Hierarchical Bias Mitigation for Semi-Supervised Medical Image Classification.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Medical Imaging

Background:

• Semi-supervised learning (SSL) leverages unlabeled data for medical image classification.
• Pseudo-labeling is a dominant SSL technique but suffers from inherent biases.
• These biases include perception, selection, and confirmation biases at different stages of the learning process.

Purpose of the Study:

• To propose a novel framework, HierArchical BIas miTigation (HABIT), to address hierarchical biases in pseudo-labeling for SSL.
• To improve the performance and reliability of medical image classification models.

Main Methods:

• Proposed the HABIT framework with three modules: Mutual Reconciliation Network (MRNet), Recalibrated Feature Compensation (RFC), and Consistency-aware Momentum Heredity (CMH).
• MRNet reconciles spatial perception bias using convolution and permutator paths with mutual information transfer.
• RFC addresses selection bias by recalibrating augmented distributions and balancing minority categories.
• CMH reduces confirmation bias by incorporating consistency across augmentations into model updates.

Main Results:

• HABIT effectively mitigates perception, selection, and confirmation biases in semi-supervised medical image classification.
• Experimental results on three datasets demonstrate state-of-the-art performance.
• The proposed modules contribute to improved feature representation, balanced training, and dependable model optimization.

Conclusions:

• The HABIT framework offers a comprehensive solution for bias mitigation in SSL for medical imaging.
• This approach enhances the accuracy and robustness of medical image classification.
• The findings suggest a promising direction for future research in bias-aware SSL.