Abstract

Semi-supervised learning for medical image segmentation, particularly within the prevalent Teacher-Student architecture, is frequently hindered by recursive error accumulation and the inherent rigidity of hard-threshold pseudo-labeling. These limitations lead to three critical bottlenecks: semantic-geometric misalignment, chronic confirmation bias, and significant loss of boundary information. To address these, we propose CSGE, a novel paradigm shifting from pseudo-label filtering to collaborative semantic-geometric refinement. CSGE synergistically leverages multimodal medical knowledge and geometric priors. By exploiting semantic cues from the textual modality, BiomedCLIP provides robust semantic anchoring, while SAM serves as a geometric refinement expert. We introduce a Conflict-aware Iterative Co-evolution module that detects discrepancies via XOR operations to drive iterative pseudo-label refinement. Furthermore, we replace hard thresholds with a spatial trust-aware soft-weighted consistency loss to maintain gradient flow within uncertain regions. Evaluated across three modalities (1%–30% labels), CSGE outperforms state-of-the-art methods, notably achieving a 4.9% average Dice improvement under the extreme 1% setting to demonstrate its robustness. Our code is available at https://github.com/yufei050201/CSGE.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3554_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/yufei050201/CSGE

Link to the Dataset(s)

N/A

BibTex

@InProceedings{WanYuf_CSGE_MICCAI2026,
        author = { Wan, Yufei AND Cheng, Hangbei AND Cui, Danchen AND Li, Junran AND Jiang, Jiahui AND Chen, Junxin AND Wei, Mingqiang AND Liu, Xueyu AND Wu, Yongfei},
        title = { { CSGE: Break the SSL Bottleneck in Medical Image Segmentation via Collaborative Semantic-Geometric Experts } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16883},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper proposes a semi supervised learning framework CSGE that integrates BiomedCLIP for semantic anchoring and SAM for geometric refinement, and introduces conflict aware iterative co evolution and spatial trust awareness to overcome semantic geometric misalignment, confirmation bias, and boundary loss.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    1.The paper synergizes BiomedCLIP’s semantic anchoring with SAM’s geometric refinement. 2.It introduces a conflict aware iterative co evolution module that uses XOR based discrepancy detection and iterative prompt resampling to correct boundaries and mitigate confirmation bias in teacher student learning. 3.The spatial trust awareness optimization replaces hard threshold pseudo label filtering with a soft weighting consistency loss, preserving gradient flow in uncertain regions and enhancing noise robustness.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    1.The framework heavily relies on two external foundation models (BiomedCLIP and SAM), which may limit its practical applicability in resource constrained clinical settings. 2.The conflict aware iterative co evolution module and spatial trust map introduce additional hyperparameters (e. g. , Gaussian kernel σ and loss weight λ) that require dataset specific tuning, potentially reducing generalizability across diverse medical imaging modalities. 3.The experimental validation is limited to standard U Net backbones and three public datasets; the paper lacks thorough analysis of multi class segmentation scenarios, cross domain shifts, or comparisons with more recent teacher student SSL methods.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    1.How is the conflict aware iterative co evolution module initialized and terminated? 2.Does the spatial trust map create circular dependency between the teacher and the pseudo label? 3.What is the computational overhead of integrating BiomedCLIP and SAM into the teacher student pipeline?

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The framework introduces several design choices that raise technical concerns. First, the conflict aware iterative co evolution module detects discrepancies via XOR operation on binarized predictions (teacher pseudo label vs. SAM mask), which discards uncertainty information and may amplify noise in early training stages. Second, the spatial trust map uses the Dice similarity between the teacher’s soft prediction and the final pseudo label as a global trust weight, potentially creating a circular dependency, since the pseudo label itself is partly derived from the teacher, this could reinforce existing biases rather than providing independent reliability quantification. Third, the method relies on pre computed gScoreCAM saliency maps from BiomedCLIP to define regions of interest, but these offline saliency maps are static and do not adapt to the student’s evolving representations, limiting dynamic refinement. Additionally, the loss formulation includes both an L2 consistency loss (L_con ) between student and teacher outputs and a spatially weighted unsupervised loss (L_unlabel ) against the refined pseudo label; the paper does not analyze potential conflicts or redundancy between these two supervisory signals.

    The experiments are confined to three 2D binary segmentation datasets (chest X ray, polyp, dermoscopic images) with no evaluation on 3D medical volumes (e.g., CT, MRI) or multi class segmentation tasks, which would better reflect real world clinical complexity. While the authors claim the framework is architecture agnostic, the main results only use a standard U Net backbone; the extension analysis briefly mentions MedSAM and SAM3 but provides no quantitative comparison or ablation on other segmentation networks (e.g., Swin UNETR, nnU Net). The paper reports average Dice/IoU over three random seeds but omits standard deviations or statistical significance tests, making it difficult to assess the robustness of the claimed 4.9% average improvement. Furthermore, the comparison with state of the art methods is incomplete, Table 1 is referenced but not shown in the provided excerpt, and no implementation details or hyperparameter tuning for competing methods are disclosed. Finally, despite claiming lightweight inference, the paper provides no runtime, memory footprint, or parameter count analysis for integrating two large foundation models (BiomedCLIP and SAM) into a teacher student pipeline, which is critical for assessing clinical deployability.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #2

  • Please describe the contribution of the paper

    This paper addresses key challenges in semi-supervised learning (SSL) by collaboratively leveraging a semantic expert (CLIP) and a geometric expert (SAM).

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
    • The paper clearly identifies and summarizes three core bottlenecks that limit the performance of semi-supervised learning, and proposes a method specifically designed to address them.
    • The overall presentation is clear, aside from a few minor details, and the experimental evaluation is extensive.
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    1.I have a few questions and comments regarding the proposed method.

    • In the conflict-aware iterative co-evolution module, what specific role do the resampled points play as prompts, in comparison with the positive and negative points? This is not clear.
    • In the sentence, ‘where d(x, y) is the Euclidean distance from pixel (x, y) to the nearest pixel in E_{SAM}’, do you mean the nearest foreground pixel? 2.It would be helpful to see the performance of a fully supervised method as an upper-bound in Table 1.
  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (5) Accept — should be accepted, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    Overall, this paper presents a well-motivated and clearly structured approach for SSL. The method is supported by extensive experiments, and the contribution appears meaningful despite a few minor points that would benefit from further clarification.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    My comments/questions primarily were about the explanation of the method, which were somewhat unclearly answered in the rebuttal. Nevertheless, I believe the paper possesses sufficient quality.



Review #3

  • Please describe the contribution of the paper

    The authors proposed CSGE, a semi-supervised medical image segmentation framework, that improves Teacher-Student learning by replacing hard-threshold pseudo-labeling with collaborative semantic-geometric refinement. It uses BiomedCLIP and SAM for semantic and geometric guidance, respectively, and introduces conflict-aware iterative refinement and soft trust-based consistency optimization, achieving strong gains, especially with extremely limited labels.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    1.Addresses an important problem in semi-supervised medical image segmentation under limited annotations. 2.Proposes a novel semantic-geometric refinement framework by integrating BiomedCLIP and SAM. 3.Shows strong empirical performance, especially in the 1% labeled data setting.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    1.Unclear linkage between stated limitations and target performance metrics, weakening the objective justification. 2.Insufficient ablation and statistical validation, making it difficult to assess the contribution of individual modules and reliability of improvements. 3.Presentation and organization issues, including scattered contributions, unclear figures, and inconsistent notation.

  • Please rate the clarity and organization of this paper

    Poor

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    Major comments 1.The manuscript should clearly specify the performance metrics targeted for improvement (e.g., Dice score, IoU, precision, recall, boundary F1-score). In particular, the authors should explicitly justify how the identified limitations, semantic-geometric misalignment, confirmation bias, and boundary information loss, affect these metrics. Without linking the limitations to measurable performance indicators, the motivation lacks a clearly defined objective. 2.The datasets and segmentation tasks used in the experiments should be clearly stated early in the abstract and introduction to improve clarity of scope and applicability. 3.The manuscript introduces multiple contributions (BiomedCLIP semantic anchoring, SAM geometric refinement, conflict-aware co-evolution, spatial trust awareness optimization) but the rationale behind each component is insufficiently consolidated. The contributions should be presented in a more focused and unified manner in both the abstract and introduction, as they currently appear scattered. 4.In Figure 1, the Student Network is shown in two locations, which may cause confusion. The figure does not appear to represent a clear generic framework. The term “co-evolution” should be explicitly defined. It is also unclear whether Figure 1(c) corresponds to the proposed method. The figure caption should be expanded to clearly explain each component. 5.In Figure 2, it is unclear whether information flows from the student network to the teacher network. Several abbreviations are used in the figure without definition, requiring the reader to infer their meaning. All terms should be explicitly defined in the caption or text. 6.The notations in Section 2 are difficult to follow. Symbols and variables should be clearly defined at first occurrence to improve readability. 7.The text modality is introduced relatively late in the manuscript, which interrupts the logical flow. Its role and motivation should be explained in the abstract and introduction. 8.The rationale for evaluating the method under different labeled data ratios should be clarified. The manuscript should explicitly explain how varying label proportions helps evaluate the stated limitations and robustness of the proposed approach. 9.Results should be reported with mean, standard deviation, and confidence intervals, along with rigorous statistical testing, to demonstrate reliability and significance of performance improvements. 10.An ablation study is required for all proposed modules to clearly identify the contribution of each component. Without ablation analysis, it is difficult to determine which part of the framework drives performance improvements.

    Minor comments 1.In Table 1, the Year and Source columns may not be necessary. Reducing table font size or simplifying table structure could provide additional space for deeper analysis. 2.Acronyms should be defined before first use, and notation formatting (including inconsistent camelCase usage) should be standardized throughout the manuscript. 3.The abbreviation SSL is commonly interpreted as self-supervised learning. To avoid ambiguity, the authors may consider using Semi-SL or consistently specifying semi-supervised learning. 4.Mention the row label in figure 3.5.The overall presentation quality can be improved through clearer structuring and stronger narrative coherence. 6.The manuscript would benefit from careful proofreading to improve grammar, clarity, and consistency.

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (2) Reject — should be rejected, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    1.Lacks clear validation, ablation, and statistical evidence for claims. 2.Presentation, notation, and figure clarity need significant improvement.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    The manuscript addresses an interesting and potentially relevant problem. I acknowledge that space limitations may have constrained the authors from presenting all methodological and experimental details, ablation study needed to fully support the proposed framework, and this should be taken into consideration. I also believe that the authors can adequately address the comments raised before submitting the final version. One additional point I missed earlier is that Fig. 3 requires a clearer explanation on p_t, M_sam, M_final.



Author Feedback

We sincerely thank the reviewers and AC for their careful reviews and insightful comments. We clarify several main issues below.

To R1 & R3 Result: All results are averaged over 3 runs; Mean±SD will be added if space permits. Preliminary tests (p<0.05, SD≤1.0%) confirm significance.

To R1 Technical Clarification: (1) Uncertainty: The co-evolution module initializes via Teacher and SAM discrepancies and runs only 1 iteration to prevent redundant overhead; XOR localizes conflicts solely for resampling and never computes the loss. Gradients flow through the continuous W(x,y), perfectly preserving uncertainty. (2) Circular Dependency: M_final is refined independently by SAM, measures consensus with external knowledge, breaks the confirmation loop. (3) Static gScoreCAM suffices since macroscopic anatomy is constant; SAM and the subsequent processes handle dynamic refinements. (4) Loss Complementarity: L_con aligns global representations, while L_unlabel enforces local boundary precision—strictly complementary. Sec. 3.3 shows identical hyperparameter across three datasets ensures robustness, making dataset-specific tuning unnecessary. Clinical Deployability: BiomedCLIP and SAM operate strictly during training (peak 16GB, V100). Inference only uses a lightweight U-Net, taking <17ms and ~1.82GB at 256*256 ensuring scalable deployment. Baselines & Efficacy: With limited space, Table 1 selectively reports four SOTAs (CVPR/TMI 2025) and four reproduced classics. CSGE consistently dominates across 5 ratios. At 1% labels (ΔDSC>4.9%), it effectively reduces annotation requirement from 10% to 1%. 2D Data & Agnosticism: We focused on 2D binary segmentation as a rigorous testbed, leveraging its status as a top-tier SSL benchmark. Sec. 3.2 validates our architecture-agnostic. Replacing SAM with MedSAM/SAM3 yields cross-dataset gains (+3.2%DSC, +2.8%IoU), or integrating nnU-Net adds a 5.02% boost. This high modularity makes future 3D/multi-class extension straightforward.

To R2 Resampling: While initial prompts provide global localization, resampled points target XOR conflict regions to resolve ambiguities and refine boundaries where Teacher and SAM disagree. Distance d: d(x,y) is the Euclidean distance to E_SAM boundary. Small d prioritizes SAM’s geometric precision, large d trusts teacher semantics for regional consistency. Upper-bound: Fully supervised results will be included.

To R3 Metrics, Ratios & Contributions: Standard Semi-SL metrics DSC and mIoU globally capture our bottlenecks: semantic incompleteness (misalignment), false positives (bias), and blurred boundaries (boundary loss). We will detail this mapping in the revision. Multi-ratio evaluations (1%-30%) demonstrate our framework’s robustness, specifically its ability to surpass SOTA at the 1% label, where traditional Semi-SL often fails. Ablation & Results: Due to space limits, Sec 3.3 highlights key components ablations. We will clarify how multiple components form unified modules: As the concurrent work SEMISAM-O1 (MedIA 2026) confirms, single foundation models face ceilings due to lacking constraints, our semantic-geometric integration is theoretically and empirically essential (+SAM/BiomedCLIP, DSC+5.1%). XOR mitigates normal region misclassification (+Conflict Iterative, DSC+3.6%); spatial trust maximally preserves boundary information (+Spatial Trust Map, DSC+1.6%). Manuscript Structure: The revised Abstract/Introduction in camera ready will specify tasks/datasets and introduce the text modality earlier to explain how BiomedCLIP’s macroscopic guidance prevents early bias from SAM. Metric impacts and contributions will be unified. Figures & Formatting: Captions will clarify framework links (Fig. 1) and EMA flow (Fig. 2), and add row labels (Fig. 3). “Co-evolution” is defined as SAM’s iterative geometric refinement of student predictions. Terminology is standardized to “Semi-SL”, with all symbols and acronyms defined at first occurrence alongside thorough proofreading.




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    Recommend rebuttal despite split reviews

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    The proposed semi-supervised medical image segmentation framework (CSGE), which collaboratively integrates BiomedCLIP for semantic anchoring and SAM for geometric refinement, is recognized as a well-motivated and empirically effective solution under extremely sparse annotations. While initial concerns were raised regarding structural circular dependency, insufficient ablation studies, and presentation clarity, consensus among reviewers post-rebuttal highlights that the authors have successfully clarified these queries and committed to resolving the presentation issues in the final camera-ready version.



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    After the rebuttal, all reviewers agreed that the paper should be accepted. The Area Chair supports this consensus and recommends acceptance.



Meta-review #3

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    In general, the reviewers provided positive comments on this work after the rebuttal. The main contribution is the integration of both BiomedCLIP and SAM to improve semi-supervised learning, which was acknowledged by all reviewers and is potentially beneficial to the community. Considering concerns remain regarding the lightweight claim and some missing details, the authors are required to address these issues in the camera-ready version. Therefore, acceptance is recommended.



back to top