CoRe: An End-to-End Collaborative Refinement Network for Medical Image Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Medical image segmentation is vital for clinical diagnosis and treatment.
• Deep learning models, particularly encoder-decoder networks, show promise but suffer from spatial detail loss due to network depth and downsampling.
• Existing methods lack targeted solutions for improving segmentation performance by addressing these inherent flaws.

Purpose of the Study:

• To propose an end-to-end collaborative refinement method (CoRe) to address the limitations of current deep networks in medical image segmentation.
• To improve the reconstruction of lost spatial detail information and refine segmentation predictions.

Main Methods:

• CoRe generates an Error-Prone Region (EPR) using uncertainty and foreground boundary maps.
• A Feature Refinement Module (FRM) refines decoder upsampling features using neighborhood-aware and foreground-boundary-enhanced features.
• A Segmentation Refinement Module (SRM) refines coarse predictions using global class centers.

Main Results:

• The proposed CoRe method achieved significant improvements in medical image segmentation across five diverse datasets.
• CoRe demonstrated competitive performance compared to current state-of-the-art methods.
• The method effectively reconstructs lost spatial detail information and refines segmentation predictions.

Conclusions:

• The end-to-end collaborative refinement method (CoRe) effectively addresses spatial detail loss in deep medical image segmentation.
• CoRe enhances segmentation accuracy and shows strong potential for clinical applications.
• The proposed approach offers a significant advancement in medical image analysis and diagnostic decision-making.