REPRESENTATION LEARNING OF FMRI DATA USING VARIATIONAL AUTOENCODER

Functional imaging data of the brain using Magnetic Resonance Imaging (MRI) – fMRI data exhibits complex but structured patterns. This fMRI data has opened a new venue for understanding the brain system at the whole-brain scale. However, the underlying origins of fMRI data are unclear and entangled. In this dissertation, I establish a variational auto-encoder, a generative model trainable with an unsupervised learning algorithm, to disentangle the unknown sources of fMRI activity. After being trained with large fMRI data in cooperation with a new reformatting strategy of input fMRI data, the model has learned the representations of cortical activity using latent variables. In Chapter 3, I found that the latent representation and its trajectory represented the spatiotemporal characteristics of fMRI activity under resting state. The latent variables reflected the principal gradients of the latent trajectory and drove activity changes in cortical networks. Latent representations were clustered by both individuals and brain states. Representational geometry captured as the covariance between latent variables, rather than cortical connectivity, was used as a more reliable feature to accurately identify subjects from a large group, even if only a short period of data was available per subjects. In Chapter 4, I further applied the VAE model pretrained with fMRI data in the resting state to new fMRI data from subjects watching naturalistic movies. I further validated that my VAE model was highly generalizable to fMRI data under different brain conditions and different scanning parameters. Additionally, I showed the task-evoked brain activity and spontaneous brain activity could be linearly separable in the VAE-derived latent space. Task-evoked latent representations and trajectory were employed to understand the dynamics of brain networks during naturalistic movie stimuli. I found that the principal gradients of the task-evoked latent trajectory were related to many aspects of the movie stimuli: low-, middle-, high-level video features. Cortical mapping of principal gradients showed the interactions between distributed cortical networks spanning from low-level sensory to high-level cognitive. Taken together, the VAE model proposed in this dissertation is a novel and effective tool that can potentially be used for understanding cortical dynamics in different brain conditions and disease conditions.