Representation learning of resting state functional MRI using a volumetric variational autoencoder model (3D VAE)

In this work, we consider whether a VAE model trained with volumetric fMRI data (rather than a cortical subset of the data) is capable of encoding fMRI into low-dimensional representations, decoding these representations back into volumetric fMRI space, and also generating new fMRI patterns from the latent space. For 3D VAE model training, validation, and testing, volumetric resting-state fMRI data was used from the Human Connectome Project minimally preprocessed pipeline. We find the 3D VAE is able to accurately represent the spatial and temporal information in the data. In addition, it is able to synthesize realistic resting-state networks.