Exploring the Connection between Uncertainty and Tissue Boundaries in Medical Image Segmentation

Area of Science:

• Medical Image Analysis
• Computer-Aided Diagnosis
• Machine Learning for Healthcare

Background:

• Automatic medical image segmentation is crucial for quantitative analysis and computer-aided diagnosis.
• Blurry boundaries in medical images, due to imaging quality or tissue properties, lead to imprecise segmentation and misclassification.
• Uncertainty is commonly observed at object boundaries during segmentation tasks.

Purpose of the Study:

• To investigate the relationship between uncertainty and tissue boundaries in medical image segmentation.
• To propose a novel loss function, Evidential Uncertainty-Guided Boundary (EUGB) loss, to address boundary segmentation errors.
• To demonstrate the effectiveness of uncertainty information in improving segmentation accuracy.

Main Methods:

• Developed the Evidential Uncertainty-Guided Boundary (EUGB) loss function.
• Incorporated evidential uncertainty to highlight challenging pixels at blurry boundaries.
• Added a regularization term to constrain uncertainty learning, penalizing incorrect predictions and reinforcing correct ones.
• Validated the EUGB loss on U-Net and TransU-Net architectures using LIDC-IDRI, ISIC 2018, and OCTA-500 datasets.

Main Results:

• The proposed EUGB loss significantly improved boundary segmentation performance compared to seven other loss functions.
• Maintained competitive region-level segmentation accuracy.
• Demonstrated superior performance in handling blurry object boundaries.

Conclusions:

• The EUGB loss effectively combats boundary segmentation errors by leveraging uncertainty information.
• Provides practical insights for selecting appropriate loss functions based on dataset characteristics and application scenarios.
• Enhances the reliability of medical image segmentation for clinical applications.