Representation learning of resting state fMRI with variational autoencoder

Area of Science:

• Neuroscience
• Machine Learning
• Medical Imaging

Background:

• Resting-state functional magnetic resonance imaging (rsfMRI) data reveals intricate patterns of brain activity.
• The origins of these complex patterns remain unclear and are often entangled within the data.
• Unsupervised learning methods are needed to decipher these underlying sources of rsfMRI activity.

Purpose of the Study:

• To develop a generative model using unsupervised learning to disentangle sources of rsfMRI activity.
• To explore the utility of latent variables for representing and generating cortical activity and connectivity patterns.
• To assess the effectiveness of representational geometry in latent space for subject identification.

Main Methods:

• Implementation of a variational auto-encoder (VAE) as a generative model.
• Training the VAE on large-scale resting-state fMRI data from the Human Connectome Project.
• Analysis of latent representations and their trajectories to capture spatiotemporal characteristics.

Main Results:

• The VAE successfully learned to represent and generate patterns of cortical activity and connectivity.
• Latent variables were found to reflect principal gradients and drive activity changes in cortical networks.
• Representational geometry (covariance/correlation of latent variables) proved more reliable than cortical connectivity for subject identification, even with limited data.

Conclusions:

• Variational auto-encoders are a valuable tool for unsupervised representation learning in rsfMRI data.
• The VAE approach effectively disentangles complex brain activity sources.
• Latent space geometry offers a robust feature for subject identification from rsfMRI data.