DUCore: Dual Uncertainty-Guided Consistency and Regional Contrastive Learning for Semi-Supervised Medical Image...

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Semi-supervised learning is crucial for medical image segmentation, but existing uncertainty estimation methods increase computational cost and may discard valuable data.
• Current approaches often miss complex structures like ambiguous lesion boundaries due to discarding uncertain regions.

Purpose of the Study:

• To introduce the Dual Uncertainty-Guided Consistency and Regional Contrastive Learning (DUCore) framework for improved medical image segmentation.
• To address the limitations of existing uncertainty estimation methods in terms of computational cost and data handling.

Main Methods:

• DUCore integrates dual uncertainty-guided consistency loss (DuCL) and Regional Contrastive Loss (ReCL).
• DuCL uses deterministic single-pass uncertainty estimation (entropy-based for aleatoric, Proxy Dirichlet for epistemic) and weights uncertain regions.
• ReCL employs boundary- and gradient-based hard negative mining for enhanced feature separability.

Main Results:

• DUCore improves segmentation robustness by adaptively calibrating prediction alignment.
• The framework effectively delineates fine structures and complex boundaries with higher precision.
• Experiments show DUCore outperforms existing consistency-based methods on medical segmentation benchmarks.

Conclusions:

• DUCore offers a more efficient and effective approach to uncertainty-aware consistency learning in medical image segmentation.
• The method preserves valuable learning signals by weighting, rather than discarding, uncertain regions.
• DUCore demonstrates superior performance in handling complex structures and ambiguous boundaries.