FCS-GVAE: A Deep Graph Unsupervised Learning Approach For The Analysis Of The Human Connectome

One of biology's fundamental concepts is that structure affects function; how something is arranged enables it to perform a particular purpose. In this context, network theory has become a standard tool in neuroscience for deciphering functional and anatomical links in the brain and identifying abnormalities in brain diseases. The primary networks of interest are structural (e.g., from DTI) and functional (e.g., from fMRI) brain networks. However, the underlying relationship between structural and functional brain networks has not been fully explored: most current research on brain networks focuses on either structural or functional connectivity, and has not investigated how complementary information between them can be used to improve our understanding. This study investigates the relationship between functional and structural brain networks by training a graph variational autoencoder system that incorporates both functional characteristics and network topology information in an unsupervised fashion. The learned low-dimensional embedding captures key information from both these connectivities and establishes a standard structure-functional spatial coordinate system for comparing different individuals, allowing for individual representation alignment across datasets, which is vital to understanding human brain organization and individual differences. In terms of numbers, this technique enables the transformation of a 132-node representation of the brain networks, i.e., a 132x132 matrix, to a 6-dimensional space. We conduct extensive testing to evaluate the proposed approach using whole-brain network data from the publicly available OASIS-3 longitudinal study dataset. This dataset includes subjects' gender, age, and diagnosis such as whether the subject is affected by Alzheimer's Dementia (AD). We've analyzed the longitudinal data of 865 participants, whose AD state has been defined as having a clinical dementia rating greater than 0. Our experimental findings demonstrate how the proposed unsupervised embedding approach can distinguish different sub-populations, mainly according to their healthy state, better than existing approaches that do not use complementary network information. When the functional connectomes' information is not included in the model, the best f1-score obtained while predicting the healthy state of the subjects has been 0.23. On the other hand, including the variability of the Functional Connectome, it has been possible to increase the outcomes of the classification task for the healthy state of the patients, noticing a boost close to 40\% in the f1-score itself.