Alternate Diverse Teaching for Semi-supervised Medical Image Segmentation

Zhen Zhao, Zicheng Wang, Longyue Wang, Dian Yu, Yixuan Yuan, Luping Zhou

TL;DR

This paper introduces AD-MT, a novel semi-supervised learning framework for medical image segmentation that tackles the "confirmation bias" problem in traditional teacher-student models. By employing two alternating teacher models with diverse training strategies and a unique conflict-combating mechanism, AD-MT significantly improves segmentation accuracy, especially when labeled data is scarce.

What Problem Does This Paper Address?

Medical image segmentation is crucial for computer-aided diagnosis, but it requires a vast amount of meticulously annotated data, which is expensive and time-consuming to obtain. Semi-supervised learning (SSL) offers a solution by leveraging abundant unlabeled data alongside limited labeled data. However, existing SSL methods, particularly those based on the teacher-student paradigm, often suffer from "confirmation bias." This occurs when the model, relying on its own predictions (even if slightly flawed), reinforces these errors, leading to suboptimal performance. The paper aims to mitigate this bias by introducing more diverse and robust supervision signals from multiple teacher models.

Methods

The core of AD-MT is a teacher-student framework with two key novel modules:

  • Random Periodic Alternate (RPA) Updating Module: Instead of a single, continuously updated teacher, AD-MT uses two non-trainable teacher models that are updated alternately. This alternation is achieved through:
  • Complementary unlabeled data batches: Each teacher is trained on different subsets of unlabeled data.
  • Distinct data augmentation strategies: Different augmentation techniques are applied to the data fed to each teacher, forcing them to learn from varied perspectives.
  • Randomized switching periods: The frequency at which teachers switch is randomized, further enhancing diversity and preventing over-reliance on any single teacher's perspective.
  • Conflict-Combating Module (CCM): Traditional methods often discard conflicting predictions between teachers. AD-MT, however, embraces these conflicts. It uses an entropy-based ensembling strategy to:
  • Separate consistent and conflicting predictions: It identifies where the teachers agree and disagree.
  • Leverage disagreements: It encourages the student model to learn from the differences between teacher predictions, especially when the student's own prediction is less confident. This is achieved by using the lower-entropy prediction (from either the teachers or the student) as the final supervision for conflicting cases.

The student model is trained using both labeled data (with a standard supervised loss) and unlabeled data (with a consistency loss guided by the teachers' diverse and conflict-aware predictions).

Datasets

The proposed AD-MT method was evaluated on the following widely-used medical image segmentation benchmarks:

  • Left Atrium (LA) dataset: A 3D dataset consisting of 100 3D gadolinium-enhanced magnetic resonance image scans.
  • Pancreas-NIH dataset: Another 3D dataset comprising 82 contrast-enhanced abdominal CT volumes.
  • Automated Cardiac Diagnosis Challenge (ACDC) dataset: A 2D benchmark with 100 cardiac MR imaging samples.

The paper does not introduce new datasets.

Key Insights and Ideas

  • Embracing conflicts is beneficial: The paper challenges the common practice of discarding conflicting predictions from multiple models. Instead, it demonstrates that learning from these disagreements can significantly improve segmentation performance, especially in challenging semi-supervised settings.
  • Diversity in teacher models is key: By using complementary data, distinct augmentations, and randomized updates, AD-MT ensures that its teacher models develop diverse "perspectives," which in turn provide richer and more robust supervision signals to the student.
  • Randomization enhances robustness: The randomized switching periods and data augmentation strategies prevent the model from becoming overly reliant on specific patterns or teacher behaviors, leading to more generalizable performance.
  • Efficiency without extra parameters: AD-MT achieves its performance gains by cleverly orchestrating the training process of existing model architectures, rather than requiring significantly more complex or numerous models.

Key Results

AD-MT consistently outperforms state-of-the-art (SOTA) semi-supervised medical image segmentation methods across all tested datasets and settings.

  • Left Atrium (LA) dataset (3D): Achieved the highest Dice Score and Jaccard Score on both 5% and 10% labeled data settings. For instance, with only 4 labeled instances, AD-MT achieved a Dice Score of 89.63%, outperforming the previous SOTA method BCP by 1.61% in Dice Score and 2.56% in Jaccard Score, without requiring additional pre-training.
  • ACDC dataset (2D): Achieved the best performance in both 5% and 10% labeled data settings. In the 5% setting, AD-MT achieved a Dice Score of 88.75%, a 1.16% improvement over BCP, and a significant ~20% improvement over other end-to-end methods without pre-training.
  • Pancreas-NIH dataset (3D): Demonstrated strong recognition performance, especially with limited labeled data. With 10% labeled data, AD-MT achieved a Dice Score of 80.21%, a 6.38% improvement over BCP. It also achieved the lowest 95% Hausdorff Distance (95HD) and Average Surface Distance (ASD), indicating more precise segmentation.

Qualitative results also show AD-MT producing segmentation results closer to the ground truth compared to other SOTA methods.

Limitations and Future Work

  • Threshold Sensitivity: The paper notes that the high-confidence threshold ($\tau$) used in the CCM module can impact performance, with optimal values varying between 2D and 3D datasets. Future work could explore adaptive thresholding mechanisms.
  • Ensembling Strategies: While the entropy-based CCM is effective, the paper mentions examining "more ensembling strategies in the experiment section," suggesting that further exploration in this area could yield additional improvements.
  • Generalizability: While tested on multiple medical imaging datasets, further validation on a wider range of modalities and anatomical structures would be beneficial.

The authors suggest that their proposed AD-MT framework, with its focus on diverse teaching and conflict utilization, provides a robust and efficient approach for semi-supervised medical image segmentation, particularly in data-scarce scenarios.

Summary updated

Abstract

Semi-supervised medical image segmentation has shown promise in training models with limited labeled data. However, current dominant teacher-student based approaches can suffer from the confirmation bias. To address this challenge, we propose AD-MT, an alternate diverse teaching approach in a teacher-student framework. It involves a single student model and two non-trainable teacher models that are momentum-updated periodically and randomly in an alternate fashion. To mitigate the confirmation bias via the diverse supervision, the core of AD-MT lies in two proposed modules: the Random Periodic Alternate (RPA) Updating Module and the Conflict-Combating Module (CCM). The RPA schedules an alternating diverse updating process with complementary unlabeled data batches, distinct data augmentation, and random switching periods to encourage diverse reasoning from different teaching perspectives. The CCM employs an entropy-based ensembling strategy to encourage the model to learn from both the consistent and conflicting predictions between the teachers. Experimental results demonstrate the effectiveness and superiority of AD-MT on the 2D and 3D medical segmentation benchmarks across various semi-supervised settings.
BibTeX
@inproceedings{zhao2024alternate,
  author = {Zhen Zhao and Zicheng Wang and Longyue Wang and Dian Yu and Yixuan Yuan and Luping Zhou},
  title = {{Alternate Diverse Teaching for Semi-supervised Medical Image Segmentation}},
  booktitle = {European Conference on Computer Vision},
  year = {2024},
  doi = {10.1007/978-3-031-72652-1\_14},
}
Figure