models for structural and functional MRI. The study finds that encoding functional connectivity via modality-aware graphs into a lower-dimensional latent space provides superior feature reconstruction.