Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Semi-Supervised Medical Image Segmentation via Dual Networks

Yunyao Lu, Yihang Wu, Reem Kateb, Ahmad Chaddad

TL;DR

The paper tackles the challenge of semi-supervised 3D medical image segmentation with limited labeled data and noisy pseudo-labels. It proposes a dual-stream teacher–student architecture that employs consistency regularization, uncertainty-guided pseudo-label weighting, and a voxel-level contrastive loss to align uncertain predictions with reliable class prototypes. Key components include a cosine-based regularization term $\mathcal{L}_{reg}$, entropy/KL-based uncertainty weighting, and the contrastive loss $\mathcal{L}_c$, all evaluated on LA MRI and Pancreas CT benchmarks, where it achieves state-of-the-art performance relative to SSL baselines. The approach reduces labeling burden while improving robustness and generalization of 3D segmentation across modalities, with code available at GitHub.

Abstract

Traditional supervised medical image segmentation models require large amounts of labeled data for training; however, obtaining such large-scale labeled datasets in the real world is extremely challenging. Recent semi-supervised segmentation models also suffer from noisy pseudo-label issue and limited supervision in feature space. To solve these challenges, we propose an innovative semi-supervised 3D medical image segmentation method to reduce the dependency on large, expert-labeled datasets. Furthermore, we introduce a dual-network architecture to address the limitations of existing methods in using contextual information and generating reliable pseudo-labels. In addition, a self-supervised contrastive learning strategy is used to enhance the representation of the network and reduce prediction uncertainty by distinguishing between reliable and unreliable predictions. Experiments on clinical magnetic resonance imaging demonstrate that our approach outperforms state-of-the-art techniques. Our code is available at https://github.com/AIPMLab/Semi-supervised-Segmentation.

Semi-Supervised Medical Image Segmentation via Dual Networks

TL;DR

The paper tackles the challenge of semi-supervised 3D medical image segmentation with limited labeled data and noisy pseudo-labels. It proposes a dual-stream teacher–student architecture that employs consistency regularization, uncertainty-guided pseudo-label weighting, and a voxel-level contrastive loss to align uncertain predictions with reliable class prototypes. Key components include a cosine-based regularization term , entropy/KL-based uncertainty weighting, and the contrastive loss , all evaluated on LA MRI and Pancreas CT benchmarks, where it achieves state-of-the-art performance relative to SSL baselines. The approach reduces labeling burden while improving robustness and generalization of 3D segmentation across modalities, with code available at GitHub.

Abstract

Traditional supervised medical image segmentation models require large amounts of labeled data for training; however, obtaining such large-scale labeled datasets in the real world is extremely challenging. Recent semi-supervised segmentation models also suffer from noisy pseudo-label issue and limited supervision in feature space. To solve these challenges, we propose an innovative semi-supervised 3D medical image segmentation method to reduce the dependency on large, expert-labeled datasets. Furthermore, we introduce a dual-network architecture to address the limitations of existing methods in using contextual information and generating reliable pseudo-labels. In addition, a self-supervised contrastive learning strategy is used to enhance the representation of the network and reduce prediction uncertainty by distinguishing between reliable and unreliable predictions. Experiments on clinical magnetic resonance imaging demonstrate that our approach outperforms state-of-the-art techniques. Our code is available at https://github.com/AIPMLab/Semi-supervised-Segmentation.

Paper Structure

This paper contains 10 sections, 12 equations, 2 figures, 3 tables.

Table of Contents

  1. Introduction
  2. PROPOSED METHOD
  3. Consistency Regularization
  4. Uncertainty Estimation
  5. Contrastive Loss for Uncertain Voxels
  6. Experiments
  7. Datasets
  8. CONCLUSION
  9. Compliance with ethical standards
  10. Acknowledgements

Figures (2)

  • Figure 1: Pipeline of our proposed model. It has three parts: consistency regularization, uncertainty estimation, and a contrastive learning method to improve student and teacher performance.
  • Figure 2: Visual comparison of segmentation results: the first and second rows show the left atrium (LA), while the third and fourth rows show the pancreas, respectively.