Anatomically-aware uncertainty for semi-supervised image segmentation.

Area of Science:

• Medical image analysis
• Computer vision
• Machine learning

Background:

• Semi-supervised learning reduces reliance on large labeled datasets for image segmentation.
• Current uncertainty estimation methods are computationally expensive and lack global context.
• Existing approaches struggle with pixel-wise disparities and global information integration.

Purpose of the Study:

• To develop a novel, computationally efficient method for estimating segmentation uncertainty.
• To leverage global information from segmentation masks for improved uncertainty estimation.
• To enhance semi-supervised image segmentation accuracy in medical imaging.

Main Methods:

• Learned an anatomically-aware representation from available segmentation masks.
• Mapped new segmentation predictions to anatomically-plausible segmentations.
• Estimated pixel-level uncertainty based on deviations from plausible segmentations using a single inference.

Main Results:

• The proposed method significantly reduces computational cost compared to traditional uncertainty estimation.
• Achieved improved segmentation accuracy on cardiac MRI and abdominal CT datasets.
• Outperformed state-of-the-art semi-supervised methods in commonly used evaluation metrics.

Conclusions:

• The anatomically-aware approach effectively estimates segmentation uncertainty using global context.
• This method offers a computationally efficient and accurate alternative for semi-supervised medical image segmentation.
• The approach demonstrates potential for broader applications in medical image analysis.