EPL: Evidential Prototype Learning for Semi-supervised Medical Image Segmentation

Zhi Jin, Fellow of Ieee and Ccf, Yuanpeng HearXiv 2024 · DOI · arXiv

TL;DR

This paper introduces Evidential Prototype Learning (EPL), a novel semi-supervised medical image segmentation method that tackles the challenge of uncertainty in both unlabeled data and model predictions. EPL fuses predictions from multiple sources using Dempster's combination rule and employs a dual uncertainty measurement mechanism to improve the learning process, achieving state-of-the-art results on several medical imaging datasets.

What Problem Does This Paper Address?

Medical image segmentation (MIS) is crucial for computer-aided diagnosis, but it's hampered by the scarcity of precisely annotated medical images, which are expensive and time-consuming to create. While semi-supervised learning (SSL) methods aim to leverage abundant unlabeled data, they often struggle with the inherent uncertainty present in both the data itself and the model's predictions. Existing SSL approaches, such as pseudo-labeling and consistency regularization, have limitations in effectively handling this dual uncertainty. Specifically, pseudo-labeling can be negatively impacted by low-quality pseudo-labels, and consistency methods often rely on assumptions about the latent space that may not hold true for complex medical data. This paper addresses the need for a more robust approach that can simultaneously explore and leverage the uncertain aspects of both unlabeled data and model predictions.

Methods

The core of EPL is its Evidential Prototype Learning framework. Here's a breakdown of the key techniques:

  • Evidential Deep Learning (EDL) Extension: EPL extends EDL by incorporating multi-objective sets, allowing for more nuanced modeling of probability distributions. This framework is used to gather "evidence" for each class, which is then used to construct a Dirichlet distribution of class probabilities.
  • Dempster's Combination Rule: To fuse predictions from multiple evidential classifiers (both from the student and teacher models), EPL employs Dempster's combination rule. This rule effectively combines mass assignments from different sources, strengthening the overall prediction and accounting for uncertainty.
  • Dual Uncertainty Measurement: A novel mechanism is introduced to measure voxel-level uncertainty. This combines the existing uncertainty measure derived from the evidential framework with belief entropy. This dual approach allows the model to self-correct its predictions on labeled data and guides learning on unlabeled data using the uncertainty of pseudo-labels generated by a teacher network.
  • Uncertainty-Based Prototype Learning: Prototypes are extracted from both labeled and unlabeled data. Crucially, the uncertainty generated by the evidential framework is used to mask unreliable features during prototype generation. This ensures that prototypes are more robust and representative.
  • Generalized Evidential Deep Learning Optimization: The optimization objective is extended to better handle uncertainty. It aims to align predictions with labels while avoiding forced optimization on highly uncertain regions, which could introduce biases.

Datasets

The proposed method was evaluated on three distinct medical imaging datasets:

  • Left Atrium (LA) dataset: Consists of 100 MR images.
  • Pancreas-CT dataset: Comprises 82 CTA scans.
  • Type B Aortic Dissection (TBAD) dataset: A multi-center dataset with 124 CTA scans, meticulously labeled by radiologists.

For all datasets, the experiments utilized a limited amount of labeled training data (5%, 10%, or 20%), normalizing voxel intensities to have zero mean and unit variance.

Key Insights and Ideas

  • Uncertainty is a Feature, Not Just Noise: The paper's central theme is that uncertainty in medical image segmentation is not merely an obstacle but a valuable signal. EPL actively leverages this uncertainty to improve model performance.
  • Fusion of Evidence for Robustness: By combining predictions from multiple sources using Dempster's rule, EPL gains robustness against individual model weaknesses and enhances the reliability of its fused predictions.
  • Dual Uncertainty for Smarter Learning: The dual uncertainty measurement (evidential + belief entropy) provides a more comprehensive understanding of prediction ambiguity, enabling more targeted and effective learning from both labeled and unlabeled data.
  • Masking for Better Prototypes: Using uncertainty to mask features before prototype generation is a clever way to prevent noisy or unreliable features from distorting the learned class representations.

Key Results

The EPL framework achieved state-of-the-art performance across multiple metrics on the LA, Pancreas-CT, and TBAD datasets, even with very limited labeled data.

  • On the TBAD dataset, EPL significantly outperformed existing methods, achieving superior performance with only 5% of labeled data compared to other methods using 20%.
  • For the LA and Pancreas-CT datasets, EPL achieved performance very close to fully supervised V-Net with only 20% labeled data, and in some cases, even reversed the ASD metric.
  • Visualizations showed that EPL's segmentation results were more coherent and consistent with the ground truth compared to comparative methods, which often produced fragmented segmentations.

Limitations and Future Work

  • Generalization of EDL: The paper notes that generalized evidential deep learning might not be as friendly to boundary overlap metrics when the proportion of labeled data is high. This suggests a potential area for further refinement.
  • Computational Cost: While not explicitly detailed as a limitation, the use of multiple classifiers and evidential fusion likely increases computational complexity. Future work could explore more efficient implementations.
  • Further Exploration of Uncertainty: The paper demonstrates the effectiveness of uncertainty-guided learning. Future work could explore more sophisticated ways to integrate and utilize uncertainty across different stages of the segmentation pipeline.

Summary updated

BibTeX
@misc{ccf2024epl,
  author = {Zhi Jin, Fellow of IEEE and CCF and Yuanpeng He},
  title = {{EPL: Evidential Prototype Learning for Semi-supervised Medical Image Segmentation}},
  year = {2024},
  eprint = {2404.06181},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2404.06181},
}
Figure