Exploring the potential of MCVAE for patients stratification and skewed data compensation across the AD continuum

Imaging Genetics (IG) focuses on integrative studies to assess the influence of genetic architecture on brain structure and function, which is particularly relevant to decrypt the underlying mechanisms of complex pathologies such as Alzheimer's Disease (AD). To this aim, in this work, we exploited a Multi-Channel Variational Autoencoder (MCVAE) for investigating the genetic underpinnings of Grey Matter (GM) and White Matter (WM) modulations in the AD continuum with the twofold goal of (i) capturing the associations among multimodal features in the latent space; (ii) assessing the generative potential of the model to be eventually exploited for data augmentation and mitigation of the data skewed problem. Three channels were considered, respectively representing morphometric, microstructural and genetic features. The resulting common latent space was well aligned across the different channels, highlighting a clustering of the latent components concordant with different stages of AD. Regarding the generative performance, reconstruction accuracy revealed that diffusion MRI led to the best performance, especially when decoded from structural MRI, while neither the microstructural nor the genetic channel allowed effective decoding of the others. This might most probably be due to the limited data numerosity and the inherent limitations of the MCVAE architecture. Our preliminary results provide evidence of the potential of multi-modal feature integration with MCAVE models for decoding subtle disease signatures and highlight some limitations of the model providing hints for further improvement.