Role Exchange-Based Self-Training Semi-Supervision Framework for Complex Medical Image Segmentation.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Accurate segmentation of complex medical images like vascular and pulmonary networks is challenging due to the need for numerous tiny target annotations.
• Fully supervised deep learning models require extensive manual annotations, hindering segmentation of intricate structures.

Purpose of the Study:

• To develop an efficient semi-supervised model for complex medical image segmentation.
• To reduce the reliance on large annotated datasets for medical image segmentation tasks.

Main Methods:

• Proposed a bidirectional self-training paradigm with dynamically exchanging teacher-student roles based on model reliability.
• Introduced asymmetric supervision (AS) and hierarchical dual student (HDS) structures to prevent model collapse on small datasets.
• Implemented bidirectional distillation loss with role exchange (RE) and global-local-aware consistency loss for stable feature matching.

Main Results:

• The proposed semi-supervised model significantly outperformed existing methods on public and private datasets.
• Achieved performance comparable to fully supervised methods with only 5% of the labeling cost.
• Demonstrated stable mutual promotion and effective matching of global and local features.

Conclusions:

• The bidirectional self-training model offers a highly effective solution for complex medical image segmentation with minimal annotation requirements.
• The AS strategy and HDS structure successfully address the challenges of training on small-scale annotated data.
• This approach represents a significant advancement in efficient and accurate medical image segmentation.