A Foundational fMRI Model for Representing Continuous Brain States.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Existing models struggle with the temporal complexity of brain signals due to fixed time windows.
• Understanding brain state dynamics is crucial for advancing neuroscience and clinical applications.

Purpose of the Study:

• To develop a novel functional magnetic resonance imaging (fMRI) foundational model, BrainSN (Brain States Network), capable of representing continuous brain state information.
• To enable diverse downstream tasks, including clinical diagnosis and mental state decoding.

Main Methods:

• Utilized a transformer-based architecture for BrainSN to reconstruct and predict brain states across multiple time scales.
• Integrated multiple embeddings and a channel gating module with an attention mechanism for feature extraction.
• Trained BrainSN on 1,256 hours of resting-state and naturalistic stimulus fMRI data.

Main Results:

• Achieved high accuracy (75.23% autism, 75.82% attention disorder) in diagnostic tasks without fine-tuning, matching leading models.
• Attained 95.31% accuracy in mental state decoding without fine-tuning, outperforming models trained on task-based fMRI data.
• Demonstrated BrainSN's ability to capture semantic content and sequence sensitivity from fMRI signals during movie stimuli analysis.

Conclusions:

• BrainSN effectively models brain state dynamics, capturing both short-term and long-term dependencies.
• The model shows significant potential for clinical diagnosis, treatment evaluation, and cognitive neuroscience research.
• BrainSN offers advantages over existing models by learning large-scale brain dynamics without task-based paradigms.