Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MetaDCSeg: Robust Medical Image Segmentation via Meta Dynamic Center Weighting

Chenyu Mu, Guihai Chen, Xun Yang, Erkun Yang, Cheng Deng

TL;DR

MetaDCSeg addresses noisy annotations and boundary ambiguity in medical image segmentation by combining pixel-wise meta-learning with a Dynamic Center Distance (DCD) boundary refinement. It learns per-pixel reliability weights to balance real and pseudo-label supervision, and employs uncertainty-aware region fragmentation with weighted center distances to emphasize hard boundary pixels. The method integrates these components into a unified training objective that includes Dice loss for global consistency, and demonstrates robust gains across four benchmarks under varying noise ratios, while maintaining strong performance on clean data. This approach offers a practical, scalable solution for robust medical image segmentation in real-world, noisy annotation settings.

Abstract

Medical image segmentation is crucial for clinical applications, but it is frequently disrupted by noisy annotations and ambiguous anatomical boundaries, which lead to instability in model training. Existing methods typically rely on global noise assumptions or confidence-based sample selection, which inadequately mitigate the performance degradation caused by annotation noise, especially in challenging boundary regions. To address this issue, we propose MetaDCSeg, a robust framework that dynamically learns optimal pixel-wise weights to suppress the influence of noisy ground-truth labels while preserving reliable annotations. By explicitly modeling boundary uncertainty through a Dynamic Center Distance (DCD) mechanism, our approach utilizes weighted feature distances for foreground, background, and boundary centers, directing the model's attention toward hard-to-segment pixels near ambiguous boundaries. This strategy enables more precise handling of structural boundaries, which are often overlooked by existing methods, and significantly enhances segmentation performance. Extensive experiments across four benchmark datasets with varying noise levels demonstrate that MetaDCSeg consistently outperforms existing state-of-the-art methods.

MetaDCSeg: Robust Medical Image Segmentation via Meta Dynamic Center Weighting

TL;DR

MetaDCSeg addresses noisy annotations and boundary ambiguity in medical image segmentation by combining pixel-wise meta-learning with a Dynamic Center Distance (DCD) boundary refinement. It learns per-pixel reliability weights to balance real and pseudo-label supervision, and employs uncertainty-aware region fragmentation with weighted center distances to emphasize hard boundary pixels. The method integrates these components into a unified training objective that includes Dice loss for global consistency, and demonstrates robust gains across four benchmarks under varying noise ratios, while maintaining strong performance on clean data. This approach offers a practical, scalable solution for robust medical image segmentation in real-world, noisy annotation settings.

Abstract

Medical image segmentation is crucial for clinical applications, but it is frequently disrupted by noisy annotations and ambiguous anatomical boundaries, which lead to instability in model training. Existing methods typically rely on global noise assumptions or confidence-based sample selection, which inadequately mitigate the performance degradation caused by annotation noise, especially in challenging boundary regions. To address this issue, we propose MetaDCSeg, a robust framework that dynamically learns optimal pixel-wise weights to suppress the influence of noisy ground-truth labels while preserving reliable annotations. By explicitly modeling boundary uncertainty through a Dynamic Center Distance (DCD) mechanism, our approach utilizes weighted feature distances for foreground, background, and boundary centers, directing the model's attention toward hard-to-segment pixels near ambiguous boundaries. This strategy enables more precise handling of structural boundaries, which are often overlooked by existing methods, and significantly enhances segmentation performance. Extensive experiments across four benchmark datasets with varying noise levels demonstrate that MetaDCSeg consistently outperforms existing state-of-the-art methods.

Paper Structure

This paper contains 40 sections, 46 equations, 4 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Medical Image Segmentation
  4. Learning with Noisy Labels
  5. Meta Learning for Label Correction
  6. Methodology
  7. Problem Formulation and Overview
  8. Meta-Learning for Pixel-wise Noise Suppression
  9. Pixel-wise Bootstrapping Loss
  10. Weight Initialization Strategy
  11. Meta-Learning Objective
  12. Dynamic Center Distance Boundary Refinement
  13. Feature Extraction
  14. Uncertainty-Aware Region Decomposition
  15. Meta-Weighted Dynamic Centers
  16. ...and 25 more sections

Figures (4)

  • Figure 1: Boundary challenges in medical image segmentation and our framework.
  • Figure 2: The proposed MetaDCSeg architecture with pixel-wise meta-learning and dynamic boundary refinement modules.
  • Figure 3: Spatial distribution and distance-dependency of uncertainty weight.
  • Figure 4: Visual comparison of segmentation methods on Kvasir-SEG, PROMISE12, MSD Heart, and MSD Brain datasets.