Predicting task-related brain activity from resting-state brain dynamics with fMRI Transformer

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Predicting brain task reactivity from resting-state fMRI is challenging.
• Traditional methods struggle with complex spatiotemporal brain patterns.
• Deep learning offers potential for improved prediction accuracy.

Purpose of the Study:

• Introduce SwiFUN (Swin fMRI UNet Transformer), a novel deep learning framework.
• Predict 3D task activation maps directly from resting-state fMRI.
• Enhance understanding and prediction of individual brain function.

Main Methods:

• Utilized a transformer-based deep learning architecture (SwiFUN).
• Employed shifted window-based self-attention for pattern recognition.
• Incorporated contrastive learning to capture individual subject differences.

Main Results:

• SwiFUN achieved higher prediction accuracy than existing methods across all contrasts.
• Demonstrated up to 27% improvement in prediction for the FACES-PLACES contrast (ABCD data).
• Revealed individual differences in task activation maps related to sex, age, and depressive symptoms.

Conclusions:

• SwiFUN provides a scalable, accurate method for predicting task reactivity from resting-state fMRI.
• This approach may reduce reliance on task-based fMRI in clinical settings.
• Offers a promising direction for future neuroscience and clinical research.