[Paper Review] Pathological Image Segmentation with Noisy Labels

Li Xiao, Yinhao Li+4 more|arXiv (Cornell University)|Mar 20, 2021

AI in cancer detection20 references4 citations

TL;DR

This paper proposes a novel label re-weighting framework and Gaussian attention focal loss (GAFL) to improve pathological image segmentation under noisy multi-expert labels. By leveraging pixel-level reliability via surrounding features and roughness-aware attention, the method boosts robustness and achieves state-of-the-art performance on the Gleason 2019 dataset, especially improving segmentation of challenging Gleason grade 5 regions by up to 29.37% Dice score.

ABSTRACT

Segmentation of pathological images is essential for accurate disease diagnosis. The quality of manual labels plays a critical role in segmentation accuracy; yet, in practice, the labels between pathologists could be inconsistent, thus confusing the training process. In this work, we propose a novel label re-weighting framework to account for the reliability of different experts' labels on each pixel according to its surrounding features. We further devise a new attention heatmap, which takes roughness as prior knowledge to guide the model to focus on important regions. Our approach is evaluated on the public Gleason 2019 datasets. The results show that our approach effectively improves the model's robustness against noisy labels and outperforms state-of-the-art approaches.

Motivation & Objective

• To address the challenge of inconsistent annotations from multiple pathologists in pathological image segmentation.
• To reduce model bias caused by noisy, inter-observer label variability in clinical pathology datasets.
• To improve segmentation robustness and accuracy, especially for high-grade Gleason patterns (e.g., grade 5), using multi-expert label integration.
• To develop a reliable, learnable mechanism that weights labels based on local image context and expert agreement.
• To validate the method on the public Gleason 2019 benchmark, demonstrating superior performance over state-of-the-art approaches.

Proposed method

• A label re-weighting framework uses a weighting CNN to assign dynamic loss weights per pixel based on surrounding features, reflecting expert reliability.
• The Gaussian attention focal loss (GAFL) module uses roughness as prior knowledge to generate a heatmap that reweights focal loss, emphasizing important regions.
• The framework combines a voting loss (from majority consensus) and weighted losses from individual experts, all modulated by the GAFL heatmap.
• The weighting CNN is trained end-to-end with the segmentation network, optimizing for reduced loss on uncertain or noisy regions.
• The loss function dynamically adjusts hyperparameters λ₁ and λ₂ during training using a schedule based on training iteration count.
• The model is trained on the Gleason 2019 dataset with HRNet18 and HRNet48 backbones, using early stopping based on validation performance.

Experimental results

Research questions

• RQ1 Can a learnable label re-weighting mechanism improve segmentation performance when expert annotations are inconsistent?
• RQ2 How does incorporating roughness as a prior in the loss function affect model attention and segmentation accuracy?
• RQ3 To what extent can the proposed GAFL and re-weighting modules reduce the negative impact of noisy multi-expert labels on pathological image segmentation?
• RQ4 Does the framework generalize across different backbone architectures and improve performance on difficult Gleason grades, particularly grade 5?
• RQ5 Can the method outperform existing state-of-the-art models on the public Gleason 2019 benchmark under real-world noisy labeling conditions?

Key findings

• The proposed method achieves a mean Dice coefficient of 66.21% on the Gleason 2019 test set using HRNet18, outperforming the previous SOTA PSPNet (55.53%).
• The addition of both the label re-weighting and GAFL modules improves the Dice score for Gleason grade 5 by 29.37% on HRNet18 and 21.19% on HRNet48, compared to baseline.
• The GAFL module alone improves the Dice score by 4.07% on HRNet18, indicating its strong impact on focusing on critical regions.
• The label re-weighting module alone improves performance by 0.36% on HRNet18, showing incremental but meaningful gains in handling label noise.
• The ablation study confirms that both modules consistently improve segmentation across different backbones, with the most significant gains in hard-to-segment regions like Gleason grade 5.
• Visualization results demonstrate that the proposed method produces more accurate and coherent segmentation maps, especially in distinguishing high-grade cancerous areas.