Semantic contrast with uncertainty-aware pseudo label for lumbar semi-supervised classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Lumbar disc herniation (LDH) diagnosis relies on MRI, but manual annotation is labor-intensive.
• Semi-supervised learning (SSL) uses limited labeled and abundant unlabeled data for deep learning.
• Existing SSL methods like pseudo labeling and consistency regularization have limitations.

Purpose of the Study:

• To introduce SeCoFixMatch, an SSL approach for efficient LDH diagnosis from MRI.
• To address limitations of current SSL techniques, specifically confirmation bias in pseudo labeling and lack of guidance in consistency regularization.
• To improve the accuracy and reduce annotation effort in deep learning-based LDH diagnosis.

Main Methods:

• SeCoFixMatch integrates semantic contrast and uncertainty-aware pseudo labeling into SSL.
• Semantic contrast ensures semantic consistency between labeled and unlabeled MRI images.
• Uncertainty-aware pseudo labeling combines predictive confidence and uncertainty, calculated via KL loss with a Dirichlet distribution.

Main Results:

• SeCoFixMatch demonstrates superior effectiveness and generalization compared to other SSL models.
• Ablation studies confirm the model's performance across varying amounts of labeled data.
• With only 40 labeled images, SeCoFixMatch surpassed a baseline model trained on 200 labels, an 80% reduction in annotation effort.

Conclusions:

• The proposed pseudo labeling algorithm generates precise labels, enhancing semantic contrastive learning.
• Semantic contrastive learning improves feature representation, boosting pseudo-label prediction accuracy.
• The synergistic interaction between pseudo labeling and semantic contrast significantly enhances SSL performance for LDH diagnosis.