Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Bidirectional Channel-selective Semantic Interaction for Semi-Supervised Medical Segmentation

Kaiwen Huang, Yizhe Zhang, Yi Zhou, Tianyang Xu, Tao Zhou

TL;DR

The paper tackles semi-supervised medical image segmentation under limited annotations by introducing BCSI, a single-model framework that enables bidirectional cross-stream interaction between labeled and unlabeled data. It combines Semantic-Spatial Perturbation with a learnable Channel-selective Router and Bidirectional Channel-wise Interaction to selectively exchange information on informative channels, guided by a weak-to-strong consistency regime. Empirical results across Left Atrium, BraTS-2019, and Pancreas-CT demonstrate that BCSI outperforms existing methods on Dice, IoU, and boundary metrics, especially at low label ratios. This approach reduces noise and model complexity while delivering robust segmentation performance in 3D medical imaging settings.

Abstract

Semi-supervised medical image segmentation is an effective method for addressing scenarios with limited labeled data. Existing methods mainly rely on frameworks such as mean teacher and dual-stream consistency learning. These approaches often face issues like error accumulation and model structural complexity, while also neglecting the interaction between labeled and unlabeled data streams. To overcome these challenges, we propose a Bidirectional Channel-selective Semantic Interaction~(BCSI) framework for semi-supervised medical image segmentation. First, we propose a Semantic-Spatial Perturbation~(SSP) mechanism, which disturbs the data using two strong augmentation operations and leverages unsupervised learning with pseudo-labels from weak augmentations. Additionally, we employ consistency on the predictions from the two strong augmentations to further improve model stability and robustness. Second, to reduce noise during the interaction between labeled and unlabeled data, we propose a Channel-selective Router~(CR) component, which dynamically selects the most relevant channels for information exchange. This mechanism ensures that only highly relevant features are activated, minimizing unnecessary interference. Finally, the Bidirectional Channel-wise Interaction~(BCI) strategy is employed to supplement additional semantic information and enhance the representation of important channels. Experimental results on multiple benchmarking 3D medical datasets demonstrate that the proposed method outperforms existing semi-supervised approaches.

Bidirectional Channel-selective Semantic Interaction for Semi-Supervised Medical Segmentation

TL;DR

The paper tackles semi-supervised medical image segmentation under limited annotations by introducing BCSI, a single-model framework that enables bidirectional cross-stream interaction between labeled and unlabeled data. It combines Semantic-Spatial Perturbation with a learnable Channel-selective Router and Bidirectional Channel-wise Interaction to selectively exchange information on informative channels, guided by a weak-to-strong consistency regime. Empirical results across Left Atrium, BraTS-2019, and Pancreas-CT demonstrate that BCSI outperforms existing methods on Dice, IoU, and boundary metrics, especially at low label ratios. This approach reduces noise and model complexity while delivering robust segmentation performance in 3D medical imaging settings.

Abstract

Semi-supervised medical image segmentation is an effective method for addressing scenarios with limited labeled data. Existing methods mainly rely on frameworks such as mean teacher and dual-stream consistency learning. These approaches often face issues like error accumulation and model structural complexity, while also neglecting the interaction between labeled and unlabeled data streams. To overcome these challenges, we propose a Bidirectional Channel-selective Semantic Interaction~(BCSI) framework for semi-supervised medical image segmentation. First, we propose a Semantic-Spatial Perturbation~(SSP) mechanism, which disturbs the data using two strong augmentation operations and leverages unsupervised learning with pseudo-labels from weak augmentations. Additionally, we employ consistency on the predictions from the two strong augmentations to further improve model stability and robustness. Second, to reduce noise during the interaction between labeled and unlabeled data, we propose a Channel-selective Router~(CR) component, which dynamically selects the most relevant channels for information exchange. This mechanism ensures that only highly relevant features are activated, minimizing unnecessary interference. Finally, the Bidirectional Channel-wise Interaction~(BCI) strategy is employed to supplement additional semantic information and enhance the representation of important channels. Experimental results on multiple benchmarking 3D medical datasets demonstrate that the proposed method outperforms existing semi-supervised approaches.
Paper Structure (14 sections, 14 equations, 4 figures, 7 tables)

This paper contains 14 sections, 14 equations, 4 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Semi-Supervised Medical Image Segmentation
  4. Weak-to-Strong Consistency Learning
  5. Methodology
  6. Semantic-Spatial Perturbation
  7. Channel-selective Router
  8. Bidirectional Channel-wise Interaction
  9. Loss Function
  10. Experiments
  11. Experimental Setup
  12. Comparison With State-of-the-art Methods
  13. Ablation Studies
  14. Conclusion

Figures (4)

  • Figure 1: A comparison of semi-supervised learning paradigms: (a) Mean Teacher framework with student and teacher networks, (b) Co-training framework with two subnets, and (c) Our proposed framework with bidirectional data-stream interaction and semantic-spatial perturbation under weak-to-strong consistency.
  • Figure 2: An overview of the proposed BCSI framework, involving an encoder-decoder structure. A channel-selection router then processes the extracted features from labeled and unlabeled data, followed by bidirectional channel-wise interaction for feature enhancement. Finally, the refined features are passed to the decoder under a weak-to-strong consistency learning paradigm.
  • Figure 3: Visualization results of our method compared to other semi-supervised methods on the LA, BraTS-2019, and Pancreas datasets. The yellow numbers represent the Dice score of the currently displayed sample.
  • Figure 4: Performance comparison of our method with foundation model-based semi-supervised approaches and the fully supervised VNet, using a $20\%$ labeled data ratio.