Evidence-based uncertainty-aware semi-supervised medical image segmentation.

Area of Science:

• Medical Image Analysis
• Artificial Intelligence
• Machine Learning

Background:

• Semi-Supervised Learning (SSL) reduces annotation needs in clinical settings.
• Pseudo-labeling in SSL can introduce performance-degrading errors.
• Existing uncertainty-aware methods face trade-offs in cost, accuracy, and theory.

Purpose of the Study:

• To introduce Evidential Inference Learning (EVIL) for uncertainty-aware SSL in medical image segmentation.
• To address limitations of current uncertainty-aware SSL methods.
• To improve the reliability and accuracy of medical image segmentation.

Main Methods:

• Integrated Dempster-Shafer Theory of Evidence (DST) into SSL.
• Developed EVIL as a consistency regularization training paradigm.
• Enabled precise uncertainty quantification within a single forward pass.
• Discarded unreliable pseudo-labels based on uncertainty estimation.

Main Results:

• EVIL demonstrated competitive performance against state-of-the-art methods.
• The approach was validated on public datasets: ACDC, MM-WHS, and MonuSeg.
• Achieved reliable pseudo-label generation and improved segmentation.

Conclusions:

• EVIL offers a theoretically assured and computationally efficient solution for uncertainty quantification in SSL.
• The method effectively improves medical image segmentation by leveraging trustworthy pseudo-labels.
• EVIL provides a promising direction for robust SSL in clinical applications.