Deep Variational Autoencoder for Mapping Functional Brain Networks

In the neuroimaging and brain mapping communities, researchers have proposed a variety of computational methods to map functional brain networks (FBNs). Recently, it has been proven that deep learning (DL) can be applied on functional magnetic resonance image (fMRI) data with superb representation power over the traditional machine learning methods. However, due to the lack of labeled data and the high dimension of fMRI volume images, DL suffers from overfitting in both supervised and unsupervised training processes. In this work, we proposed a novel generative model: deep variational autoencoder (DVAE) to tackle the challenge of insufficient data and incomplete supervision. The experimental results showed that the representations learned by DVAE are interpretable and meaningful compared to those learned from well-known sparse dictionary learning (SDL). Besides, the organization of some FBN patterns derived from different layers in DVAE was observed in a hierarchical fashion. Furthermore, we showed that DVAE has better performance on small dataset over autoencoder (AE). By using attention deficit hyperactivity disorder (ADHD)-200 dataset as our test bed, we constructed a DVAE-based modeling and classification pipeline in which all subjects' functional connectivities estimated by FBNs were taken as input features to train a classifier. Finally, the results achieved by our pipeline reached state-of-the-art classification accuracies on three ADHD-200 sites compared with other fMRI-based methods.