CSBNC-PAL: Consistency Semi-Supervised Brain Network Classification Framework With Prototypical-Adversarial Learning.

Area of Science:

• Neuroscience
• Machine Learning
• Data Science

Background:

• Semi-supervised learning (SSL) shows promise for functional brain network (FBN) classification by utilizing unlabeled multisite data.
• Existing SSL methods face challenges with distributional differences across sites, limiting feature extraction and classification performance.

Purpose of the Study:

• To propose a novel consistency semi-supervised FBN classification framework with prototypical-adversarial learning (CSBNC-PAL).
• To address the limitations of existing SSL methods in handling multisite data variations for improved FBN classification.

Main Methods:

• A contrastive consistency module (CCM) for effective exploitation of unlabeled data and preliminary feature representation learning.
• A prototype alignment module (PAM) for site-aware prototype computation and inter-site feature alignment.
• An adversarial alignment module (AAM) using gradient reversal for intra-site alignment and learning site-invariant features.

Main Results:

• The proposed CSBNC-PAL framework integrates CCM, PAM, and AAM for end-to-end optimization.
• The method effectively learns from both labeled and unlabeled data while mitigating multisite data distribution differences.
• Experiments on ABIDE I, ABIDE II, and ADHD-200 datasets show superior performance compared to state-of-the-art SSL methods.

Conclusions:

• CSBNC-PAL offers a robust solution for semi-supervised functional brain network classification in multisite settings.
• The framework successfully addresses inter-site distributional variations, leading to enhanced classification accuracy.
• The findings highlight the potential of prototypical-adversarial learning for advancing brain network analysis.