Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Accurate Segmentation of Optic Disc And Cup from Multiple Pseudo-labels by Noise-aware Learning

Tengjin Weng, Yang Shen, Zhidong Zhao, Zhiming Cheng, Shuai Wang

TL;DR

This paper tackles the problem of label noise arising from boundary ambiguity in optic disc and cup segmentation. It introduces the Multiple Pseudo-labels Noise-aware Network (MPNN), combining the MPGGD module (which generates $K$ pseudo-labels via diverse initializations and thresholds the clean versus noisy pixels by consensus) with a Mean-Teacher style learning framework that uses an EMA-updated teacher to enforce uncertainty-guided consistency. Clean pixels drive supervised learning while noisy pixels contribute through a contrastive-like consistency loss, controlled by per-pixel uncertainty estimates and a ramped loss weight $\lambda(t)$. On the RIGA dataset, MPNN achieves state-of-the-art performance for both ground-truth references (Majority Vote and Rater1), demonstrating robust denoising of noisy annotations and improved segmentation accuracy, with code released for reproducibility.

Abstract

Optic disc and cup segmentation plays a crucial role in automating the screening and diagnosis of optic glaucoma. While data-driven convolutional neural networks (CNNs) show promise in this area, the inherent ambiguity of segmenting objects and background boundaries in the task of optic disc and cup segmentation leads to noisy annotations that impact model performance. To address this, we propose an innovative label-denoising method of Multiple Pseudo-labels Noise-aware Network (MPNN) for accurate optic disc and cup segmentation. Specifically, the Multiple Pseudo-labels Generation and Guided Denoising (MPGGD) module generates pseudo-labels by multiple different initialization networks trained on true labels, and the pixel-level consensus information extracted from these pseudo-labels guides to differentiate clean pixels from noisy pixels. The training framework of the MPNN is constructed by a teacher-student architecture to learn segmentation from clean pixels and noisy pixels. Particularly, such a framework adeptly leverages (i) reliable and fundamental insight from clean pixels and (ii) the supplementary knowledge within noisy pixels via multiple perturbation-based unsupervised consistency. Compared to other label-denoising methods, comprehensive experimental results on the RIGA dataset demonstrate our method's excellent performance. The code is available at https://github.com/wwwtttjjj/MPNN

Accurate Segmentation of Optic Disc And Cup from Multiple Pseudo-labels by Noise-aware Learning

TL;DR

This paper tackles the problem of label noise arising from boundary ambiguity in optic disc and cup segmentation. It introduces the Multiple Pseudo-labels Noise-aware Network (MPNN), combining the MPGGD module (which generates pseudo-labels via diverse initializations and thresholds the clean versus noisy pixels by consensus) with a Mean-Teacher style learning framework that uses an EMA-updated teacher to enforce uncertainty-guided consistency. Clean pixels drive supervised learning while noisy pixels contribute through a contrastive-like consistency loss, controlled by per-pixel uncertainty estimates and a ramped loss weight . On the RIGA dataset, MPNN achieves state-of-the-art performance for both ground-truth references (Majority Vote and Rater1), demonstrating robust denoising of noisy annotations and improved segmentation accuracy, with code released for reproducibility.

Abstract

Optic disc and cup segmentation plays a crucial role in automating the screening and diagnosis of optic glaucoma. While data-driven convolutional neural networks (CNNs) show promise in this area, the inherent ambiguity of segmenting objects and background boundaries in the task of optic disc and cup segmentation leads to noisy annotations that impact model performance. To address this, we propose an innovative label-denoising method of Multiple Pseudo-labels Noise-aware Network (MPNN) for accurate optic disc and cup segmentation. Specifically, the Multiple Pseudo-labels Generation and Guided Denoising (MPGGD) module generates pseudo-labels by multiple different initialization networks trained on true labels, and the pixel-level consensus information extracted from these pseudo-labels guides to differentiate clean pixels from noisy pixels. The training framework of the MPNN is constructed by a teacher-student architecture to learn segmentation from clean pixels and noisy pixels. Particularly, such a framework adeptly leverages (i) reliable and fundamental insight from clean pixels and (ii) the supplementary knowledge within noisy pixels via multiple perturbation-based unsupervised consistency. Compared to other label-denoising methods, comprehensive experimental results on the RIGA dataset demonstrate our method's excellent performance. The code is available at https://github.com/wwwtttjjj/MPNN
Paper Structure (15 sections, 7 equations, 3 figures, 4 tables)

This paper contains 15 sections, 7 equations, 3 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Methods
  3. Multiple Pseudo-labels Generation and Guided Denoising (MPGGD)
  4. Pixel-level Semi-supervised Pattern Learning
  5. Clean Pixels Loss
  6. Noisy Pixels Loss
  7. Final Loss
  8. Experiments
  9. Dataset and Pre-processing
  10. Implementation and Evaluation Metrics
  11. Experiments on the RIGA Dataset
  12. Parameter Ablation Study
  13. Component Ablation Study
  14. Conclusions
  15. Acknowledge

Figures (3)

  • Figure 1: Illustration of our method. (a) the architecture of MPNN; (b) the MPGGD module. The MPGGD generates multiple pseudo-labels $\{\mathbf{Y}_{p1},...,\mathbf{Y}_{pK}\}$ by multiple different initialization networks ($\{LinkNet(1),..., LinkNet(K)\}$) to fit its $DSC_{m}$ on true label $\mathbf{Y}$ to a certain threshold $\varphi$, which subsequently distinguishes the clean pixels set $\mathbf{O}_{cl}$ from the noisy pixels set $\mathbf{O}_{no}$ by consistent information. The MPNN is constructed by a model consisting of a student network and a teacher network. The student network learns by (i) minimizing the segmentation loss guided by $\mathbf{O}_{cl}$ and (ii) minimizing the consistency loss guided by $\mathbf{O}_{no}$ in the teacher model with multiple uncertainties. The teacher network is updated using the EMA.
  • Figure 2: Visualized segmentation results on the RIGA dataset of different methods. The ground truth mask and predicted mask are represented by red and green, respectively.
  • Figure 3: Visualization of the identified noisy pixels (green), the number underneath indicates the total number of noisy pixels included.