Adaptive Label Correction for Robust Medical Image Segmentation with Noisy Labels

TL;DR

This paper tackles the challenge of robust medical image segmentation when labels are noisy by introducing Adaptive Label Correction (ALC), a Mean Teacher-based self-ensemble framework. ALC combinesHQ-label learning with low-quality label learning through adaptive label refinement and uncertainty-guided sample selection, complemented by a consistency loss that aligns student and teacher predictions under perturbations. The approach demonstrates that dynamic refinement and selective supervision substantially improve segmentation accuracy and boundary delineation in synthetic and real-world noisy-label settings, outperforming both label-agnostic and label-aware baselines across multiple datasets. The work offers a practical strategy for leveraging imperfect annotations in clinical contexts, reducing the reliance on perfectly labeled data while maintaining high performance.

Abstract

Deep learning has shown remarkable success in medical image analysis, but its reliance on large volumes of high-quality labeled data limits its applicability. While noisy labeled data are easier to obtain, directly incorporating them into training can degrade model performance. To address this challenge, we propose a Mean Teacher-based Adaptive Label Correction (ALC) self-ensemble framework for robust medical image segmentation with noisy labels. The framework leverages the Mean Teacher architecture to ensure consistent learning under noise perturbations. It includes an adaptive label refinement mechanism that dynamically captures and weights differences across multiple disturbance versions to enhance the quality of noisy labels. Additionally, a sample-level uncertainty-based label selection algorithm is introduced to prioritize high-confidence samples for network updates, mitigating the impact of noisy annotations. Consistency learning is integrated to align the predictions of the student and teacher networks, further enhancing model robustness. Extensive experiments on two public datasets demonstrate the effectiveness of the proposed framework, showing significant improvements in segmentation performance. By fully exploiting the strengths of the Mean Teacher structure, the ALC framework effectively processes noisy labels, adapts to challenging scenarios, and achieves competitive results compared to state-of-the-art methods.

Figures (6)

• Figure 1: Visualization of noisy labels: The first and third rows display LQ labels from ground truth expansion, while the second and fourth rows depict those from ground truth erosion.
• Figure 2: Overview of our proposed ALC approach, which employs a student-teacher network to integrate label correction and selection. The total loss comprising high-quality label loss $\mathcal{L}_{hs}$, low-quality label loss $\mathcal{L}_{ls}$, noisy label loss $\mathcal{L}_{n}$, and consistency loss $\mathcal{L}_{c}$.
• Figure 3: Quantitative comparisons of pseudo and corrected label quality across iterations under varying HQ and LQ label ratios on the LA dataset (a)-(b) and NIH dataset (c)-(d).
• Figure 4: Visual comparison of segmentation results on the NIH dataset under the 10% HQ and 90% LQ label setting across different methods
• Figure 5: The impact of different $k$ values on the LA and NIH datasets under 10% HQ label data and 90% LQ label data.