Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Diverse Teaching and Label Propagation for Generic Semi-Supervised Medical Image Segmentation

Wei Li, Pengcheng Zhou, Linye Ma, Wenyi Zhao, Huihua Yang, Yuchen Guo

TL;DR

The paper addresses annotation scarcity and domain shift in medical image segmentation by introducing DTLP-Net, a unified framework that simultaneously tackles SSMIS, UMDA, and Semi-MDG. It leverages a dual-teacher scheme with a diffusion-based and a mean-teacher branch, combined with entropy-based ensembling to produce reliable pseudo-labels under domain shift. The model further enforces global-local consistency through cross-set CutMix and masked image modeling, employs knowledge distillation and masked reconstruction to mitigate noisy pseudo-labels, and introduces voxel-level label propagation to capture voxelwise correlations. Extensive experiments across five benchmarks demonstrate state-of-the-art performance and strong robustness across tasks, with thorough ablations confirming the contribution of each component. The approach advances practical SSL in medical imaging by providing a versatile, data-structure-aware framework that remains effective under varied domain shifts and label availability scenarios.

Abstract

Both limited annotation and domain shift are significant challenges frequently encountered in medical image segmentation, leading to derivative scenarios like semi-supervised medical (SSMIS), semi-supervised medical domain generalization (Semi-MDG) and unsupervised medical domain adaptation (UMDA). Conventional methods are generally tailored to specific tasks in isolation, the error accumulation hinders the effective utilization of unlabeled data and limits further improvements, resulting in suboptimal performance when these issues occur. In this paper, we aim to develop a generic framework that masters all three tasks. We found that the key to solving the problem lies in how to generate reliable pseudo labels for the unlabeled data in the presence of domain shift with labeled data and increasing the diversity of the model. To tackle this issue, we employ a Diverse Teaching and Label Propagation Network (DTLP-Net) to boosting the Generic Semi-Supervised Medical Image Segmentation. Our DTLP-Net involves a single student model and two diverse teacher models, which can generate reliable pseudo-labels for the student model. The first teacher model decouple the training process with labeled and unlabeled data, The second teacher is momentum-updated periodically, thus generating reliable yet divers pseudo-labels. To fully utilize the information within the data, we adopt inter-sample and intra-sample data augmentation to learn the global and local knowledge. In addition, to further capture the voxel-level correlations, we propose label propagation to enhance the model robust. We evaluate our proposed framework on five benchmark datasets for SSMIS, UMDA, and Semi-MDG tasks. The results showcase notable improvements compared to state-of-the-art methods across all five settings, indicating the potential of our framework to tackle more challenging SSL scenarios.

Diverse Teaching and Label Propagation for Generic Semi-Supervised Medical Image Segmentation

TL;DR

The paper addresses annotation scarcity and domain shift in medical image segmentation by introducing DTLP-Net, a unified framework that simultaneously tackles SSMIS, UMDA, and Semi-MDG. It leverages a dual-teacher scheme with a diffusion-based and a mean-teacher branch, combined with entropy-based ensembling to produce reliable pseudo-labels under domain shift. The model further enforces global-local consistency through cross-set CutMix and masked image modeling, employs knowledge distillation and masked reconstruction to mitigate noisy pseudo-labels, and introduces voxel-level label propagation to capture voxelwise correlations. Extensive experiments across five benchmarks demonstrate state-of-the-art performance and strong robustness across tasks, with thorough ablations confirming the contribution of each component. The approach advances practical SSL in medical imaging by providing a versatile, data-structure-aware framework that remains effective under varied domain shifts and label availability scenarios.

Abstract

Both limited annotation and domain shift are significant challenges frequently encountered in medical image segmentation, leading to derivative scenarios like semi-supervised medical (SSMIS), semi-supervised medical domain generalization (Semi-MDG) and unsupervised medical domain adaptation (UMDA). Conventional methods are generally tailored to specific tasks in isolation, the error accumulation hinders the effective utilization of unlabeled data and limits further improvements, resulting in suboptimal performance when these issues occur. In this paper, we aim to develop a generic framework that masters all three tasks. We found that the key to solving the problem lies in how to generate reliable pseudo labels for the unlabeled data in the presence of domain shift with labeled data and increasing the diversity of the model. To tackle this issue, we employ a Diverse Teaching and Label Propagation Network (DTLP-Net) to boosting the Generic Semi-Supervised Medical Image Segmentation. Our DTLP-Net involves a single student model and two diverse teacher models, which can generate reliable pseudo-labels for the student model. The first teacher model decouple the training process with labeled and unlabeled data, The second teacher is momentum-updated periodically, thus generating reliable yet divers pseudo-labels. To fully utilize the information within the data, we adopt inter-sample and intra-sample data augmentation to learn the global and local knowledge. In addition, to further capture the voxel-level correlations, we propose label propagation to enhance the model robust. We evaluate our proposed framework on five benchmark datasets for SSMIS, UMDA, and Semi-MDG tasks. The results showcase notable improvements compared to state-of-the-art methods across all five settings, indicating the potential of our framework to tackle more challenging SSL scenarios.

Paper Structure

This paper contains 39 sections, 18 equations, 4 figures, 11 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Semi-supervised medical image segmentation
  4. Unsupervised Medical Domain Adaptation
  5. Semi-supervised Medical Domain Generalization
  6. Methodology
  7. Preliminary baseline
  8. Diffusion for Capturing Invariant Features with $D(x^l;\xi)$
  9. Supervised Training with $D(x^l;\psi)$
  10. Pseudo Labeling with Diverse Teaching
  11. Pseudo Labeling with Reparameterize & Smooth (RS) Strategy
  12. Pseudo Labeling with Mean Teacher model
  13. Unsupervised Training with $\mathcal{D}(x^u;\theta)$
  14. Global-local consistency learning
  15. Global consistency learning with Cross-set CutMix
  16. ...and 24 more sections

Figures (4)

  • Figure 1: An illustration of our Diverse Teaching and Label Propagation Network (DTLP-Net). The training process of the decoders using labeled data and unlabeled data is decoupled. During inference, only the decoder $\mathcal{E}(x^u;\theta)$ is employed.
  • Figure 2: Visualization results with 5% labeled data on the LA dataset.
  • Figure 3: Visualization results with 20% labeled on the Synapse dataset.
  • Figure 4: Parameters sensitivity analysis on Synapse data with 20% labeled data for training.