Synergy-Guided Regional Supervision of Pseudo Labels for Semi-Supervised Medical Image Segmentation

Tao Wang, Xinlin Zhang, Yuanbin Chen, Yuanbo Zhou, Longxuan Zhao, Tao Tan, Tong Tong

Abstract

Semi-supervised learning has received considerable attention for its potential to leverage abundant unlabeled data to enhance model robustness. Pseudo labeling is a widely used strategy in semi supervised learning. However, existing methods often suffer from noise contamination, which can undermine model performance. To tackle this challenge, we introduce a novel Synergy-Guided Regional Supervision of Pseudo Labels (SGRS-Net) framework. Built upon the mean teacher network, we employ a Mix Augmentation module to enhance the unlabeled data. By evaluating the synergy before and after augmentation, we strategically partition the pseudo labels into distinct regions. Additionally, we introduce a Region Loss Evaluation module to assess the loss across each delineated area. Extensive experiments conducted on the LA dataset have demonstrated superior performance over state-of-the-art techniques, underscoring the efficiency and practicality of our framework.

Introduction

Medical image segmentation is critical for modern radiology but is often hindered by the high cost of expert annotations. Semi-supervised learning (SSL) addresses this by utilizing small labeled datasets alongside larger unlabeled ones. A common SSL approach is pseudo-labeling, where a model generates its own training targets. However, this often leads to "noise contamination," where the model reinforces its own errors, particularly in ambiguous regions like anatomical boundaries.

Existing noise-reduction techniques typically rely on simple confidence thresholds. These methods often suffer from over-suppression (discarding useful data) or over-smoothing (losing fine details). SGRS-Net proposes a more granular approach by evaluating prediction "synergy" to categorize pseudo-labels into regions of varying reliability.

Methodology: The SGRS-Net Framework

SGRS-Net extends the Mean Teacher architecture with three primary modules designed to manage label noise surgically.

The SGRS-Net architecture, illustrating the Teacher-Student framework, Mix Augmentation, Synergy Evaluation, and Regional Loss Evaluation modules.

1. Mix Augmentation (MA) Module

The MA module generates hybrid images by interpolating unlabeled data with labeled samples. This process encourages the model to learn robust features and provides the necessary variance to test prediction consistency.

2. Synergy Evaluation (SE) Module

The SE module acts as the framework's decision engine. It compares predictions on original unlabeled images against their augmented counterparts. Based on this synergy, pseudo-labels are partitioned into three regions:

  • $\Delta$ (Disregarded): High-entropy regions where the model is confused; these are excluded from training.
  • $\Omega$ (Consistent): Stable, low-entropy regions that are treated as reliable "trusted" labels.
  • $\Theta$ (Inconsistent): Regions where predictions change under augmentation despite high confidence. These are identified as potentially noisy but informative.

3. Regional Loss Evaluation (RLE) Module

Instead of a global loss function, the RLE module applies tailored strategies. For the inconsistent $\Theta$ region, a noise-robust loss with label smoothing is employed to extract anatomical signals without allowing noise to dominate the gradient.

Experimental Results

The framework was evaluated on the Left Atrium (LA), Pancreas-CT, and BraTS2019 datasets. SGRS-Net consistently outperformed state-of-the-art (SOTA) methods, particularly in low-label regimes.

Quantitative comparison of SGRS-Net against SOTA methods on the LA dataset using 5% and 10% labeled data.

On the LA dataset, SGRS-Net achieved a Dice score of 89.70% with only 5% labeled data, surpassing the previous best method (CAML) by 2.36%. With 10% labels, it reached 90.76%, nearly matching the performance of fully supervised models (91.62%). Similar gains were observed in the Pancreas-CT dataset, where SGRS-Net showed a 9.55% improvement over the MCF method in the 5% label setting.

Conclusion

SGRS-Net introduces a sophisticated method for managing pseudo-label noise through regional synergy evaluation. By distinguishing between uncertainty and inconsistency, the model effectively leverages unlabeled data that simpler thresholding methods might discard. Future work will focus on optimizing hyperparameter selection for different medical modalities and reducing the computational overhead associated with multiple forward passes.

Reference: Wang, T., et al. (2024). Synergy-Guided Regional Supervision of Pseudo Labels for Semi-Supervised Medical Image Segmentation. arXiv:2411.04493v2.

Summary updated

Abstract

Semi-supervised learning has received considerable attention for its potential to leverage abundant unlabeled data to enhance model robustness. Pseudo labeling is a widely used strategy in semi supervised learning. However, existing methods often suffer from noise contamination, which can undermine model performance. To tackle this challenge, we introduce a novel Synergy-Guided Regional Supervision of Pseudo Labels (SGRS-Net) framework. Built upon the mean teacher network, we employ a Mix Augmentation module to enhance the unlabeled data. By evaluating the synergy before and after augmentation, we strategically partition the pseudo labels into distinct regions. Additionally, we introduce a Region Loss Evaluation module to assess the loss across each delineated area. Extensive experiments conducted on the LA dataset have demonstrated superior performance over state-of-the-art techniques, underscoring the efficiency and practicality of our framework.
BibTeX
@misc{wang2024synergy,
  author = {Tao Wang and Xinlin Zhang and Yuanbin Chen and Yuanbo Zhou and Longxuan Zhao and Tao Tan and Tong Tong},
  title = {{Synergy-Guided Regional Supervision of Pseudo Labels for Semi-Supervised Medical Image Segmentation}},
  year = {2024},
  eprint = {2411.04493},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2411.04493},
}
Figure