Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

3D medical image segmentation is critical for disease diagnosis and treatment planning, yet requires costly large-scale fine-grained annotations. Semi-Supervised Learning (SSL), combining limited labeled data with massive unlabeled data, offers an ideal solution. However, mainstream Mean Teacher (MT) methods suffer from two core limitations: (1) The single-teacher paradigm is prone to confirmation bias during iteration, causing error accumulation especially in low-contrast challenging regions; (2) Limited labels fail to provide sufficient anatomical priors. To this end, we propose a dual-teacher knowledge distillation framework for semi-supervised 3D medical segmentation, integrating an internal teacher, an external teacher and a learnable prototype guidance mechanism. An EMA-updated internal teacher provides stable consistency supervision, while a frozen SAM-Med3D external teacher generates high-quality segmentation maps via prompt-driven generation to distill anatomical shape and boundary priors. We introduce a learnable prototype module to calibrate internal pseudo-labels via residual fusion, and design an entropy-guided loss to mine hard examples in high-uncertainty regions to boost feature discriminability.Experiments on LA, Pancreas-CT and BraTS19 benchmarks demonstrate the proposed method achieves state-of-the-art performance on most metrics.The code is available at https://github.com/Aguang094/Dualteacher-with-SAM.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3395_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/Aguang094/Dualteacher-with-SAM.

Link to the Dataset(s)

N/A

BibTex

@InProceedings{HuoShu_DualTeacher_MICCAI2026,
        author = { Huo, Shuaiguang AND Sang, Zifeng AND Luo, Xu AND Zhu, Jinghua AND Xi, Heran},
        title = { { Dual-Teacher Knowledge Distillation with SAM and Learnable Prototype Guidance for Semi-supervised 3D Medical 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

    The authors improve medical data segmentation tasks by including a pre-trained Med-SAM prediction as pseudo-label.

  • 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 mathematical notation is concise. The reviewer particularly likes that lower-case variables (b,h,w,d) are used to index Upper-case counts (B,H,W,D), which makes it easier to follow the math.
    • An ablation study shows the impact of removing some submodules.
    • Compared to related methods, the results show improved prediction values.
  • 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.There are several parts in the paper that are not clearly stated: a) Most importantly, the exact task to be solved is mentioned nowhere. Is the goal to predict voxel-based class labels (and which classes would this be), or just binary foreground-background prediction? Or is the task tumor localization/segmentation? The authors should update their title accordingly, which proposes improvements in general segmentation tasks. b) What exactly do the evaluation metrics measure? DICE and Jaccard are binary metrics, so the task is binary foreground/background prediction? c) The role of Med-SAM could be explained more explicitly. It appears as if only the foreground prediction mask (\tilde M_u) of Med-SAM is used to guide the results. This should be stressed more. Also, it is of utmost importance to state whether MED-SAM was trained on any of the datasets that the authors used in their experiments. d) There is absolutely no detail given on the backbone that the authors used to perform the prediction. e) The role of the prototypes is not stated clearly. To the reviewer’s understanding, prototypes are bound to classes, and independent 0f voxel locations. f) The proposed method has several hyperparamters (the number of prototypes, loss weights) that are mentioned nowhere. Did the authors optimize these parameters on the test sets?

    2.The mathematical notation is incompletely described: a) Section 2.2: There are many variables (B, H, W, D, d_q, C) and operations (\phi, \psi) that are defined nowhere. Why do the authors predict C+1 classes, and not C classes? b) Section 2.3: What is z used in (7), and how is S_b,c^loc defined? What does L_b,q,: in (8) refer to? What is q indexing over (wasn’t this l in section 2.2)? What is \ell_b,n in (11) – shouldn’t this somehow sum over voxel locations (h,w,d)? How are high-confidence anchors defined? c) Section 2.4: At least two loss functions from (13) are not (explicitly) defined anywhere.

    3.The experimental evaluation is not very clear: a) The method requires large amounts of unlabeled data, but it is not mentioned anywhere where this data is coming from. Which of the compared methods make use of the same data sources? b) The motivation on why the specific datasets are selected would be important. What can we learn from the experiments on the different datasets? c) While it is mentioned that the authors ran 100 bootstrap resamples to compute means and standard deviations, the tables only present the means, and no statistical analysis is performed to provide significance tests.

    4.The conclusion states that improvements are “especially in challenging regions with low contrast and ambiguous boundaries”, but the paper does not provide any evidence for such a claim.

  • 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

    • Given the page limit, the abstract is unnecessarily long. The space could be better used to provide all details of the proposed method.

    • Some of the abbreviations capitalize the words (Mean Teacher, MT), while some don’t (exponential moving average, EMA). The authors should be consistent.

    • Graphics are stored and imported as pixel graphics (. jpg/. png). Better options would be vector graphics (. eps/. pdf).

    • The manuscript uses a mixture of British (especially in Figure 1) and American English. The authors should select one language, and stick to it.

    • The term “Distillation” typically refers to predict/replicate internal states of the teacher. In this work the predicted class (argmax in (11)) of the SAM model is targeted, which would be called pseudo-labels.

    • Table 1 should highlight the best results in bold, not the authors’ own results – this is misleading. Additionally highlighting the second-best value (for example in italics) could emphasize that the method performs well on average.

    • From figure 2, only the Positive predictions and False Negatives can be observed. When adding a third color where predictions and ground truth overlap would also highlight True Positives and False Positives. In that case, the Ground Truth column could be avoided.

    • Space could be gained by not repeating the result numbers from the tables in the text.

  • 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.

    (3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

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

    While the reviewer sees potential in the proposed method, the literal presentation of the paper needs to be improved to increase reproducibility. It needs to be verified that the Med-SAM model was not trained on the datasets/samples that the authors use in their evaluation.

  • 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 major concerns were clarified sufficiently well. I hope that the authors actually implement the changes as proposed in the rebuttal, which would clearly improve the paper.



Review #2

  • Please describe the contribution of the paper

    This paper proposes a semi-supervised 3D medical image segmentation framework. Building upon Mean Teacher (MT), the method introduces an external teacher model (based on SAM-Med3D) and incorporates a prototype mechanism with contrastive learning to inject categorical semantic information, thereby enhancing feature discriminability. The method achieves state-of-the-art (SOTA) performance.

  • 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 integration of external anatomical priors into the SSL pipeline via the external teacher model. (2) The introduction of learnable prototype guidance reduces bias and promotes result stability. (3) The methodology section is clearly presented with effective use of figures and text, and the experimental design is relatively comprehensive.

  • 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) Due to the introduction of a large external teacher model and complex prototype modules, the computational overhead may increase significantly. The training and inference costs should be reported. (2) The paper lacks explanations for hyperparameter selection and necessary parameter details. Clarifying these aspects would enhance the reproducibility of the results. (3) Reporting necessary statistical significance metrics alongside the results would further strengthen the validation of the findings. (4) Additional explanation is required regarding how the prompt points for SAM-Med3D are specifically generated.

  • 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.

    (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, the method proposed in this paper is novel, effective, and constitutes a meaningful contribution. However, the manuscript lacks some necessary information that requires 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.

    Authors sufficiently address the points raised in the reviews.



Review #3

  • Please describe the contribution of the paper
    • Proposes a dual-teacher distillation framework for semi-supervised 3D medical image segmentation.
    • Integrates an internal exponential moving average teacher with an external frozen SAM-Med3D teacher.
    • Introduces a learnable “prototype guidance module” to calibrate pseudo-labels, and to mitigate confirmation bias.
    • Develops an entropy-guided prototype calibration loss for hard-example mining in high-uncertainty regions (a form of active learning).
    • Validates the approach on LA, Pancreas-CT, and BraTS19 datasets showing state-of-the-art performance.
  • 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.
    • Integration of SAM-Med3D as an external teacher successfully injects rich anatomical priors into the learning process.
    • Learnable prototype mechanism is an elegant solution to the well-known confirmation bias problem in Mean Teacher setups.
    • Strong approach also to incorporate hard examples
    • Comprehensive experimental validation across three distinct and standard 3D medical datasets.
    • Ablation studies are thorough and clearly demonstrate the isolated performance gains of the SAM teacher, prototype guidance, and uncertainty calibration.
    • Achieves strong quantitative improvements, particularly in boundary metrics like Average Surface Distance.
  • 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.
    • Thinking about the analogy to active learning, the paper could benefit from a discussion/positioning w.r.t. active learning frameworks like MONAI-Label (https://doi.org/10.1016/j.media.2024.103207) or nnInteractive (https://arxiv.org/abs/2503.08373).
    • The paper could also benefit from a discussion on the computational overhead introduced by the dual-teacher setup and prototype module during training.
    • Missing comparison of inference times and computational complexity against baseline models.
    • What would happen if the frozen SAM-Med3D model fails on highly specific out-of-distribution modalities? Paper would benefit from addressing potential limitations in the SAM-Med3D ”oracle”.
  • 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

    product operator - especially formula 7 threw me off. Consider leaving it away, the way the sentence continues implies the full-stop, leasving it away still makes sense, but would be more readable.

    • Perhaps consider adding a brief section or table in the supplementary material detailing training memory requirements and inference speeds (if space allows)
    • Discuss failure cases. It would be helpful to know if & where the SAM-Med3D priors might misguide the student model?
  • 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 core methodological proposal of combining an internal Mean Teacher with an external foundation model teacher is sound and well-motivated, even though a bit incremental.
    • Experimental results are robust and consistently outperform competing methods on standard benchmarks.
  • 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.

    I am maintaining my weak accept rating. Authors provided a table with numbers on computational overhead and inference times. Their positioning against active learning frameworks makes sense, the explanation of how the system deals with SAM-Med3D failures (e.g. on out-of-distribution data) sounds reasonable too.



Author Feedback

We sincerely thank all reviewers for their constructive feedback. We address the major concerns below.

1.Methodological Clarity and Implementation Details (R1-Q1a,b,d)Task & Metrics and Backbone: All tasks are binary foreground/background segmentation:left atrium (LA), pancreas (Pancreas-CT), whole tumor (BraTS19, FLAIR only). Dice and Jaccard are binary;ASD and 95HD are foreground-specific, consistent with all compared methods.For fair comparison, VNet is used on LA and Pancreas-CT, and 3D U-Net on BraTS19, following prior work. (R1-Q1c,R2-Q4)SAM Instructions & Prompt Generation:SAM-Med3D provides shape-aware, voxel-wise supervision, not a binary mask. EMA teacher generates a pseudo-mask for rough localization; point prompts are sampled from pseudo-mask with fixed ratio and count, then fed into SAM to produce a probability map. Student’s prediction is aligned via MSE loss (Eq.6). Crucially, SAM-Med3D was not trained on any evaluation dataset—its pre-training uses completely disjoint public data. (R1-Q1f,R2-Q2)Hyperparameters: Number of prototypes L was ablated over 0,…,16,with L=8 chosen as optimal.λ1, λ2, λ3, μ, τ were selected via grid search on the validation set;we will report the chosen values and brief sensitivity remarks in the revision.

2.Prototype Mechanism & Notation (R1-Q1e, R1-Q2a,b,c)The reviewer correctly notes that prototypes are class-bound and location-independent.Learnable prototypes are refined by self-attention and interact with voxel features via cross-attention. They are assigned class distributions, allowing multiple prototypes per class. They serve as global semantic anchors for pseudo-label calibration (Sec.2.2) and feature contrast (Sec.2.3). Our framework predicts C foreground classes + one background channel, ensuring multi-class generality (even on binary benchmarks).We will define all symbols (B, F, H′, W′, D′, dq, ϕ, ψ) and missing losses (L_sup, L_int-dist) in revision.

3.Evidence for Improved Boundaries (R1-Q4)Evidence in Table 4: with entropy-guided calibration, ASD on LA drops from 1.62 to 1.46 and on BraTS19 from 1.45 to 1.20 – directly capturing boundary accuracy. Dedicated boundary visualizations, omitted due to space, will be included in the revision.

4.Computational Overhead (R2-Q1, R3-Q2,Q3)As shown in the table below, our method introduces a moderate increase in computation compared to the MT baseline. This overhead originates almost entirely from the frozen forward pass of SAM-Med3D, while the proposed prototype module adds negligible parameters and cost. | Method | Params(M) | FLOPs (G) |Train (s/iter) |Infer (s/vol) | | MT | 5.88 | 122.28 |0.51 |1.32| | Ours | 8.46 | 227.79 |0.91 |1.89|

5.Discussion & Limitations (R3-Q1)Our method and MONAI-Label/nnInteractive aim to cut annotation cost with complementary designs. The latter adopts interactive prompt refinement with human annotators, while our semi-supervised method uses fixed labeled/unlabeled data, needs no input after initial labeling.Notably, our frozen SAM-Med3D prior is conceptually similar to nnInteractive’s foundation model priors. (R3-Q4)If SAM-Med3D fails under severe modality mismatch, its predictions degrade but do not misguide the student:EMA teacher remains the primary supervision, learnable prototype module corrects conflicting signals via target-domain semantic residual fusion, SAM distillation activates only after warm-up, and λ2 can be reduced.

6.Evaluation Rigor & Dataset Choice (R1-Q3a,b,c; R2-Q3)All methods use identical labeled/unlabeled splits. Datasets test complementary challenges:LA—standard SSL; Pancreas-CT—large anatomical variability, low contrast, small targets, testing prototype guidance under label scarcity;BraTS19—heterogeneous tumors with irregular boundaries, evaluating boundary robustness.Means and SDs from 100 bootstrap resamples (e.g., BraTS19, 20% labels: Dice 87.71±1.19, Jaccard 78.57±1.52, ASD 1.20±0.16, HD95 4.97±0.64); full SDs and significance tests will be added in revision.



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.

    N/A

  • 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 dual-teacher distillation framework is recognized as a technically sound and empirically robust solution for semi-supervised 3D medical image segmentation. While initial concerns were raised regarding notation ambiguities, computational overhead, and the generalizability of the external teacher, consensus post-rebuttal highlights that the authors have successfully resolved these queries through detailed mathematical corrections, efficiency metrics, and comprehensive clarifications.



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.

    All reviewers raised positive comments for this work after the rebuttal. This work is a bit engineering-oriented and has validated the method’s effectiveness via serveral widely-used datasets. The performance is good, and the authors should address the reviewer’s suggestions in the camera-ready version. A final acceptance recommendation is given.



Meta-review #3

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

    Reject

  • Please justify your recommendation.

    Despite positive reviewer updates, I recommend rejection. The method is evaluated only on binary single-organ/lesion segmentation and substantially lacks comparison to highly relevant SAM-based SSL/barely supervised SOTA baselines such as SemiSAM+ (MedIA’25) and GM-ABS (TMI’25), which also target extremely limited labels. The current SSL baselines are also not SOTA. Reported gains appear very marginal under an incomplete baseline set. Given the foundation-model SSL framing, these omissions substantially weaken the claimed novelty and performance advantage.



back to top