Modeling and augmenting of fMRI data using deep recurrent variational auto-encoder.

Area of Science:

• Neuroimaging
• Machine Learning
• Computational Neuroscience

Background:

• Deep learning models show promise in functional magnetic resonance imaging (fMRI) analysis.
• Challenges include overfitting due to limited data and modeling temporal dynamics in fMRI.
• fMRI data augmentation techniques are underexplored.

Purpose of the Study:

• To develop a deep recurrent variational auto-encoder (DRVAE) to address challenges in fMRI deep learning.
• To improve the modeling of temporal dynamics and mitigate overfitting in fMRI data.
• To explore the utility of DRVAE for fMRI data augmentation.

Main Methods:

• A deep recurrent variational auto-encoder (DRVAE) was constructed, combining variational auto-encoder and recurrent neural network components.
• The DRVAE encoder extracts generalized temporal features, while the decoder generates synthetic data for augmentation.
• LASSO regression was applied to estimate spatial networks from temporal features and fMRI data.

Main Results:

• The DRVAE-LASSO framework successfully identified meaningful temporal patterns and spatial networks from real and generated fMRI data.
• Experimental results on the HCP dataset demonstrated the model's ability to learn from group and single-subject data.
• Data augmentation using DRVAE on ADHD-200 resting-state fMRI datasets significantly improved classification performance.

Conclusions:

• The proposed DRVAE method effectively derives temporal features and spatial networks from fMRI data.
• DRVAE serves as a powerful tool for generating high-quality synthetic fMRI data, enhancing data augmentation strategies.
• This approach has broad implications for advancing fMRI analysis and related applications.