Abstract
Deep learning models for medical image segmentation often struggle with "noisy" training labels—imprecise or inconsistent annotations common in clinical practice. We present GSD-Net, a novel framework that integrates geometric distance weighting, superpixel-based structural refinement, and cross-sample knowledge transfer. By leveraging the intrinsic anatomical structure of images, GSD-Net effectively filters annotation errors, outperforming state-of-the-art methods across multiple imaging modalities and noise types.
Introduction: The Challenge of Noisy Annotations
High-quality, pixel-perfect annotations are the bedrock of medical image segmentation models. However, obtaining such "ground truth" is difficult. Radiologists often produce coarse delineations to manage high workloads, and different experts frequently disagree on boundary locations (inter-observer variability). Standard convolutional neural networks (CNNs) tend to "memorize" these label errors, leading to poor generalization.
Existing noise-robust techniques often rely on simple statistical heuristics, such as the "small-loss" strategy, which ignores the spatial and anatomical context. GSD-Net addresses this by focusing on spatially structured noise—errors that typically cluster around complex anatomical boundaries.
Methodology: The GSD-Net Framework
GSD-Net employs a dual-network co-regularization strategy where two models train simultaneously, cross-verifying each other's predictions. This process is enhanced by three specialized modules designed to handle the unique characteristics of medical image noise.
1. Geometric Distance-Aware (GDA) Module
Label noise is most prevalent near object boundaries. The GDA module calculates a distance map from noisy boundaries and dynamically assigns higher loss weights to pixels further from the boundary (the "safe" zones). This suppresses the influence of potentially erroneous boundary pixels during the early stages of training.
2. Structure-Guided Label Refinement (SGLR) Module
To ensure anatomical plausibility, the SGLR module utilizes SLIC superpixels—clusters of pixels with similar intensities. It enforces a structural prior: pixels within the same superpixel should likely share the same label. This allows the image's own intrinsic structure to "clean" noisy labels.
3. Knowledge Transfer (KT) Module
The KT module swaps local regions between pairs of training images. This forces the networks to learn local structural patterns rather than overfitting to specific noise patterns in individual images, thereby increasing data diversity and robustness.
Experimental Evaluation
The framework was rigorously evaluated on six public datasets across diverse modalities:
- Simulated Noise: Kvasir (Endoscopy), Shenzhen (X-ray), BU_SUC (Ultrasound), and BraTS2020 (MRI). Realistic under-annotation ($S_R$) and over-annotation ($S_E$) were simulated using Markov-based boundary perturbation.
- Real-World Noise: LIDC (CT) and MMIS-2024 (MRI), which contain annotations from multiple independent experts, capturing genuine clinical subjectivity.
Key Insights and Results
GSD-Net achieved state-of-the-art performance, particularly in high-noise scenarios. In the Shenzhen (X-ray) dataset, the model improved the Dice score by 22.76% under simulated under-annotation. On real-world multi-rater data (LIDC and MMIS-2024), GSD-Net significantly outperformed models trained on single-expert labels.
A key insight from this work is that medical segmentation noise is spatially structured at boundaries. By treating anatomy as a regularizer through superpixels, the model uses the image itself to correct the labels. Ablation studies confirmed that the synergy between geometric weighting (GDA) and structural refinement (SGLR) is critical for preventing early overfitting. Furthermore, GSD-Net is architecture-agnostic, showing performance gains when applied to various backbones, including UNet++ and Vision Transformers (ViT).
Limitations and Future Directions
While GSD-Net is highly effective for binary segmentation, extending it to multi-class scenarios with overlapping organs remains a challenge. Additionally, the dual-network strategy and superpixel processing increase training complexity, although inference speed is unaffected. Future work will focus on multi-center validation to ensure robustness against hardware-induced artifacts.
Summary updated