From Noisy Labels to Intrinsic Structure: A Geometric-Structural Dual-Guided Framework for Noise-Robust Medical Image Segmentation

Abstract

The effectiveness of convolutional neural networks in medical image segmentation relies on large-scale, high-quality annotations, which are costly and time-consuming to obtain. Even expert-labeled datasets inevitably contain noise arising from subjectivity and coarse delineations, which disrupt feature learning and adversely impact model performance. To address these challenges, this study propose a Geometric-Structural Dual-Guided Network (GSD-Net), which integrates geometric and structural cues to improve robustness against noisy annotations. It incorporates a Geometric Distance-Aware module that dynamically adjusts pixel-level weights using geometric features, thereby strengthening supervision in reliable regions while suppressing noise. A Structure-Guided Label Refinement module further refines labels with structural priors, and a Knowledge Transfer module enriches supervision and improves sensitivity to local details. To comprehensively assess its effectiveness, we evaluated GSD-Net on six publicly available datasets: four containing three types of simulated label noise, and two with multi-expert annotations that reflect real-world subjectivity and labeling inconsistencies. Experimental results demonstrate that GSD-Net achieves state-of-the-art performance under noisy annotations, achieving improvements of 1.58% on Kvasir, 22.76% on Shenzhen, 8.87% on BU-SUC, and 1.77% on BraTS2020 under SR simulated noise. The codes of this study are available at https://github.com/ortonwang/GSD-Net.

Figures (11)

• Figure 1: Illustration of label noise sources: (a) coarse annotations, (b) intra-observer variability (inconsistencies by the same expert under different conditions), and (c) inter-observer variability (annotations from different experts shown in different colors).
• Figure 2: Illustration of collaborative learning among modules. Small Loss Criterion (Sec. \ref{['method_jocor']}); GDA: Geometric Distance-Aware (Sec. \ref{['Probabilistic_aware']}); SGLR: Structure-Guided Label Refinement (Sec. \ref{['dynamic_fusion']}); KT: Knowledge Transfer (Sec. \ref{['transfer_section']}). The texts along the arrows describe the functional effects that one module exerts on another. $S_{\mathrm{1}}$-$S_{\mathrm{4}}$ denote different stages.
• Figure 3: Schematic diagram of the proposed GSD-Net framework: (a) overall workflow, (b) Geometric Distance-Aware module, and (c) Structure-Guided Label Refinement module. In subfigure (b), $\mathbb{D}^{clean}$ represents the reliable labeled regions selected by the small-loss criterion, while $\overline{\mathbb{D}^{clean}}$ represents the complementary set.
• Figure 4: The schematic diagram of the Geometric Distance-Aware Module.
• Figure 5: Visualization of simulated label noise ($S_R$: foreground-reducing, $S_E$: foreground-expanding, $S_{DE}$: simulated via dilation or erosion) and inter-expert variability. In the upper panel, red contours represent ground truth boundaries, and green contours indicate simulated noisy labels.