TOWARDS ZERO-SHOT TASK-GENERALIZABLE LEARNING ON FMRI.

Area of Science:

• Neuroimaging
• Machine Learning
• Computational Neuroscience

Background:

• Functional MRI (fMRI) using Blood-Oxygen-Level-Dependent (BOLD) signal is crucial for understanding brain function and disorders.
• Task-based fMRI offers richer, task-specific neural activity data compared to resting-state fMRI.
• Aggregating diverse task-based fMRI datasets for generalizable models is challenging due to varied experimental designs.

Purpose of the Study:

• To address the difficulty of aggregating diverse task-based fMRI data.
• To propose a novel supervised network, TA-GAT, for learning generalizable brain patterns from task-based fMRI.
• To enable the integration of task-specific prior knowledge into fMRI analysis.

Main Methods:

• Developed a supervised task-aware network (TA-GAT).
• TA-GAT jointly learns a general-purpose encoder and task-specific contextual information.
• Combines encoder embeddings with contextual information for downstream tasks.

Main Results:

• The proposed TA-GAT architecture facilitates the incorporation of fMRI task prior knowledge.
• The network learns general-purpose embeddings and task-specific contextual information.
• This approach enhances the ability to capture functional brain patterns.

Conclusions:

• TA-GAT offers a flexible, plug-and-play solution for improving task-based fMRI analysis.
• The method enhances the generalizability of models trained on diverse fMRI datasets.
• This work advances the application of machine learning in neuroimaging for studying brain disorders.