Uncertainty-guided mutual consistency learning for semi-supervised medical image segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Medical image segmentation is crucial for clinical applications.
• Semi-supervised learning (SSL) reduces annotation burden by using unlabeled data.
• Existing SSL methods struggle to fully utilize shape and boundary information from unlabeled data.

Purpose of the Study:

• To propose a novel uncertainty-guided mutual consistency learning framework for semi-supervised medical image segmentation.
• To effectively exploit unlabeled data by integrating intra-task and cross-task consistency learning.
• To improve the utilization of region-level shape and boundary-level distance information.

Main Methods:

• Developed an uncertainty-guided mutual consistency learning framework.
• Integrated intra-task consistency learning for self-ensembling.
• Incorporated cross-task consistency learning with task-level regularization for geometric shape information.
• Utilized estimated segmentation uncertainty to select reliable predictions for consistency learning.

Main Results:

• Achieved significant performance improvements using unlabeled data, with Dice coefficient increases of up to 4.13% (left atrium) and 9.82% (brain tumor) over supervised baselines.
• Demonstrated superior segmentation performance compared to other SSL methods on benchmark datasets.
• Showcased the method's effectiveness, robustness, and potential transferability to other medical image segmentation tasks.

Conclusions:

• The proposed framework effectively leverages unlabeled data for enhanced medical image segmentation.
• Uncertainty guidance improves the reliability of information extracted from unlabeled data.
• The method offers a robust and transferable solution for semi-supervised medical image segmentation challenges.