Learning brain representation using recurrent Wasserstein generative adversarial net.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Understanding brain cognition and disorders requires modeling the mind-brain connection.
• Functional brain networks (FBNs) and their temporal features are key to brain representation.
• Deep learning excels at fMRI analysis but struggles with overfitting due to limited data.

Purpose of the Study:

• To develop a deep learning model for robust brain representation learning from fMRI data.
• To address overfitting issues in deep learning models for fMRI analysis.
• To enable fMRI data augmentation through generative modeling.

Main Methods:

• Applied a recurrent Wasserstein generative adversarial net (RWGAN) to volumetric fMRI data.
• Utilized GANs to capture data distribution and extract generalized features, mitigating overfitting.
• Incorporated recurrent layers for temporal feature modeling and LASSO regression for feature decomposition.

Main Results:

• RWGAN successfully learned interpretable temporal features and FBNs from fMRI data.
• The model demonstrated superior performance on small datasets compared to other deep learning methods.
• Generated synthetic fMRI data using RWGAN also yielded meaningful representations.

Conclusions:

• This work represents an early application of generative deep learning for fMRI data modeling.
• The proposed RWGAN provides a novel method for learning brain representations from fMRI.
• RWGAN can generate high-quality synthetic data, offering potential for fMRI data augmentation.