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

Abstract

Common prototype-based medical image few-shot segmentation (FSS) methods model foreground and background classes using class-specific prototypes. However, given the high variability of the background, a more promising direction is to focus solely on foreground modeling, treating the background as an anomaly—an approach introduced by ADNet. Yet, ADNet faces three key limitations: dependence on a single prototype per class, a focus on binary classification, and fixed thresholds that fail to adapt to patient and organ variability. To address these shortcomings, we propose the Tied Prototype Model (TPM), a principled reformulation of ADNet with tied prototype locations for foreground and background distributions. Building on its probabilistic foundation, TPM naturally extends to multiple prototypes and multi-class segmentation while effectively separating non-typical background features. Notably, both extensions lead to improved segmentation accuracy. Finally, we leverage naturally occurring class priors to define an ideal target for adaptive thresholds, boosting segmentation performance. Taken together, TPM provides a fresh perspective on prototype-based FSS for medical image segmentation. The code can be found at https://github.com/hjk92g/TPM-FSS.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/2931_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/hjk92g/TPM-FSS

Link to the Dataset(s)

ABD-MRI dataset: https://chaos.grand-challenge.org/Combined_Healthy_Abdominal_Organ_Segmentation/ ABD-CT dataset: https://www.synapse.org/Synapse:syn3193805/wiki/89480

BibTex

@InProceedings{KimHye_Tied_MICCAI2025,
        author = { Kim, Hyeongji and Hansen, Stine and Kampffmeyer, Michael},
        title = { { Tied Prototype Model for Few-Shot Medical Image Segmentation } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15962},
        month = {September},
        page = {660 -- 670}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This paper proposes TPM, reformulating ADNet with tied prototypes for multi-class segmentation in few-shot medical imaging.

  • 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.

    This paper is claimed to solve the main drawbacks of ADNet and achieve promising results in multi-class segmentation in few-shot medical imaging scenarios. The start point is interesting to me, and the solution has practical value.

  • 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 abstract lacks a description of the experimental results.

    2. Fig.1 does not effectively convey the motivation and design of the TPM method. Specifically, the differences between TPM and ADNet are not clearly illustrated. Additionally, the figures are somewhat confusing, and the axes are not clearly labeled.

    3. The first two drawbacks of ADNet mentioned in the page.2 appear somewhat similar. I’d like to hear about the disccusion.

    4. The manuscript does not adequately explain why a shared prototype center is better. A clearer explaination, better combined with Fig.1, is encouraged.

    5. How to obtain P_B and P_F in Eq.3?

    6. Theorem1 proves equivalence to ADNet in the single-prototype case. However, this alone does not necessarily demonstrate that the proposed method inherits ADNet’s advantages.

    7. The motivation and trade-offs of EM and GMM in Sec.3.2 are not clear.

    8. The manuscript introduces several complex operations but lacks an analysis of computational complexity.

    9. The experimental section does not include comparisons with PANet and PPNet.

    10. The robustness of the “Ideal threshold” in sec.3.4 is not well validated.

  • 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.

  • 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/2025/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.

    (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?

    Insufficient clarity in motivation and methodology design. Incomplete experimental comparisons. Unclear theoretical justification.

  • 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.

    The authors have basically addressed my concerns, except that R1W7 seems misunderstood by authors. My question was aimed at understanding why EM and GMM were chosen to obtain prototypes and distributions, and whether there are potentially better alternatives.



Review #2

  • Please describe the contribution of the paper

    This paper introduces the Tied Prototype Model (TPM), a new approach for medical image segmentation that improves upon existing methods like ADNet. TPM binds foreground and background prototype positions through probabilistic reconstruction while distinguishing them with different parameters to effectively separate background anomalies. The model extends to multi-prototype capability (capturing foreground diversity using Gaussian mixture models) and multi-class segmentation (enhancing representation through joint modeling of class relationships). The paper also proposes a dynamic threshold estimation method based on class priors and feature distances, addressing the adaptability issues of fixed thresholds across different patients and organs. Experiments on abdominal MRI and CT datasets demonstrate TPM’s significant performance advantages in both binary and multi-class segmentation tasks.

  • 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 Tied Prototype Model (TPM) introduces a novel approach to medical image segmentation by using shared center points with different spreads for foreground and background. This counter-intuitive method proves remarkably effective at separating complex anatomical features.
    2. TPM elegantly extends to handle multiple prototypes and multi-class segmentation, addressing key limitations in previous methods. This enables better representation of organ variations and relationships, significantly improving segmentation accuracy when dealing with multiple organs.
    3. The paper’s adaptive thresholding approach matches predicted pixel counts to actual counts, optimizing segmentation performance. Experiments on abdominal scans demonstrate TPM’s superior results with minimal architectural complexity, offering a simple yet powerful solution for few-shot medical image segmentation.
  • 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. Considering replacing ResNet-101 with a more specialized backbone like Swin-transformer that has shown better performance in segmentation.
    2. Providing explicit formulation for the EM algorithm implementation used for multiple prototype extraction.
    3. Author should compare more SoTA few-shot medical image segmentation models in experiments.
  • 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.

  • 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/2025/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 — could be accepted, dependent on rebuttal

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

    Using a tied (shared) center position based prototype p is novel in few-shot medical image segmentation field. And classification angle of prototype generation method is special.

  • 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 #3

  • Please describe the contribution of the paper

    This paper focuses on one shot image segmentation. It extends ADNet from binary segmentation to multi-class and multi-prototype segmentation mathematically with rigorous derivations. The benchmark results showed better results than baseline in two datasets on both binary and multiclass segmentation.

  • 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 derivation strongly supports the claim of extending ADNet.
    • It is one framework for both binary and multi-class few shot segmentation, surpassing ADNet for binary and ADNet++ for multi-class.
    • The evaluation includes two different modalities, MR and CT.
    • The code is released, making the results potentially reproducible
  • 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.
    • Authors mentioned T loss, but did not clarify what’s the loss, L1, L2 or any other loss functions? and what is the weight.
    • It is unclear why the authors removed T loss when training CE-T as the threshold is learned as well. There is no ablation studies or empirical conclusions.
    • As the evaluation is one shot, it is unclear if the authors have sampled all possible support/query combinations to make the evaluation unbiased towards specific choices of support and query. (this is how it was done in ADNet++).
  • 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.

  • 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/2025/en/REVIEWER-GUIDELINES.html
    • There is no report of Hausdorff distance, which is a common metric in segmentation tasks, also used in ADNet++.
    • It would be helpful if the definition of AvgEst and LinEst are defined in Method section with more details. Currently, it may not be feasible to fully understand without checking the code.
    • The title is on few-shot, but the experiments is only on one support image slice. Maybe worth consider clarifying it’s one-shot.
    • One potential limitation would be the domain of the images, all images are of abdomen.
  • 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?

    Both the mathematical derivations and strong results suggest an accept for this paper. The released code and comparison across two datasets of different modalities helped to demonstrate reproducibility.

  • 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 thank the reviewers for the feedback and will add clarifications. We are glad they liked our empirical studies(R2,R3), theory(R2), effectiveness(R3), unified model(R2,R3), and practical applicability(R1). “W” denotes “weakness” below.

*TPM: Motivation+advantage(R1) Background(BG) usually has heterogeneity and irregularity, while foreground(FG) is more homogeneous. Geometrically, this aligns with inside-outside(IO) classification, with FG points forming the limited inside region and BG points outside. Standard prototype models have class boundaries defined by Voronoi cells, making IO modelling difficult with a fixed number of prototypes.

W2: To show TPM settings, Fig.1a depicts the pdf ϕ(p,σ;r) in 1D (x-axis is location, y-axis the density function). Fig.1b–1e shows FG probabilities p(F|x) produced by various classification models, where the xy-axis shows the 2D Euclidean feature space positions, and points denote prototypes. Fig.1b shows that standard models perform directional (left vs right) and not IO classification. Instead, all TPM models(Fig.1c-e) perform IO separation of features based on distances to tied prototypes. This enables the desired FG and BG behavior. As ADNet is limited to spherical space, a direct visual comparison is not feasible. Fig.1c visualizes the ADNet equivalent(TPM-SP) in Euclidean space.

W4,W6: The core advantage of ADNet is IO separation, allowing distinction of typical from non-typical features. While standard models struggle with this, all TPM models achieve it (see Fig.1c-e), thus inheriting ADNet’s advantages. As shown in Fig.1a, given sigma_F<sigma_B with tied prototypes, the FG probability will be higher than the BG one near the prototype and lower away from it, achieving IO separation.

*Clarification: Multi-prototypes(EM,GMM), multi-class, priors(R1,R3) R1-W3: It seems confusion arises from the terms: “class” and “prototype”. A prototype is a representative feature vector of a class, and a single class can be modeled with multiple prototypes. Thus, the first drawback of ADNet concerns its reliance on a single prototype per class, limiting representational expressiveness, while the second is the classification task itself—not supporting multi-class classification. We address both by providing formulations for the multi-prototype and multi-class settings(Fig.1d and e).

R1-W7: EM is a way of estimating parameters (prototypes+weights) for the GMM and is needed as each class in the multi-prototype setting is a Gaussian mixture. Hence, EM is not an alternative/tradeoff to the GMM—it is the method used to fit it.

R3-W2: We leverage [22] but will add more details.

R1-W5: Class priors p_F (=p(F)) and p_B (=p(B)) in Eq.(3) are usually unknown. For CE-T, we thus use a non-informative prior(p_F=p_B=0.5). To improve performance, for AvgEst and LinEst, we estimate them based on training data(see Sec. 3.4).

R1-W8: For C classes, M prototypes per class, I EM iterations, then for each query image: ADNet or single-prototype TPM take O(H’W’d); EM fitting O(IH’W’Md); multi-prototype TPM inference O(H’W’Md); multi-class(ADNet++ or TPM-MC) O(H’W’Cd); ICP calculation O(HW(d+log(HW))).

R1-W10: We indirectly show the robustness of ICP through improved performance of AvgEst and LinEst.

*Limited backbone+baseline(R1,R3) RV1-W1,W9; RV3-W1,W3: We propose a novel perspective on few-shot segmentation to inspire new directions rather than focusing on SoTA performance. Thus, we consider the most relevant baseline(ADNet) using its setup. ADNet outperformed PANet and PPNet-supporting their omission (see [6]). In future work, we will add comparisons with more backbones and other SoTA.

*T loss+inference(R2) R2-W1,W2: The T loss (L_T=T_s/alpha) has the same weight as the segmentation loss. It reduces the threshold T_S, corresponding to a reduced distance threshold T_D. As AvgEst and LinEst provide more principled thresholds, we remove it in TPM.

R2-W3: We use all support-query combinations as in ADNet++.



Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.

    Reviewers have found this submission to be a novel idea, whose solution is of practical value, and its motivation properly supported. Nevertheless, reviewers have identified several major concerns that must be addressed. In particular, a key limitation is the insufficient empirical validation in terms of compared baselines, where authors only benchmark the proposed approach to one existing few-shot segmentation method (i.e., ADNet), making it hard to put this work in context with prior literature (in terms of empirical benefits). Furthermore, the motivation for several components lack clarity, and requires further explanation (e.g., why proposed method inherits ADNet’s advantages, benefits of using a shared prototype center, or the motivation and trade-offs of EM and GMM, among others). Last, other several points needs further clarification in the main paper, such as the definition of AvgEst and LinEst, or how the class priors P_B and P_F in Eq.3 are defined.

  • 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



back to top