Abstract

Semi-supervised medical image segmentation (SSMIS) is critical, yet suffers from pseudo-label noise and boundary uncertainty in low-label regimes. While foundation models like SAM offer strong priors, their high-confidence semantic bias can amplify errors when naively integrated into pseudo-supervision loops. We propose DiCoP- SAM, a two-stage collaborative framework that closes the loop between a CNN and SAM via automatic, error-aware corrective prompting that requires neither GT masks nor manual prompts for unlabeled images. Stage-I (TriGate-CPS) augments cross pseudo-supervision with a fore- ground/background/uncertainty tri-gate mechanism to produce reliable fused pseudo-labels and explicitly localize high-risk disagreement bands. Stage-II (DiCoP) harvests these bands to generate such corrective posi- tive/negative point prompts, driving parameter-efficient SAM adaptation; refined predictions are distilled back to the CNN, forming a self-correcting “predict → correct → relearn” loop. On CT, MRI, and ultrasound bench- marks, DiCoP-SAM consistently outperforms semi-supervised and SAM- assisted baselines, approaching fully supervised performance with limited labels (e.g., 83.13 Dice vs. 83.30 for TRFE+ on TN3K with 20% labels).

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1048_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/Fzuwxy/DiCoP-SAM

Link to the Dataset(s)

TN3K/TG3K thyroid ultrasound: https://github.com/haifangong/TRFE-Net-for-thyroid-nodule-segmentation ACDC cardiac MRI: https://www.creatis.insa-lyon.fr/Challenge/acdc/ EAT cardiac CT: in-house anonymized dataset; not publicly available due to institutional/privacy restrictions.

BibTex

@InProceedings{WanXin_DiCoPSAM_MICCAI2026,
        author = { Wang, Xinyu AND Liu, Jiayao AND Wang, Youlin AND Su, Jiawei AND Luo, Zhiming AND Lian, Sheng},
        title = { { DiCoP-SAM: Disagreement-Band Guided Corrective Prompting for Collaborative Semi-Supervised Medical Image Segmentation } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16880},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper proposes a two-stage semi-supervised medical image segmentation (SSMIS) framework—DiCoP-SAM. This method aims to address the problem that base models (such as SAM) easily amplify high-confidence errors in semi-supervised pseudo-label loops. The authors conducted experiments on three datasets, demonstrating the superior performance of this method with low label ratios.

  • 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 core idea of using disagreement between models in a semi-supervised setting to generate corrective signals is intuitive and well-motivated. The proposed two-stage framework is technically sound and well-structured. Stage-I (TriGate-CPS) is a sensible extension of cross pseudo-supervision that explicitly models uncertainty to improve pseudo-label quality. Stage-II (DiCoP) presents a clever, GT-free mechanism to translate model disagreement into actionable, corrective prompts for SAM, which is a key contribution.

  • 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 overall framework is a sophisticated combination of existing ideas: cross pseudo-supervision (CPS), knowledge distillation, and prompt-based finetuning of SAM. While the integration is novel, the core building blocks are not. For example, using model disagreement or uncertainty to guide learning is a known concept in active learning and co-training. 2.The point generation strategy in DiCoP (Section 2.2) relies on several manually set hyperparameters. How were these values chosen? The paper mentions these values in Section 3 but does not provide a sensitivity analysis. This is a significant weakness. 3.The claim of being “GT-free” for prompt generation is slightly misleading. While the point prompts for unlabeled data are indeed generated without GT, the entire system is trained on a labeled set where GT is used to supervise the foreground, background, and uncertainty. 4.The computational overhead of the proposed two-stage training process is not discussed. The framework involves running two CNN branches, fusing their outputs, running SAM with generated prompts, and performing distillation. This seems computationally expensive compared to standard SSL methods. A discussion of training time, memory usage, and inference speed would be valuable for assessing the practical applicability of the method. 5.The description of the loss function for the uncertainty gate is vague. On page 4, it states it uses BCE with a “soft high-risk target derived from the GT boundary band. “ The exact formulation of this soft target is not provided, which hinders reproducibility. 6.The paper mentions applying “entropy-guided patch-level CutMix” in the late training stage. This seems like an important detail for achieving the reported performance, yet it is mentioned only briefly and not included in the main formulation or ablation studies.

  • 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 submission does not provide sufficient information for reproducibility.

  • 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.

    (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?

    This paper proposes DiCoP-SAM, a two-stage collaborative framework for semi-supervised medical image segmentation (SSMIS). The core idea of using disagreement between models in a semi-supervised setting to generate corrective signals is intuitive and well-motivated.

  • Reviewer confidence

    Very confident (4)

  • [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 proposes a two-stage training framework to fully leverage the promptable nature of SAM as a refinement process for semi-supervised segmentation, forming a self-correcting “predict->correct->relearn” loop. Experimental results on three datasets with different modalities demonstrate the effectiveness of the proposed method.

  • 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.Reasonable method design. The authors propose a novel integration of SAM within a semi-supervised segmentation framework through a two-stage training scheme. In the first stage, high-risk regions are identified; in the second, SAM is provided with corrective point prompts (both positive and negative) and a mask prompt to generate refined predictions, which are subsequently used as supervision signals. 2.Comprehensive experiments. Extensive experimental results on three datasets with two different ratios of labeled data demonstrate the effectiveness of the proposed method.

  • 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.Incomplete SAM-based semi-supervised comparison methods. The experimental comparisons are currently limited to two SAM-based semi-supervised methods (SemiSAM and CPCSAM), both introduced in 2024.To more rigorously position the proposed contribution within the rapidly evolving research landscape, it is important to include comparisons with more recent state-of-the-art methods, such as KnowSAM [R1] and SemiSAM+ [R2]. It is also well worth noting that KnowSAM is also validated on the ACDC dataset and TN3K dataset. 2.Unfair comparison with other SAM-based methods. The two compared baseline methods (SemiSAM and CPCSAM) employ the general-purpose SAM as their backbone. In contrast, this work utilizes SAM-Med2D, which benefits from large-scale medical image pre-training. This fundamental discrepancy in model initialization creates an uneven starting point, making it difficult to attribute performance gains solely to the proposed architectural or algorithmic contributions. 3.Lack of experiments with extremely limited labeled data. While the paper validates the method under 10% and 20% labeled data settings, its performance under extremely low-data regimes remains unexplored. Given the practical challenges of acquiring expert annotations in medicine, and the potential of foundation models like SAM-Med2D to learn from limited supervision, it is crucial to evaluate the method in few-shot scenarios, for instance, with only one or two labeled examples per class. Such experiments would more comprehensively demonstrate the method’s utility in real-world, data-scarce clinical environments.

    References: [R1] Huang, Kaiwen, et al. “Learnable prompting sam-induced knowledge distillation for semi-supervised medical image segmentation.” IEEE Transactions on Medical Imaging 44.5 (2025): 2295-2306.[R2] Zhang, Yichi, et al. “SemiSAM+: rethinking semi-supervised medical image segmentation in the era of foundation models.” Medical Image Analysis (2025): 103733.

  • 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 does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • 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.

    (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?

    This method design is reasonable and effective to some extent on three datasets. However, more recent state-of-the-art SAM-based methods are not included for a comprehensive comparison. In addition, the leverage of the medical foundation model SAM-Med2D leads to unfair comparison with other methods.

  • Reviewer confidence

    Very confident (4)

  • [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 #3

  • Please describe the contribution of the paper

    1)Proposing DiCoP-SAM, a collaborative paradigm that breaks the error amplification loop in semi-supervised medical image segmentation by establishing a bidirectional “CNN ↔ SAM” interaction with GT-free, error-aware prompting. 2)Enhancing pseudo-label reliability and enable targeted error correction through TriGate-CPS, which explicitly localizes high-risk disagreement regions via tri-gate uncertainty modeling. 3)Automating corrective prompting with DiCoP, which converts disagreement bands into sparse, error-aware point prompts without requiring manual annotations, thereby enabling lightweight SAM adaptation through bidirectional distillation. 4)Extensive experiments across three imaging modalities demonstrate promising performance, where our method consistently outperforms both SSL specialists and SAM-based variants under low-label regimes.

  • 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 precisely identifies the core challenges in semi-supervised medical image segmentation: pseudo-label noise and boundary uncertainty, particularly in low-label regimes where the high-confidence semantic bias of foundation models such as SAM may amplify errors. This issue holds significant practical relevance in clinical applications. The methodological innovation is substantial: the TriGate-CPS module enhances cross pseudo-supervision through a foreground/background/uncertainty tri-gate mechanism, explicitly localizing high-risk disagreement regions and improving the reliability of pseudo-labels. The DiCoP module introduces, for the first time, the ability to automatically generate corrective point prompts from disagreement regions without requiring ground truth, enabling lightweight SAM adaptation and forming a self-correcting “predict → correct → relearn” loop. The bidirectional distillation mechanism, in which refined SAM predictions are distilled back to the CNN, enables mutual enhancement between the CNN and SAM, overcoming the limitations of traditional single-model approaches or simple concatenation. The experimental design is comprehensive, and the results demonstrate superior performance.

  • 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 channel mixing operator ϕ(⋅) in the TriGate module is not explicitly defined (e.g., whether it is a 1×1 convolution or an MLP), which affects reproducibility. Similarly, the implementation details of spatial suppression Top-K selection in DiCoP point sampling, such as the mathematical form of the distance weighting function φ(d(i,D)), are not specified. 2) There is a lack of sensitivity analysis for hyperparameters, including τ− and τ+​ ; the distillation temperature T; the suppression radius r; and the number of sampled points K, loss trade-off factor λ_sam,λ_kd and λ_fuse, among others. 3) Inconsistent parameter notations appear across the text: for instance, τl​ and τh​ in Equation (5) are later referred to as τ− and τ+. 4) Although the authors claim “lightweight SAM adaptation,” the comparative experiments do not report efficiency metrics such as training time, inference time, number of parameters, or FLOPs.

  • 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 submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • 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.

    (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 paper demonstrates strong novelty and holds certain clinical relevance. However, the experimental section requires further completeness, and some descriptions in the methodology part need to be more detailed. If the authors can supplement these aspects, I believe the paper could be considered for acceptance.

  • 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



Author Feedback

We appreciate the reviewers’ recognition of our motivation, method design, and experimental results, as well as their constructive feedback.We would like to clarify several concerns raised in the reviews.

First, the term “GT-free” was intended to describe only the corrective prompt generation process on unlabeled data, rather than the entire training pipeline. DiCoP-SAM still follows the standard semi-supervised setting, where a small labeled set is used for supervised training. Specifically, for unlabeled images, DiCoP generates positive and negative corrective point prompts from disagreement bands without using ground-truth masks or manual prompts. We will rephrase this section in the final version for better clarity.

Second, we agree that the notation and several implementation details can be further clarified for reproducibility. We will correct the related symbols and add the missing implementation details in the camera-ready version. The exact settings, hyper-parameters, and configuration files will also be released with the code after publication. Regarding computational cost, we clarify that our SAM adaptation is parameter-efficient: the image encoder is frozen, and only the prompt encoder and mask decoder are updated.

We sincerely appreciate the reviewers’ insightful comments, which guide us in improving the clarity, reproducibility, and presentation of the final manuscript.




Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • 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.

    All three reviewers recommend Weak Accept, and the reviews consistently recognize the paper as addressing an important problem in semi-supervised medical image segmentation. The reviewers generally agree that the proposed two-stage DiCoP-SAM framework is well motivated and experimentally effective across multiple datasets and low-label settings. The main concerns are mostly addressable, which are important for revision but do not undermine the core contribution. Therefore, given the consistent reviewer support and the paper’s clear methodological and empirical value, an acceptance is justified.



back to top