Deep Generative Analysis for Task-Based Functional MRI Experiments.

Area of Science:

• Neuroscience
• Brain Imaging Analysis
• Statistical Modeling

Background:

• Functional magnetic resonance imaging (fMRI) is crucial in neuroscience but presents significant data analysis challenges.
• The standard mass univariate approach, while successful, relies on potentially inaccurate assumptions and struggles with individual variability.
• Existing methods face limitations in linearity assumptions, noise modeling, and voxel-wise statistical independence.

Purpose of the Study:

• To develop a more flexible and interpretable approach for analyzing task-based fMRI data.
• To address the limitations of the mass univariate method in capturing individual differences and statistical accuracy.
• To propose a generative model with a decomposable structure as an alternative for fMRI analysis.

Main Methods:

• Implemented a generalized additive model (GAM) for direct modeling of entire brain volumes.
• Utilized a variational autoencoder to generate spatial maps and Gaussian Processes for covariate-specific gain modeling.
• Ensured separability of regressor effects to maintain interpretability.

Main Results:

• The proposed GAM approach produced group-level effect maps comparable or superior to standard fMRI analysis software.
• Generated single-subject effect maps that effectively capture individual differences in brain activity.
• Demonstrated increased model flexibility while preserving the interpretability of experimental effects.

Conclusions:

• Generative models with decomposable structures offer a promising and flexible alternative for task-based fMRI data analysis.
• The GAM approach enhances the ability to model complex brain activity patterns and individual variability.
• This method advances the analysis of fMRI data, providing more nuanced insights into brain function.