Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Adaptive Knowledge Transferring with Switching Dual-Student Framework for Semi-Supervised Medical Image Segmentation

Thanh-Huy Nguyen, Hoang-Thien Nguyen, Ba-Thinh Lam, Vi Vu, Bach X. Nguyen, Jianhua Xing, Tianyang Wang, Xingjian Li, Min Xu

TL;DR

This work tackles limitations of Mean-Teacher in semi-supervised medical image segmentation by introducing a switching Dual-Student framework with entropy-guided Student Selection and a Loss-Aware Exponential Moving Average (LA-EMA). The method uses Dual-Student Cross-Sample CutMix to diversify pseudo-labels, selects the most reliable student for teacher updates, and adaptively weights EMA updates based on student loss, all underpinned by PAC-Bayes-based reasoning. Empirical results on 3D medical datasets (LA and ACDC) and general datasets (CIFAR-10, Pascal VOC) show state-of-the-art accuracy and robust performance under limited supervision, with ablations confirming the effectiveness of CSC, SS, and LA-EMA. The approach offers a practical, plug-and-play improvement to semi-supervised segmentation pipelines with substantial impact for clinical imaging tasks where annotations are scarce.

Abstract

Teacher-student frameworks have emerged as a leading approach in semi-supervised medical image segmentation, demonstrating strong performance across various tasks. However, the learning effects are still limited by the strong correlation and unreliable knowledge transfer process between teacher and student networks. To overcome this limitation, we introduce a novel switching Dual-Student architecture that strategically selects the most reliable student at each iteration to enhance dual-student collaboration and prevent error reinforcement. We also introduce a strategy of Loss-Aware Exponential Moving Average to dynamically ensure that the teacher absorbs meaningful information from students, improving the quality of pseudo-labels. Our plug-and-play framework is extensively evaluated on 3D medical image segmentation datasets, where it outperforms state-of-the-art semi-supervised methods, demonstrating its effectiveness in improving segmentation accuracy under limited supervision.

Adaptive Knowledge Transferring with Switching Dual-Student Framework for Semi-Supervised Medical Image Segmentation

TL;DR

This work tackles limitations of Mean-Teacher in semi-supervised medical image segmentation by introducing a switching Dual-Student framework with entropy-guided Student Selection and a Loss-Aware Exponential Moving Average (LA-EMA). The method uses Dual-Student Cross-Sample CutMix to diversify pseudo-labels, selects the most reliable student for teacher updates, and adaptively weights EMA updates based on student loss, all underpinned by PAC-Bayes-based reasoning. Empirical results on 3D medical datasets (LA and ACDC) and general datasets (CIFAR-10, Pascal VOC) show state-of-the-art accuracy and robust performance under limited supervision, with ablations confirming the effectiveness of CSC, SS, and LA-EMA. The approach offers a practical, plug-and-play improvement to semi-supervised segmentation pipelines with substantial impact for clinical imaging tasks where annotations are scarce.

Abstract

Teacher-student frameworks have emerged as a leading approach in semi-supervised medical image segmentation, demonstrating strong performance across various tasks. However, the learning effects are still limited by the strong correlation and unreliable knowledge transfer process between teacher and student networks. To overcome this limitation, we introduce a novel switching Dual-Student architecture that strategically selects the most reliable student at each iteration to enhance dual-student collaboration and prevent error reinforcement. We also introduce a strategy of Loss-Aware Exponential Moving Average to dynamically ensure that the teacher absorbs meaningful information from students, improving the quality of pseudo-labels. Our plug-and-play framework is extensively evaluated on 3D medical image segmentation datasets, where it outperforms state-of-the-art semi-supervised methods, demonstrating its effectiveness in improving segmentation accuracy under limited supervision.

Paper Structure

This paper contains 18 sections, 26 equations, 4 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Preliminary
  5. Dual-Student Cross-Sample CutMix
  6. Student Selection
  7. Loss-Aware Exponential Moving Average
  8. Theoretical Analysis
  9. Proposition 1: Variance Suppression through Loss-Aware EMA
  10. Proposition 2: Reliable Descent via Entropy-Guided Student Selection
  11. Proposition 3: Generalization Improvement under PAC-Bayes Bound
  12. Summary
  13. Experiments
  14. Experimental Settings
  15. Comparison with state-of-the-art methods
  16. ...and 3 more sections

Figures (4)

  • Figure 1: The pipeline of our proposed framework consists of two parts. The training part involves the learning process of the dual-student model, where both the supervised and unsupervised losses are computed and optimized based on the predictions from the students and the pseudo-labels generated by the teacher through Dual-Student Cross-Sample CutMix module. The knowledge transfer part includes the Student Selection process, which identifies the best-performing student at each iteration and calculates its Loss-Aware Exponential Moving Average (LA-EMA) to update the teacher model.
  • Figure 2: Qualitative results of several samples from the ACDC test set.
  • Figure 3: Qualitative results on sample cases from the LA test set. From left to right, the columns display the Image, Ground Truth (GT), and predictions from various state-of-the-art methods, including BCP, AD-MT, and Ours.
  • Figure 4: The validation score (DICE) of teacher model of BCP and Ours over training.