AFoCo: Ambiguous Focus and Correction for Semi-Supervised Medical Image Segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate medical image segmentation is vital for disease diagnosis and treatment planning.
• Deep learning in semi-supervised segmentation struggles with ambiguous regions of high predictive volatility.
• Ambiguous regions in unlabeled data offer valuable complementary information for improving segmentation models.

Purpose of the Study:

• To propose an innovative ambiguous focusing and correction (AFoCo) framework to address limitations in semi-supervised medical image segmentation.
• To accurately capture and refine ambiguous regions with high predictive volatility.
• To enhance the overall stability and accuracy of medical image segmentation.

Main Methods:

• Developed a dual-network framework: an ambiguous focus network and an ambiguous correction network.
• The focus network uses historical prediction changes and information entropy to identify ambiguous regions.
• The correction network redistributes pixel labels in ambiguous areas using a weight-weighted similarity strategy and task-aware asymmetric cross-supervision.

Main Results:

• The proposed AFoCo framework demonstrated superior performance compared to state-of-the-art methods on four medical image datasets.
• AFoCo significantly improved segmentation accuracy.
• The framework effectively reduced the proportion of ambiguous regions in the segmentation output.

Conclusions:

• The AFoCo framework offers a novel and effective solution for semi-supervised medical image segmentation.
• By precisely focusing on and correcting ambiguous regions, AFoCo enhances segmentation quality and reliability.
• This approach holds significant potential for advancing clinical applications requiring precise medical image analysis.