The student-teacher framework guided by self-training and consistency regularization for semi-supervised medical...

Area of Science:

• Medical Image Analysis
• Machine Learning
• Computer Vision

Background:

• Semi-supervised learning is highly effective for medical image segmentation.
• Existing methods often focus on single frameworks, limiting performance gains.
• Integrating multiple semi-supervised techniques can potentially improve segmentation accuracy.

Purpose of the Study:

• To propose a novel semi-supervised framework, Pseudo-Label Mean Teacher (PLMT), for medical image segmentation.
• To synergize self-training with pseudo-labeling and consistency regularization.
• To investigate the impact of adaptive loss weights on segmentation performance.

Main Methods:

• The proposed PLMT framework integrates a student-teacher structure with consistency loss into a self-training pipeline.
• Pseudo-labels are generated for both self-training and the student-teacher framework.
• Adaptive loss weights dynamically adjust the contribution of different semi-supervised losses during training.

Main Results:

• Experiments on three public datasets show PLMT achieves the best performance.
• PLMT outperforms five other semi-supervised methods.
• The integrated approach demonstrates mutually beneficial enhancements between self-training and consistency regularization.

Conclusions:

• The PLMT framework offers an innovative perspective in semi-supervised image segmentation.
• Combining self-training and consistency regularization yields significant performance improvements.
• PLMT provides a robust and effective approach for medical image segmentation tasks.