Contrastive learning for neural fingerprinting from limited neuroimaging data

Area of Science:

• Neuroscience
• Machine Learning
• Biometrics

Background:

• Neural fingerprinting identifies individuals via unique brain activity patterns.
• Deep learning excels but requires retraining for new subjects and struggles with limited neuroscience data.
• Existing methods face scalability and data limitations.

Purpose of the Study:

• To develop a scalable neural fingerprinting method using contrastive learning and data augmentation.
• To eliminate the need for retraining deep learning models with new subjects.
• To enhance model robustness in low-sample-size conditions.

Main Methods:

• Utilized the LEMON dataset (138 subjects, 3T MRI, resting-state fMRI).
• Computed functional connectivity for baseline correlation metrics.
• Adapted a deep learning model with data augmentation and contrastive triplet loss.

Main Results:

• Deep learning significantly improved fingerprinting performance over correlation-based methods (98% accuracy).
• Identified single subjects out of 138 using 39 functional connectivity profiles.
• Contrastive method demonstrated flexibility and robustness in 'leave subject out' scenarios.

Conclusions:

• Contrastive learning and data augmentation provide a scalable solution for neural fingerprinting.
• The proposed method is robust across varying data sizes, addressing limited sample challenges.
• This approach enhances the practical application of deep learning in individual identification via brain activity.