Self-Supervised Transformer-Based Foundation Model for functional Magnetic resonance Imaging.

Area of Science:

• Neuroimaging
• Machine Learning
• Computational Neuroscience

Background:

• Functional Magnetic Resonance Imaging (fMRI) offers insights into brain function but faces challenges with high dimensionality and data variability.
• Developing robust methods to analyze complex fMRI time series is crucial for advancing neuroscience and clinical applications.

Purpose of the Study:

• To introduce a self-supervised transformer-based foundation model for learning generalizable representations of fMRI time series.
• To evaluate the model's performance in cognitive task classification and neuroticism prediction using various probe architectures and learning settings.

Main Methods:

• Utilized a masked autoencoder approach within a transformer architecture for self-supervised learning on fMRI data.
• Trained the model on the Human Connectome Project (HCP) S1200 dataset.
• Evaluated performance using linear, MLP, and ConvLSTM probes in zero-shot and fine-tuning scenarios.

Main Results:

• The proposed model significantly outperformed training from scratch.
• Achieved over 90% accuracy in cognitive task classification.
• Demonstrated improved correlation in neuroticism prediction compared to baseline methods.
• Architectural enhancements like contrastive loss and spatiotemporal attention further improved representation quality.

Conclusions:

• Self-supervised transformers show significant potential for analyzing fMRI data.
• The developed model enables scalable and generalizable representations for neuroscience research.
• This approach paves the way for advanced clinical applications in brain disorder diagnosis and monitoring.