Meta Transfer of Self-Supervised Knowledge: Foundation Model in Action for Post-Traumatic Epilepsy Prediction

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Deep learning for brain activity analysis faces challenges with heterogeneous functional patterns and scarce data, particularly for predicting Post-Traumatic Epilepsy (PTE) after traumatic brain injury (TBI).
• Foundation models, while promising, struggle with scarce clinical data, limiting their ability to identify relevant clinical features from functional Magnetic Resonance Imaging (fMRI).

Approach:

• Introduced a novel foundation model training strategy integrating meta-learning with self-supervised learning to enhance generalization from healthy to clinical brain imaging data.
• Utilized self-supervised training on control datasets to focus on inherent features and meta-learning with bi-level optimization to improve model generalizability.
• Applied the trained foundation model to predict PTE on an unseen TBI dataset using zero-shot learning.

Key Points:

• The proposed strategy significantly improves task performance on small-scale clinical datasets for neurological disorder classification.
• Demonstrated enhanced generalizability of the foundation model in downstream applications, specifically for PTE prediction.
• Meta-learning combined with self-supervised learning effectively addresses data scarcity and heterogeneity in clinical neuroimaging analysis.

Conclusions:

• The novel training strategy enables foundation models to generalize effectively from abundant healthy control data to scarce clinical data for tasks like PTE prediction.
• The developed foundation model shows strong potential for improving the prediction of neurological disorders from limited neuroimaging datasets.
• This approach offers a promising direction for leveraging large-scale datasets to advance clinical applications of deep learning in neuroscience.