BrainNPT: Pre-Training Transformer Networks for Brain Network Classification.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning in brain imaging is limited by scarce labeled data.
• Pre-training on unlabeled data shows promise in various AI fields.
• Its application in brain network analysis remains underexplored.

Purpose of the Study:

• To introduce BrainNPT, a Transformer-based network for brain functional network classification.
• To develop a pre-training framework leveraging unlabeled data for improved feature learning in brain networks.
• To enhance classification performance by utilizing structural information from unlabeled brain data.

Main Methods:

• Proposed BrainNPT, a Transformer network utilizing a token for classification embedding.
• Developed a pre-training framework for BrainNPT using unlabeled brain network data.
• Conducted classification experiments comparing pre-trained and non-pre-trained models against state-of-the-art methods.

Main Results:

• BrainNPT without pre-training achieved state-of-the-art performance.
• The pre-trained BrainNPT model significantly outperformed existing methods, showing an 8.75% accuracy improvement.
• Analysis included comparisons of pre-training strategies, data augmentation, and model parameter influence.

Conclusions:

• Transformer-based pre-training is highly effective for brain functional network classification.
• Leveraging unlabeled data with BrainNPT substantially boosts classification accuracy.
• The study validates the efficacy of the proposed pre-training framework and model architecture.