acceptodds
Under review as a conference paper at ICLR 2027

CHCL-Seg: A Competitive Hierarchical Contrastive Learning Framework for Semi-Supervised 3D Medical Image Segmentation

Abstract

3D medical image segmentation plays an important role in intelligent clinical applications. Although existing deep learning methods have achieved substantial progress in this field, they are limited by the high cost of 3D manual annotation. Semi-supervised learning, which combines a small amount of labeled data with a large amount of unlabeled data for training, is an effective approach to alleviate label scarcity. However, existing semi-supervised methods for 3D medical image segmentation still face the following problems: confirmation bias is difficult to effectively mitigate, target boundary segmentation accuracy is insufficient, and cross-dataset semantic representation is difficult. To address these challenges, this paper proposes a competitive hierarchical contrastive learning framework for semi-supervised 3D medical image segmentation, termed CHCL-Seg. We first construct a dual-branch network with shared parameters, and introduce an uncertainty-aware competitive loss (UnCL) to balance noise filtering and boundary information utilization, thereby mitigating confirmation bias. We then propose a hierarchical contrastive learning mechanism, where at the high level, the Dense-OT module performs class prototype alignment to alleviate class imbalance, and at the low level, the V-DACL module mines hard boundary samples to enhance boundary feature discrimination. Extensive experiments on three public datasets, LA, BraTS-2019, and ACDC demonstrate that the proposed method comprehensively outperforms existing state-of-the-art semi-supervised segmentation methods. Our code will be released upon acceptance.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
3D medical image segmentation, semi-supervised learning, contrastive learning, uncertainty-aware learning, hierarchical representation, confirmation bias

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_aSZLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors proposed a medical image semi-supervised segmentation method, CHCL-Seg. The network is composed of ABCP, UnCL loss function with hierarchical contrastive learning. Experiments prove the effectiveness of the proposed modules, especially on boundary-wise metrics.

Critical strengths
  1. Semi-supervised learning in medical imaging segmentation is a difficult problem for a long time. This method addressed the problem by introducing dual students, and significantly improves the boundary accuracy.
  2. The experiments are abundant to prove the effectiveness of CHCL-Seg.
Critical weaknesses
  1. The UnCL chooses the student branch with higher confidence, while higher confidence does not necessarily indicate higher accuracy, particularly when both students make wrong predictions. The paper lacks sufficient analysis to prove that this competitive strategy effectively reduces confirmation bias.
  2. V-DACL assumes that low-density features correspond to informative hard samples near object boundaries, but such features may also represent noises. Furthermore, confidence-based filtering may exclude uncertain boundary voxels. The paper lacks enough evidence of this sampling strategy.
  3. The paper claims that ABCP is a prerequisite for feature-level contrastive learning. However, the ablation results do not fully support this conclusion. As to LA, HCL alone improves Dice from 89.53 to 89.96, and similarly on BraTS-2019 where Dice increases from 85.05 to 85.81. Furthermore, adding HCL to ABCP improves Dice on LA (90.82 to 91.69) but slightly decreases Dice on BraTS-2019 (87.09 to 86.97). The combination of UnCL and HCL without ABCP also shows opposite trends across the two datasets. These inconsistent interactions suggest that ABCP is not necessarily a prerequisite for HCL, and the claimed dependency requires further justification.
  4. Since the pipeline contains students using shared parameters but different Dropouts, it is necessary to validate across different random seeds settings to prove the robustness and consistent improvements of this method.
Questions for authors
  1. How does UnCL ensure that selecting the student branch with higher confidence leads to more reliable predictions, especially when both branches make incorrect predictions? It is suggested to provide quantitative evidences.
  2. How do the authors verify that the low-density features selected by V-DACL correspond to informative boundary samples rather than noise or outliers? Additionally, does the confidence threshold exclude potentially useful uncertain boundary voxels?
  3. The ablation results need more clarifications, like is ABCP truly a prerequisite for HCL, or does it only provide complementary benefits under certain conditions?
  4. How stable is the method under different initialization settings?
Reviewer_MZBsOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces CHCL-Seg, a hierarchical contrastive learning framework designed for semi-supervised medical image segmentation. The method addresses challenges such as noisy pseudo-labels, class imbalance, and inaccurate boundary segmentation when annotations are limited. CHCL-Seg incorporates an EMA teacher and parameter-shared student branches, and combines annealed bidirectional copy-paste with uncertainty-aware competitive supervision. Its hierarchical contrastive learning mechanism enhances feature discrimination through class-prototype alignment using optimal transport and density-aware voxel sampling. Experiments on LA, BraTS-2019, and ACDC datasets at various labeling ratios demonstrate improved overlap metrics compared to existing methods. Ablation studies further evaluate the impact of each component.

Critical strengths
  1. The experiments demonstrate the method’s effectiveness with limited annotations. In the BraTS-2019 ablation study, adding hierarchical contrastive learning improves both overlap and boundary metrics at 5% and 10% labeled data. These results show that this component can improve segmentation when annotated scans are scarce.

  2. CHCL-Seg achieves higher reported Dice scores than the recent baselines on LA, BraTS-2019, and ACDC at both 10% and 20% labeled data. This demonstrates that the framework is competitive across different anatomical segmentation tasks and amounts of labeled data.

Critical weaknesses
  1. The ACDC experiments do not directly validate the proposed voxel-level boundary-mining method. Appendix B.2 states that ACDC uses instance-level Masked Average Pooling DACL, whereas LA and BraTS-2019 use voxel-level V-DACL, and Table 5 distinguishes these formulations as separate variants. Although Figures 1 and 3 and Table 3 show boundary improvements on ACDC, no evaluation of voxel-level V-DACL on ACDC is reported. These results therefore do not establish that the proposed voxel-level mechanism drives those improvements. The authors should make the dataset-specific formulations explicit in the main text and provide ACDC ablations to identify the contributions of the components actually used. They should also either evaluate voxel-level V-DACL on ACDC or qualify claims that the ACDC results support that contribution.

  2. The robustness and reliability of the reported gains are insufficiently established. Although the authors claim that fine-grained retuning is unnecessary, Table 8 reports Dice values ranging from approximately 85.4 to 88.49 as the contrastive weights vary. This variation is greater than HCL’s 0.49-point gain over the no-HCL baseline on BraTS-2019 with 20% labels, and both sweep averages fall below that baseline. In addition, Table 8’s standard deviations reflect variation across weight settings rather than repeated runs, so small improvements such as the 0.17-point Dice gain over BCSI on BraTS-2019 (10% labels) remain uncertain. The authors should clarify the evaluation split and selection procedure, report means and standard deviations across training seeds, and test whether HCL’s advantage is consistent under small weight changes.

Questions for authors
  1. Are there controlled ACDC ablation studies that isolate the effects of ABCP, UnCL, Dense-OT, and instance-level DACL? Which components are responsible for the boundary improvements shown in Table 3?

  2. Which data split was used to select the contrastive weights and generate the sensitivity results in Table 8? Were these weights fixed prior to evaluating the final model on the test set?

  3. Do the current sensitivity experiments demonstrate that HCL continues to outperform the no-HCL baseline with minor changes to the selected weights? Please specify the tested settings and scores that support the claim that fine-grained retuning is unnecessary.

  4. Appendix B.2 specifies training seed 1337. Do all CHCL-Seg comparisons and ablations use this seed? Are multi-seed means and standard deviations available to assess the robustness of the 0.17-point Dice gain over BCSI (10% labels) and the 0.49-point gain from HCL (20% labels) on BraTS-2019?

Reviewer_U2jbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CHCL-Seg, a competitive hierarchical contrastive learning framework for semi-supervised medical image segmentation. It introduces uncertainty-aware competitive learning (UnCL), dense optimal transport-based prototype alignment (Dense-OT), and voxel-wise density-aware contrastive learning (V-DACL). Experiments on LA, BraTS-2019, and ACDC demonstrate competitive segmentation performance, particularly in boundary accuracy.

Critical strengths
  1. The integration of uncertainty-aware learning and hierarchical contrastive learning is well motivated. The proposed framework addresses important challenges in semi-supervised medical image segmentation.
  2. The method achieves promising results on three benchmark datasets, particularly on ACDC. The ablation studies provide useful evidence supporting the overall framework.
Critical weaknesses
  1. The proposed framework combines several established techniques, including competitive learning, optimal transport, and contrastive learning. Its fundamental differences from closely related methods, particularly CDS, require clearer justification.
  2. The individual contributions of Dense-OT and V-DACL are not fully isolated. Moreover, ACDC adopts instance-level DACL rather than the proposed voxel-level formulation, making it difficult to attribute the reported improvements to the new design.
Questions for authors

Q1. Could the authors clarify the main methodological differences between CHCL-Seg and CDS, particularly regarding competitive learning and hierarchical contrastive learning?

Q2. Could separate ablation studies for Dense-OT and V-DACL be provided to better establish their individual contributions?

Q3. Why is instance-level DACL used on ACDC instead of the proposed voxel-level V-DACL? How does this choice affect performance?

Q4. Could the authors provide quantitative evidence that low-density feature samples correspond to difficult boundary regions?

Q5. Could the authors clarify how UnCL preserves informative boundary voxels while suppressing unreliable high-uncertainty predictions?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.