Contour-Aware Multi-Expert Model for Ambiguous Medical Image Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Computational Biology

Background:

• Medical image segmentation faces challenges due to ambiguous regions and expert variability, particularly affecting critical boundary areas.
• Existing methods struggle to accurately segment these boundaries, limiting diagnostic utility.
• Previous approaches often focus on pixel-wise segmentation, neglecting contour-level nuances.

Purpose of the Study:

• To address limitations in medical image segmentation, especially concerning boundary accuracy and expert variations.
• To propose a novel contour-based regression approach for generating diverse and detailed segmentation results.
• To develop a Contour-aware Multi-expert Segmentor (ContourMS) for improved medical image analysis.

Main Methods:

• Formulated medical image segmentation as a contour-based regression problem, moving beyond pixel-wise methods.
• Developed ContourMS, a coarse-to-fine framework utilizing SegmentNet for initial mask prediction and multi-expert knowledge.
• Introduced LatentNet and ContourNet in the fine stage to learn expert-specific latent spaces and refine contours based on expert styles.

Main Results:

• ContourMS successfully generates diverse segmentation variants with rich boundary details.
• The proposed method achieves competitive performance on multiple public multi-expert medical segmentation datasets.
• Demonstrated the effectiveness of the contour-based approach in handling expert variations and improving boundary segmentation.

Conclusions:

• ContourMS offers a novel and effective approach to medical image segmentation by focusing on contour refinement.
• The method successfully addresses the challenges of ambiguous regions and expert knowledge variations, particularly at boundaries.
• ContourMS provides a promising direction for enhancing the accuracy and robustness of medical image analysis in clinical practice.