Mapping Neuronal Activity in the Brain Using Hemodynamic Activity to Quantify a Disconnect Between Normal and Abnormal Coupling

Abstract

In this project we tried to develop a machine learning method for learning the neu- ronal activation patterns based on hemodynamic activity in the mouse brain, both represented as functional connectivity optical intrinsic signal (fcOIS) imaging. The goal of this method is to quantify the hemodynamic coupling in the brain and use it to look for a disconnect in brains that are subject to pathology or aged. For this, two deep learning architectures were developed which are both suited to recreating data based on an encoded input. These architectures are the varia- tional autoencoder (VAE) and the variational autoencoding generative adversarial networks (VAE-GAN). These architectures use the hemodynamic data to create a re- construction of the neuronal activity. As the fcOIS data consist of data which have both spatial and temporal elements, a combination of convolutional and recurrent layers is used in the architectures to try and learn the features in the data. Unfortunately, the architectures did not produce satisfactory results. The samples generated by the VAE suffered from mode collapse and the VAE-GAN produced blurry results. The samples generated by the networks are not suited for further investigation into the quantification of the coupling. We do show, however, that the VAE performs slightly better in this task with the current parameters. Improvements might be made by obtaining more data from the mice and by pre-training parts of the architectures used.