Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Yuanpeng He

TL;DR

This work tackles semi-supervised medical image segmentation under uncertainty by introducing Evidential Prototype Learning (EPL), which extends the probabilistic framework of evidential deep learning to fuse voxel-wise predictions from labeled and unlabeled data via Dempster–Shafer combination. It combines a teacher-student setup with dual uncertainty evaluation based on belief entropy and mass-based evidence, and it employs uncertainty-guided prototype learning to mask unreliable features during prototype construction. The approach yields state-of-the-art performance on Left Atrium, Pancreas-CT, and TBAD datasets across varying labeled data ratios, including strong gains with very limited annotations. EPL’s uncertainty-aware fusion and prototypes provide a robust framework for leveraging unlabeled data in medical image segmentation, with significant practical impact for data-scarce clinical settings.

Abstract

Although current semi-supervised medical segmentation methods can achieve decent performance, they are still affected by the uncertainty in unlabeled data and model predictions, and there is currently a lack of effective strategies that can explore the uncertain aspects of both simultaneously. To address the aforementioned issues, we propose Evidential Prototype Learning (EPL), which utilizes an extended probabilistic framework to effectively fuse voxel probability predictions from different sources and achieves prototype fusion utilization of labeled and unlabeled data under a generalized evidential framework, leveraging voxel-level dual uncertainty masking. The uncertainty not only enables the model to self-correct predictions but also improves the guided learning process with pseudo-labels and is able to feed back into the construction of hidden features. The method proposed in this paper has been experimented on LA, Pancreas-CT and TBAD datasets, achieving the state-of-the-art performance in three different labeled ratios, which strongly demonstrates the effectiveness of our strategy.

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

TL;DR

This work tackles semi-supervised medical image segmentation under uncertainty by introducing Evidential Prototype Learning (EPL), which extends the probabilistic framework of evidential deep learning to fuse voxel-wise predictions from labeled and unlabeled data via Dempster–Shafer combination. It combines a teacher-student setup with dual uncertainty evaluation based on belief entropy and mass-based evidence, and it employs uncertainty-guided prototype learning to mask unreliable features during prototype construction. The approach yields state-of-the-art performance on Left Atrium, Pancreas-CT, and TBAD datasets across varying labeled data ratios, including strong gains with very limited annotations. EPL’s uncertainty-aware fusion and prototypes provide a robust framework for leveraging unlabeled data in medical image segmentation, with significant practical impact for data-scarce clinical settings.

Abstract

Although current semi-supervised medical segmentation methods can achieve decent performance, they are still affected by the uncertainty in unlabeled data and model predictions, and there is currently a lack of effective strategies that can explore the uncertain aspects of both simultaneously. To address the aforementioned issues, we propose Evidential Prototype Learning (EPL), which utilizes an extended probabilistic framework to effectively fuse voxel probability predictions from different sources and achieves prototype fusion utilization of labeled and unlabeled data under a generalized evidential framework, leveraging voxel-level dual uncertainty masking. The uncertainty not only enables the model to self-correct predictions but also improves the guided learning process with pseudo-labels and is able to feed back into the construction of hidden features. The method proposed in this paper has been experimented on LA, Pancreas-CT and TBAD datasets, achieving the state-of-the-art performance in three different labeled ratios, which strongly demonstrates the effectiveness of our strategy.
Paper Structure (17 sections, 10 equations, 2 figures, 5 tables)

This paper contains 17 sections, 10 equations, 2 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Semi-supervised Medical Image Segmentation
  4. Evidential Deep Learning and Evidence Theory
  5. Method
  6. Generalized Probabilistic Framework
  7. Evidential Fusion-based Prediction Synthesis
  8. DUE: Dual Uncertainty Evaluation with Belief Entropy
  9. Generalized Evidential Deep Learning
  10. Uncertainty-based Prototype Learning
  11. Experiments
  12. Datasets
  13. Implementation Details
  14. Results on the Left Atrium Dataset and Pancreas-CT Dataset
  15. Results on the Aortic Dissection Dataset
  16. ...and 2 more sections

Figures (2)

  • Figure 1: Overview of EPL framework (using Left Atrium dataset for illustration). The labeled and unlabeled images are sent to the student model while the unlabeled images are sent to the teacher model. The output of the student model on the labeled and unlabeled images are supposed to be optimized by utilizing the uncertainty from model prediction and pseudo-labels. The evidential predictions of the teacher model are fused by using the Dempster's combination rule. Besides, the prototype fusion process utilizes masked attention pooling with the generated uncertainty-based reliability map.
  • Figure 2: Comparisons of visualized results on LA, Pancreas-CT and Aortic Dissection datasets. ①, ② and ③ represent the ground truth, BCP (LA, Pancreas-CT dataset) and UPCoL (TBAD dataset), the proposed method respectively.