Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

The 1st Solution for CARE Liver Task Challenge 2025: Contrast-Aware Semi-Supervised Segmentation with Domain Generalization and Test-Time Adaptation

Jincan Lou, Jingkun Chen, Haoquan Li, Hang Li, Wenjian Huang, Weihua Chen, Fan Wang, Jianguo Zhang

TL;DR

This work tackles accurate liver segmentation in contrast-enhanced MRI under annotation scarcity and cross-site domain shifts. It introduces CoSSeg-TTA, a semi-supervised framework built on nnU-Netv2 that integrates a mean-teacher paradigm, histogram-based domain adaptation, a contrast-aware module mapping $X_{T1}$ to $\hat{X}_{GED4}$, and continual test-time adaptation with CoTTA, plus post-processing. By leveraging both labeled and unlabeled data and generating pseudo-labels for further fine-tuning, the approach achieves superior Dice and Hausdorff distances compared to the baseline, with strong generalization to unseen domains. The contributions offer a practical, robust pipeline for cross-domain liver segmentation in clinical MRI, though limitations include the need for paired modalities and substantial computation for real-time deployment.

Abstract

Accurate liver segmentation from contrast-enhanced MRI is essential for diagnosis, treatment planning, and disease monitoring. However, it remains challenging due to limited annotated data, heterogeneous enhancement protocols, and significant domain shifts across scanners and institutions. Traditional image-to-image translation frameworks have made great progress in domain generalization, but their application is not straightforward. For example, Pix2Pix requires image registration, and cycle-GAN cannot be integrated seamlessly into segmentation pipelines. Meanwhile, these methods are originally used to deal with cross-modality scenarios, and often introduce structural distortions and suffer from unstable training, which may pose drawbacks in our single-modality scenario. To address these challenges, we propose CoSSeg-TTA, a compact segmentation framework for the GED4 (Gd-EOB-DTPA enhanced hepatobiliary phase MRI) modality built upon nnU-Netv2 and enhanced with a semi-supervised mean teacher scheme to exploit large amounts of unlabeled volumes. A domain adaptation module, incorporating a randomized histogram-based style appearance transfer function and a trainable contrast-aware network, enriches domain diversity and mitigates cross-center variability. Furthermore, a continual test-time adaptation strategy is employed to improve robustness during inference. Extensive experiments demonstrate that our framework consistently outperforms the nnU-Netv2 baseline, achieving superior Dice score and Hausdorff Distance while exhibiting strong generalization to unseen domains under low-annotation conditions.

The 1st Solution for CARE Liver Task Challenge 2025: Contrast-Aware Semi-Supervised Segmentation with Domain Generalization and Test-Time Adaptation

TL;DR

This work tackles accurate liver segmentation in contrast-enhanced MRI under annotation scarcity and cross-site domain shifts. It introduces CoSSeg-TTA, a semi-supervised framework built on nnU-Netv2 that integrates a mean-teacher paradigm, histogram-based domain adaptation, a contrast-aware module mapping to , and continual test-time adaptation with CoTTA, plus post-processing. By leveraging both labeled and unlabeled data and generating pseudo-labels for further fine-tuning, the approach achieves superior Dice and Hausdorff distances compared to the baseline, with strong generalization to unseen domains. The contributions offer a practical, robust pipeline for cross-domain liver segmentation in clinical MRI, though limitations include the need for paired modalities and substantial computation for real-time deployment.

Abstract

Accurate liver segmentation from contrast-enhanced MRI is essential for diagnosis, treatment planning, and disease monitoring. However, it remains challenging due to limited annotated data, heterogeneous enhancement protocols, and significant domain shifts across scanners and institutions. Traditional image-to-image translation frameworks have made great progress in domain generalization, but their application is not straightforward. For example, Pix2Pix requires image registration, and cycle-GAN cannot be integrated seamlessly into segmentation pipelines. Meanwhile, these methods are originally used to deal with cross-modality scenarios, and often introduce structural distortions and suffer from unstable training, which may pose drawbacks in our single-modality scenario. To address these challenges, we propose CoSSeg-TTA, a compact segmentation framework for the GED4 (Gd-EOB-DTPA enhanced hepatobiliary phase MRI) modality built upon nnU-Netv2 and enhanced with a semi-supervised mean teacher scheme to exploit large amounts of unlabeled volumes. A domain adaptation module, incorporating a randomized histogram-based style appearance transfer function and a trainable contrast-aware network, enriches domain diversity and mitigates cross-center variability. Furthermore, a continual test-time adaptation strategy is employed to improve robustness during inference. Extensive experiments demonstrate that our framework consistently outperforms the nnU-Netv2 baseline, achieving superior Dice score and Hausdorff Distance while exhibiting strong generalization to unseen domains under low-annotation conditions.

Paper Structure

This paper contains 15 sections, 5 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Semi-Supervised Learning with mean teacher
  4. Domain Adaptation via Appearance translation
  5. Contrast-aware Module
  6. Pseudo-Labels Learning
  7. Test-Time Adaptation Module and post-processing
  8. Experiment
  9. Dataset and metrics
  10. Setup
  11. Segmentation results analysis
  12. Baseline Comparison
  13. Ablation analysis
  14. Conclusion
  15. Acknowledgments.

Figures (3)

  • Figure 1: An overview of the proposed semi-supervised liver segmentation framework.
  • Figure 2: A) Original GED4 modality image; B) Style transferred with random matching histogram; C) Style transferred with GAN. The red triangle areas show the comparison of artifacts, while the blue triangle areas show the comparison of blurred pattern
  • Figure 3: Visualization of typical segmentation results comparing different methods on the CARE-Liver track dataset of CARE 2025. In A) and B), from the $2nd$ to the $5th$, we add mean teacher, pseudo-labels learning, contrast-aware module and CoTTA in sequence. C) and D) show the optimization of segmentation results by post-processing module.