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

Area of Science:

• Neuroimaging
• Brain Mapping
• Computational Neuroscience

Background:

• Functional brain networks (FBNs) characterization is crucial in neuroimaging.
• Traditional fMRI analysis methods face challenges with signal-to-noise ratio and spatial resolution.
• Deep learning models for FBNs often struggle with high-dimensional fMRI data and overfitting due to limited labeled datasets, neglecting spatial information.

Purpose of the Study:

• To propose and utilize a 3D variational autoencoder (3DVAE) for identifying spatio-temporal features in task-based fMRI (tfMRI) data.
• To address limitations of traditional methods and supervised deep learning in fMRI analysis.
• To develop an unsupervised deep learning model capable of learning from high-dimensional fMRI data without labels.

Main Methods:

• Utilized a 3D variational autoencoder (3DVAE), a generative deep learning model.
• Applied unsupervised learning to fMRI volumetric data, eliminating the need for labeled data.
• Tested the model on Human Connectome Project (HCP) Q1 data.

Main Results:

• The 3DVAE effectively learned more meaningful and interpretable FBNs compared to state-of-the-art methods.
• The model captured significant spatio-temporal features from tfMRI signals at both subject and group levels.
• 3DVAE demonstrated robust performance in deriving meaningful features even on small datasets, indicating resilience to limited sample sizes.

Conclusions:

• The proposed 3DVAE is a powerful, data-driven tool for unsupervised identification of functional brain networks from tfMRI data.
• This approach overcomes key challenges in fMRI analysis, including data scarcity and the need for spatial information.
• 3DVAE offers a promising avenue for advancing brain mapping and understanding brain function through neuroimaging data.