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

Abstract

The emergence of foundation models enables a training-free paradigm for medical image segmentation. While previous frameworks relied on heavy prompt engineering for the original SAM, SAM2 enables direct mask propagation from support sets, obviating manual prompt extraction. However, naïve propagation in a fixed field of view often fails to adapt to subtle anatomical variations and intensity shifts. To fully harness the sequence propagation capability of SAM2, we propose Dual-view Propagation and Prototypical Assessment for Training-free Few-shot Medical Segmentation (MedPro), which integrates three synergistic strategies. First, histogram matching aligns intensity distributions across subjects, improving cross-subject statistical consistency prior to propagation. Building upon this, a focus zoom-in mechanism generates dual-view pseudo-video sequences that incorporate both global contextual information and localized ROI details. These sequences are then processed via SAM2 for bidirectional mask propagation, providing a fine-grained perspective while mitigating spatial drift. Finally, a prototypical assessment is developed for automated semantic confirmation. By generating representative prototypes from DINOv2 features, it evaluates candidate masks via semantic similarity to reliably identify the optimal segmentation. Extensive experiments demonstrate that MedPro achieves competitive results across multiple datasets. Code is available at: https://github.com/CVL-hub/MedPro.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1552_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/CVL-hub/MedPro

Link to the Dataset(s)

N/A

BibTex

@InProceedings{ZhaJin_MedPro_MICCAI2026,
        author = { Zhang, Jingyi AND Liu, Mingyi AND Shu, Xiangbo AND Zhang, Xu-Yao AND Xie, Guo-Sen},
        title = { { MedPro: Dual-View Propagation and Prototypical Assessment for Training-Free Few-Shot Medical Segmentation } },
        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

    This manuscript presents an innovative training-free framework by integrating SAM2 and DINOv2, specifically, authors designed three modules, Pair Construction and Alignment, Dual-view Propagation and Prototypical Assessment for medical few-shot segmentation tasks, achieving the SOTA in three fss benchmarks.

  • 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.Cleverly combines SAM’s spatial segmentation with DINOv2’s semantic understanding. 2.Experiments are also quite extensive in fss benchmarks. 3.Performance is impressive.

  • 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.In this pipeline, SAM2 has to run both sequences separately, and after SAM2 does its thing, we still have to call the DINOv2 encoder for feature extraction. Both SAM2(sam2-hierasmall) and DINOv2(dinov2-base) are large models, hence, the authors should offer the average inference time and computational cost per slice in the manuscript.

    2.Unlike CT scans with absolute HUs, MRI values are relative. In MRI, if the histogram matching mess with the contrast, you might be erasing a lesion or creating a fake one.

    3.Another Histogram Matching concern, the authors align intensity based on the whole slice. If the Support slice has a lot of “air” (background) and the Query doesn’t, the intensity gets warped. This ruins the target organ’s features.

    4.Eq3,4 use the Support image’s coordinates to crop the Query image. However, patients aren’t clones. The organ in Image A is rarely in the exact same spot as Image B. If the organ shifts even a little bit, the crop misses it entirely.

    5.The authors only report Dice, however, SAM-based models often have “jagged” edges. Other metrics, like HD95, should be offered in such settings.

    6.There’s a data leakage concern. The authors follows the traditional few-shot paradigm, but things change when you bring in foundation models like SAM and DINOv2.This is no longer a ‘strict’ Setting 1.It’s highly likely these models have already ‘seen’ similar medical data during their massive pre-training.

  • 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 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.

    (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?

    Methodological novelty

  • 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

    The paper proposes MedPro, a training-free framework for few-shot medical image segmentation that improves SAM2-based propagation by combining three components: histogram matching for intensity alignment, dual-view propagation (global and zoomed-in local views) to reduce spatial drift, and a prototypical assessment module using DINOv2 features to automatically select the best segmentation mask based on semantic similarity.

  • 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 addresses important practical challenges in medical segmentation such as domain shift, spatial drift, and weak semantic understanding. It presents a fully training-free pipeline, which is highly valuable for real-world clinical deployment. The method is simple, interpretable, and well-motivated, with each component contributing meaningfully as shown in ablation studies. The dual-view propagation idea effectively balances global context and local detail. Experimental results are strong, outperforming prior training-free methods and even surpassing some supervised approaches on multiple datasets.

  • 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.
    • Limited methodological novelty: Most components are incremental and based on existing techniques. Histogram matching is a classical preprocessing method widely used in medical imaging. ROI cropping / zoom-in strategies are standard in segmentation pipelines. Prototypical similarity is not a new technique (e.g., prototypical networks in SSL-ALPNet [1], Q-Net [2], RPT [3], etc.). The closest prior work, FS-MedSAM2 [4] (2024), already uses SAM2 for training-free propagation, so MedPro mainly extends it rather than introducing a fundamentally new paradigm.
    • Lack of efficiency analysis: The method requires dual propagation (global + local) with SAM2 and additional DINOv2 feature extraction. There is no reporting of inference time, memory usage, or computational overhead, which is important for real-world deployment.
    • The paper does not provide a deeper analysis of why dual-view propagation or prototype-based selection works beyond empirical evidence. For example, there is no study of failure cases or conditions where the method breaks down.

    [1] Ouyang, C., Biffi, C., Chen, C., Kart, T., Qiu, H. and Rueckert, D., 2020, August. Self-supervision with superpixels: Training few-shot medical image segmentation without annotation. In European conference on computer vision (pp. 762-780). Cham: Springer International Publishing. [2] Shen, Q., Li, Y., Jin, J. and Liu, B., 2023, September. Q-net: Query-informed few-shot medical image segmentation. In Proceedings of SAI Intelligent Systems Conference (pp. 610-628). Cham: Springer Nature Switzerland. [3] Zhu, Y., Wang, S., Xin, T. and Zhang, H., 2023, October. Few-shot medical image segmentation via a region-enhanced prototypical transformer. In International conference on medical image computing and computer-assisted intervention (pp. 271-280). Cham: Springer Nature Switzerland. [4] Bai, Y., Yu, Q., Yun, B., Jin, D., Xia, Y. and Wang, Y., 2024.Fs-medsam2: Exploring the potential of sam2 for few-shot medical image segmentation without fine-tuning. arXiv preprint arXiv:2409.04298, 3.

  • 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.

    (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 presents a well-motivated and practically useful approach to training-free few-shot medical segmentation. The integration of dual-view propagation and prototypical assessment is simple but effective, and the method achieves strong empirical results, even outperforming some supervised baselines. The problem is relevant and the solution is clearly described, with solid ablation and qualitative analysis. However, the technical novelty is limited. Most components are adaptations of existing ideas (e.g., histogram matching, ROI cropping, prototype similarity), and the method heavily relies on powerful pretrained models (SAM2, DINOv2). The contribution is more of a system-level improvement rather than a fundamentally new method. Despite limited novelty, the paper is practically impactful, empirically strong, and relevant to the community, which justifies acceptance at a weak accept level.

  • 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 authors addressed my concern about efficiency analysis. Despite the limited novelty introduced by the use of existing techniques, I believe the method is effective and exceeds the acceptance threshold.



Review #3

  • Please describe the contribution of the paper

    This paper proposes MedPro for medical image segmentation by integrating three complementary strategies: Pair Construction and Alignment (PCA), Dual-view Propagation (DP), and Prototypical Assessment (PA). Extensive experiments on multiple datasets show that MedPro achieves competitive 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 paper is generally well written and easy to follow.

    2.Extensive experiments are conducted to validate 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.The necessity of Prototypical Assessment (PA) is not yet fully justified. It would be helpful to include an ablation study examining whether comparable performance can be achieved by using only a single-branch mask (either the normal branch or the zoom-in branch), or by simply averaging the predictions of the two branches.

    2.Since the proposed framework involves multiple backbone models and a dual-branch inference pipeline, it is recommended to provide a comparison of computational complexity with competing methods, such as in terms of parameter count, FLOPs, inference speed, or memory consumption.

    3.The paper states that the input image resolution is 256 × 256, whereas the actual input to SAM appears to be 1024 × 1024.This discrepancy should be clarified to avoid confusion.

  • 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.

    (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?

    Compared with existing work, MedPro introduces several novel perspectives for medical image segmentation, such as Dual-view Propagation.

  • Reviewer confidence

    Somewhat confident (2)

  • [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 authors’ rebuttal has resolved my concerns.



Author Feedback

Thank all reviewers for their valuable comments. Our detailed responses are below. Shared Q: Efficiency analysis (Reviewer #1,2,3): By checking original exp. records, we find MedPro achieves avg. inference time of 144.62ms per image on a single NVIDIA GTX 3090, requiring only 1795MB VRAM. Furthermore, processing entire Abd-MRI dataset takes ~2.5 min, demonstrating its capability to meet the real-time clinical requirements. Actually, since 1) rebuttal policy barring, i.e., not allowed additional experiments, and 2) SynPo has not released its source code or computational efficiency details, we shall perform additional efficiency comparisons in final version.

Reviewer #1 Q2,Q3.Histogram matching on MRI. Sorry for misleading, we’d like to demonstrate that our task specifically targets whole-organ segmentation rather than lesion-level diagnosis. As illustrated by the MRI images in Fig.3, our method aligns support slices intensity to be more consistent with the query slices without inducing any intensity warping, lesion erasure, or artifacts. In addition, HM achieves 1.73% gain in MRI, further confirming the benefits of this alignment. Q4.Organ shifts. To accommodate potential spatial shifts, we apply a generous padding (50px) to ensure the target organs are well-contained. Meanwhile, our dual-branch architecture combines an original-size branch alongside the cropped one; thus, even if an object exceeds the cropped area, the full-image branch maintains reliable segmentation. Q5.Evaluation metrics. The aliasing artifacts in SAM stem from its default low-res. mask output. To address this, the SAM2 we employed incorporates high-res. skip connections in the decoder, which have refined boundary details and mitigated the sawtooth effect to some extent. We appreciate your insightful suggestion regarding the HD95 metric, which is important in evaluating edge precision, and we will include it in our revised eval. Q6.Question on setting. After carefully referring to the original papers, neither involved medical datasets during training.

Reviewer #2 Q1.Novelty of MedPro. MedPro’s novelty stems from the synergy between dual-view propagation and PA, which tailors the SAM2 architecture for FSMIS. -FS_MedSAM2 [4] adapts SAM2 into FSMIS without fully exploiting the inherent strengths of SAM2 or addressing its potential limitations. In contrast, based on it, MedPro builds upon it by introducing dual-view propagation to mitigate organ drift at the global scale. Furthermore, PA module compensates for SAM2’s limited semantic perception ability. Meanwhile, HM ensures enhanced temporal consistency to meet SAM2’s operational requirements, while Focus Zoom-in provides the basis for the dual-view perspective. Together, we achieve 2%~5% performance gain (Tab.1) beyond [4]. -Conventional prototypical networks [1,2,3] use prototype similarity for classification. Different from them, our PA uses prototype similarity as a selector, which computes the global score rather than pixel-wise classification. Q3.Deeper analysis. Both experimental results (Tab. 1) and visualizations (Fig. 3) demonstrate effectiveness. In the revised version, we will add a discussion on challenging cases & boundary conditions, e.g. low-contrast regions and blurred organ boundaries, where medical image segmentation itself is inherently more challenging. This will further clarify the applicability and robustness of MedPro.

Reviewer #3 Q1.Necessity of PA. Sorry for misleading. Tab.3 (+Focus Zoom-in) shows the optimal weighted result of the dual-branch setup. By further incorporating our PA as a selector, the model gets 1.23% gains. We will also provide single-branch ablation studies in the future eval. Q3.Input size. We apologize for the confusion. To clarify, the images are maintained at 256×256 resolution throughout our pipeline; however, SAM2 uses an internal preprocessor to upscale them to 1024×1024.This implementation detail will be explicitly addressed in the revised version.



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.

    The three reviewers converge on a borderline recommendation (3 / 4 / 4), agreeing that the empirical results are strong but raising several concerns that warrant a rebuttal. The most consistent concern, raised by all three reviewers, is the absence of any efficiency / computational analysis. Given that the pipeline runs SAM2 in two passes (global + zoomed-in) and additionally invokes DINOv2, the authors should report inference time, memory footprint, and parameter / FLOP comparisons against the main competing methods. Two reviewers further question the methodological novelty, noting that histogram matching, ROI cropping, and prototypical similarity are all well-established techniques and that the work largely extends FS-MedSAM2.The rebuttal should clarify the conceptual contribution beyond system-level integration of existing components. A few individually raised but technically substantive points should also be addressed: The validity of histogram matching for MRI (relative intensities) and its sensitivity to background/air differences between support and query. The assumption in Eq. 3–4 that support-image coordinates can be reused to crop the query, despite inter-patient anatomical variation. Clarifying the 256×256 vs. 1024×1024 input resolution discrepancy.

  • 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 reading the rebuttal, I find the authors’ responses to be reasonable and largely satisfactory. While the reviews were split (two accepts, one reject), the single reject was only a weak reject given before the rebuttal, and that reviewer did not update or provide any post-rebuttal assessment. Having examined the rebuttal carefully myself, I agree that it adequately addresses the main concerns raised. On balance, I lean towards acceptance.



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.

    The authors have provided comprehensive and convincing clarifications in their rebuttal, successfully addressing the critical concerns raised during the review process. Specifically, the authors have clearly answered the questions regarding inference time, GPU memory footprint, the effectiveness of histogram matching on MR images, the resolution discrepancies ($256 \times 256$ vs. $1024 \times 1024$), and the geometric rationality of cross-image cropping under anatomical variations. Furthermore, the authors have committed to incorporating detailed analyses on efficiency, computational complexity, FLOPs, parameter counts, and background sensitivity into the revised manuscript. Consequentially, the reviewer has raised the evaluation from “Weak Accept” to “Accept”. Following the post-rebuttal consensus, the AC fully believes the paper meets the high standards of the venue and therefore 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.

    Two reviewers agreed on acceptance after authors’ rebuttal. One reviewer who rated weak reject pre-rebuttal did not provide final rating, but his/her concerns on computational cost, Histogram Matching on MRI, etc., have been clarified in the rebuttal. The algorithm is relatively efficient. Despite the use of existing techniques as indicated by R2, the work introduced several novel perspectives for medical image segmentation and showed strong empirical results.



back to top