Student AIM Seminar Seminar

A variational autoencoder (VAE) based on a convolutional neural network (CNN) architecture has been applied to resting state functional magnetic resonance imaging (rsfMRI) data to learn compressed latent representations. The VAE method can reconstruct and generate rsfMRI images using the latent representations as the input, and the latent representations can be used to discover brain networks and characterize individual variation. However, a CNN architecture fails to account for long-range connectivity and interactions among different brain regions. To overcome this limitation, we have further refined the VAE model by adding a spatial transformer module to the CNN. The transformer module uses a self-attention mechanism to learn dynamic and long-range interactions between brain regions. After training the model with data from the Human Connectome Project, we present tests comparing the performance of compressing and reconstructing the patterns of rsfMRI activity for various levels of compression. We also compare this method to principal component analysis (PCA). Analysis of the self-attention matrices from the transformer may provide insight into the dynamic functional connectivity of different brain regions of interest (ROI). This model is also expected to open new ways for using rsfMRI data to classify and characterize patients with different conditions of neurological disorders. Speaker(s): Amaya Murguia (University of Michigan)