Uncertainty-guided decoupled network improves semi-supervised medical image segmentation

UGDC-Net outperformed state-of-the-art semi-supervised methods for binary segmentation on left atrium, Pancreas-CT, and ISIC datasets. It uses uncertainty-guided decoupling, dynamic competition, and complementary pseudo-labels.

• UGDC-Net uses a contrastive mechanism to filter high-uncertainty regions and a dynamic competition mechanism to reduce teacher-student weight coupling.
• Reliable and supplementary pseudo-labels are combined to provide diverse supervision and improve small-target exploration.
• Validation was performed on left atrium, Pancreas-CT, and ISIC datasets; the abstract reports state-of-the-art outperformance for binary segmentation.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.