Uncertainty Co-Estimator for Improving Semi-Supervised Medical Image Segmentation.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Semi-supervised learning enhances medical image segmentation using limited labeled data.
• Consistency regularization and uncertainty estimation are key strategies.
• Existing methods suffer from imprecise uncertainty estimation due to single-model reliance.

Purpose of the Study:

• To propose a novel Uncertainty Co-estimator (UnCo) framework for improved semi-supervised medical image segmentation.
• To address the limitations of imprecise uncertainty estimation in current methods.
• To enhance segmentation performance through more accurate uncertainty quantification.

Main Methods:

• UnCo employs two mean-teacher modules (teacher-student pairs) inspired by co-training.
• It estimates three types of uncertainty from multi-source predictions.
• Fused uncertainty maps enhance cross-consistency regularization, while internal regularization and adversarial constraints promote diversity.

Main Results:

• UnCo achieves new state-of-the-art performance on both 2D and 3D semi-supervised segmentation tasks.
• The framework demonstrates improved accuracy in uncertainty estimation.
• Experimental validation across four medical image datasets confirms effectiveness.

Conclusions:

• The proposed UnCo framework significantly advances semi-supervised medical image segmentation.
• Accurate uncertainty estimation via multi-source prediction is crucial for performance gains.
• UnCo offers a robust and effective approach for medical image segmentation challenges.