TriCSART: Semi-Supervised Medical Image Segmentation with Triple-Level Contrastive Learning and Selective Active...

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Semi-supervised medical image segmentation faces challenges with limited data and high annotation costs.
• Existing methods lack hierarchical contrastive learning and robust pseudo-label quality assessment.
• Annotation biases can propagate due to errors in conventional pseudo-labeling.

Purpose of the Study:

• To introduce a novel semi-supervised medical image segmentation framework addressing limitations of current methods.
• To enhance feature representation and mitigate annotation bias through advanced learning strategies.
• To achieve state-of-the-art performance in medical image segmentation.

Main Methods:

• Proposed a teacher-student architecture incorporating Triple-Level Contrastive (TLC) Learning and Selective Active Re-Training (SART).
• TLC module uses three complementary loss functions for multilevel semantic consistency, improving inter-class and intra-class representation.
• SART module employs a metric-driven mechanism for salient sample selection and curriculum-guided re-training.

Main Results:

• Achieved state-of-the-art performance on five diverse benchmarks, including public and private datasets.
• Consistently outperformed existing semi-supervised medical image segmentation strategies.
• Ablation studies validated the effectiveness of individual components (TLC and SART).

Conclusions:

• The proposed TLC and SART strategy significantly advances semi-supervised medical image segmentation.
• The method effectively addresses data scarcity and annotation cost challenges.
• Demonstrated superior performance and robustness across multiple datasets.