Identification of Spatio-Temporal Features in Volumetric Task-Based fMRI Data Using 3D Variational Autoencoder

Characterizing functional brain networks (FBNs) has always been an important issue in the field of neuroimaging and brain mapping. Traditional methods often suffer from limitations such as low signal-to-noise ratio and restricted spatial resolution when processing fMRI data. In contrast, deep learning has been demonstrated to exhibit superior performance over traditional methods and can be effectively applied to fMRI data. However, due to the high dimensionality of fMRI volumetric data and the lack of labeled data, deep learning is prone to overfitting in supervised training processes. Most deep learning-based models for functional brain networks identification neglect the critical spatial information of fMRI volumes. In order to tackle these problems, our study proposed and utilized a 3D variational autoencoder (3DVAE) to identify the spatio-temporal features from task-based fMRI (tfMRI) data. As a generative model, 3DVAE can learn fMRI volumetric data in an unsupervised manner without the requirement of any labels, hence resolving the issue of data scarcity. Using the Human Connectome Project (HCP) Q1 data as test-bed, 3DVAE can effectively learn more meaningful and interpretable FBNs than state-of-art method. Furthermore, as a complete data-driven model, 3DVAE can capture meaningful temporal features embedded in tfMRI signals at both subject-level and group-level. We also evaluated the performance of 3DVAE on small datasets and found that it can still derive meaningful spatio-temporal features of tfMRI data even with limited samples.