Breaking error coupling via divergent-convergent coordination for semi-supervised medical image segmentation

Area of Science:

• Medical Image Analysis
• Artificial Intelligence
• Machine Learning

Background:

• Semi-supervised medical image segmentation faces challenges like error coupling and model homogenization.
• Existing methods often rely on strict prediction consistency, limiting performance.

Purpose of the Study:

• To introduce a novel Divergent-Convergent Framework (DCF) for semi-supervised medical image segmentation.
• To address error coupling and model homogenization by dynamically managing model disagreement.

Main Methods:

• Proposed a Divergent-Convergent Framework (DCF) utilizing a Guidance Mask (GM) to quantify model confidence and disagreement.
• Implemented a Convergence Stabilization Mechanism for high-confidence regions and a Divergent Exploration Mechanism for low-confidence regions.
• Tested on six public datasets (2D and 3D) and fine-tuned foundation models like MedSAM and MedSAM2.

Main Results:

• DCF significantly outperformed state-of-the-art semi-supervised methods across 5%, 10%, and 20% labeled data ratios.
• Demonstrated advantages in segmenting ambiguous boundaries and robustness to domain shifts.
• Showcased consistent performance improvements when fine-tuning foundation models like MedSAM and MedSAM2.

Conclusions:

• DCF offers a practical, label-efficient strategy for semi-supervised medical image segmentation.
• The framework effectively maintains model diversity while suppressing shared errors.
• DCF shows significant potential for clinical deployment and improving foundation model performance.