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

Abstract

Limited labeled data and domain shifts present significant challenges for accurate medical image segmentation. Semi-supervised learning (SSL) and unsupervised domain adaptation (UDA) methods address these challenges individually. Existing SSL methods do not perform well in UDA scenarios, and vice versa. We observe that excelling in SSL requires effective learning from limited labeled data while avoiding overfitting, whereas in UDA, the domain gap must be effectively reduced. To design a novel unified framework that tackles both the scarcity of labeled data and domain shift, it is essential to address both objectives. To accomplish this, we introduce Wavelet Frequency Exchange (WFE), which decomposes encoder features into low and high-frequency components and exchanges high-frequency features between labeled and unlabeled data. WFE provides two key benefits: it disrupts overfitting by preventing the model from memorizing details from limited labeled data in SSL, and it reduces the domain gap in UDA. To improve the representation of exchanged features, we propose a Learnable Parametric Feature Network (LPFN), which includes downsampling and upsampling blocks. These blocks include Parametric Spline (PS) layers, which map the relationships between the exchanged features using a spline function. Evaluations on two publicly available medical datasets demonstrate the effectiveness of our method.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/3573_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{KumSur_Addressing_MICCAI2025,
        author = { Kumari, Suruchi and Singh, Pravendra},
        title = { { Addressing Label Scarcity and Domain Shift in Medical Image Segmentation } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15966},
        month = {September},
        page = {34 -- 44}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This paper proposes a co-training-based framework equipped with wavelet frequency mixing for semi-supervised learning (SSL) and unsupervised domain adaptation (UDA). Experiments were conducted on the LA and ACDC datasets for SSL and the MMWHS dataset for UDA.

  • 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.

    Overall, the paper is easy to follow. The idea of mixing high-frequency wavelet components for co-training is new, but the motivation behind using wavelet transformation and its subsequent operations remains unclear.

  • 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.

    The motivation remains unclear, particularly regarding the authors’ choice of wavelet transformation and the subsequent LPFN.

    Also, the paper would benefit from a more focused discussion of the specific limitations in prior work, rather than the general challenges, e.g., overfitting and domain shifts, to strengthen the contribution of this paper.

    The novelty of the method is somewhat weakened by the existing methods [R1, R2, R3], which performs frequency mixup for UDA, DG, or, SSL, respectively.

    There are some grammatical errors, typos, as well as unclear expressions. A thorough proofreading is recommended to improve readability and clarity.

    [R1] Yang, Yanchao, and Stefano Soatto. “Fda: Fourier domain adaptation for semantic segmentation.” Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2020.

    [R2] Xu, Qinwei, et al. “A fourier-based framework for domain generalization.” Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2021.

    [R3] Zhou, Yanfeng, et al. “Xnet: Wavelet-based low and high frequency fusion networks for fully-and semi-supervised semantic segmentation of biomedical images.” Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023.

  • 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.

  • 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.

    (3) Weak Reject — could be rejected, dependent on rebuttal

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

    Please refer to the weaknesses mentioned above.

  • Reviewer confidence

    Very confident (4)

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

    Reject

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

    About the choice of Wavelet transformation: performance improvement is a post-hoc test and should be considered only one of factors in selecting a technique to address the underlying problems. There exist many off-the-shelf techniques that could also be suitable for the tasks investigated in this paper, such as random convolution and style transfer, etc. “Does the quality of pseudo labels lead to overfitting? Wavelet transformation can enhance the quality of pseudo labels when applied to a co-training framework? Is Wavelet transformation effective in mitigating specific components of domain shifts?” These are important questions that warrant further investigation.

    Minor: In addition, regarding prior work, there are many studies aimed at addressing both label scarcity and domain shift simultaneously, e.g., the line of research on semi-supervised domain generalization (Semi-DG). The authors might have noted that reference [22], i.e., A&D, includes an experiment on Semi-DG.



Review #2

  • Please describe the contribution of the paper

    This paper addresses two key challenges in medical image segmentation—label scarcity and domain shift—by proposing a unified framework that combines semi-supervised learning (SSL) with unsupervised domain adaptation (UDA). The method introduces Wavelet Frequency Exchange (WFE), which decomposes encoder features into high- and low-frequency components using discrete wavelet transform (DWT) and exchanges the high-frequency information to reduce overfitting on limited labeled data and mitigate source-target domain discrepancies. Additionally, a Learnable Parametric Feature Network (LPFN) is designed using learnable spline activation functions in the upsampling and downsampling modules to better encode and reconstruct the exchanged features. Evaluations on benchmark datasets including LA, ACDC, and MMWHS show that the proposed approach consistently outperforms state-of-the-art methods in both SSL and UDA tasks, with ablation studies confirming the effectiveness of its components.

  • 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. Methodologically, the exchange of wavelet high-frequency features between labeled and unlabeled samples addresses both overfitting and domain shift, offering a clever frequency-domain intervention that is both theoretically novel and practically feasible.
    2. The proposed LPFN module, based on parametric B-splines and enhanced with residual activation functions and exponential moving average optimization, effectively models nonlinear high-order feature relationships, holding significant research value for medical image segmentation.
    3. In terms of experiments, the method is validated on multiple datasets (LA, ACDC, MMWHS), covering both SSL and UDA tasks.
  • 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 description of the research background and current developments is insufficient. For example, in the first paragraph of the introduction, the sentence “Several SSL techniques have been developed…” lacks specific discussions of these techniques. Additionally, in the second paragraph, only one existing technique that integrates SSL and UDA is mentioned, which raises concerns about the authors’ depth of understanding regarding the studied topic. 2.The paper does not specify what metrics are used to measure model diversity. Simply applying different augmentations is not rigorous enough to justify diversity. 3.The choice of wavelet basis functions significantly impacts the selection of high- and low-frequency features, and this aspect warrants a more detailed discussion and quantitative analysis. 4.The complexity and interpretability of the spline functions in the LPFN layer are not thoroughly discussed. 5.The experiments lack explanations of some abbreviations (e.g., AA, LAC), which may confuse readers. 6.It is unclear how the quality of the images generated after model transfer in the UDA experiments is evaluated. 7.The UDA evaluation only involves modality transfer between MR and CT, which is insufficient to demonstrate the generalizability of the proposed method across domains. 8.There are also minor formatting and layout issues throughout the paper, such as Table 1 being presented too early, and citation failures in the experimental section (e.g., “Table ??”).

  • 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.

  • 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?

    The proposed method demonstrates a good level of innovation; however, the validation, particularly in the UDA aspect, is insufficient. Moreover, the paper lacks a comprehensive and in-depth description of the research motivation and current state of the field. The abstract also requires further improvement to clearly convey the key contributions.

  • 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

    This paper conducts relevant research work in the fields of semi-supervised learning and unsupervised domain adaptation, and proposes a unified framework aiming to simultaneously address the scarcity of labeled data and the problem of domain shift. Specifically, the paper first introduces Wavelet Frequency Exchange. By exchanging the high-frequency features of labeled and unlabeled data, the overfitting phenomenon of labeled data is avoided, and the domain difference problem is alleviated to a certain extent. Subsequently, the paper constructs a learnable parametric feature network. With the help of spline functions, the relationships between the exchanged features are mapped, further optimizing the representation of the exchanged features and improving the representation of the exchanged features.

  • 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.

    From the perspective of the overall structure, the paper has a clear logic. The focused problems have certain practical significance, and the proposed methods are somewhat innovative. Moreover, a series of experiments are designed to fully demonstrate the feasibility and effectiveness of the methods.

  • 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 the paper, there are cases where some formulas are not numbered, and the arrangement and layout of the formulas are rather chaotic, lacking clear organization. On page 7, the text “in Tables 1 and ??”appears.
    2. On page 6 of the paper, it is mentioned that “The symbol 𝐾 represents the number of LPFN layers.”However, according to Figure 2, 𝐾 seems to be the number of PS Layers. Shouldn’t “the output of the PatchEmbed operation 𝑧𝑑 is added to the output of the combined LPFN and DepthConv outputs” be “the combined PS Layer and DepthConv outputs”? LPFN seems to be a combination of LPFN-DB and LPFN-UB. Formula (7) is inconsistent with the figure. The input of the LN layer in Figure 2 should be similar to Formula (5). It is recommended that the author re-sort out and check the content of Section 2.5.
    3. In Section 2.6 of the paper, the spline function seems to only introduce relevant knowledge and supplement Section 2.5, without reflecting the author’s innovative work and unique insights in this area. Is it necessary to have this as a separate section?
    4. The SSL comparison experiment in the paper lacks visual results, and the analysis of the experimental results is insufficient.
    5. On page 7, the paper states that “We conducted experiments on two semi - supervised benchmark datasets, namely Left Atrium [26] and ACDC [2]…”, but in fact, only the LA dataset is used.
    6. What is the impact of introducing LPFN on the number of model parameters?
  • 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.

  • 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?

    The overall structure of the paper has a certain logic, but the expression of some contents is rather chaotic, and the elaboration of the methods is ambiguous. Although some experiments have been carried out, there are still some deficiencies. In particular, for the SSL scenario, the explanation is insufficient.

  • 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 sincerely thank the reviewers for their positive feedback: novelty and practical value [R1, R2, R3], clear structure [R2, R3], unified SSL-UDA framework [R1, R2], LPFN for modeling complex feature relationships [R1, R2], and strong experimental validation across multiple datasets [R1, R2, R3].

R1W1, R3W2: Prior Work: We have cited and compared relevant works on individual SSL and UDA methods. However, due to space limits, we could not elaborate on each in detail. Since our work focuses on a unified SSL-UDA framework, we discussed the only well-established technique known at submission time that integrates both.

R1W2: We apologize for not clearly stating this in the manuscript. One model receives a standard augmented image, while the other is given a CutMix-augmented version instead of conventional augmentations. This strategy is well-known and has been used in prior work such as CPS [5]. [Ref.] refers to the main paper citations.

R1W3, R3W1:Choice of wavelet transform: Following common practices in image analysis, we initially experimented with both the Haar wavelet and the Fourier transform. As the Haar wavelet yielded better performance, we proceeded with it for our method.

R1W4, R2W3: The PS layer, instead of directly predicting per-pixel labels, learns a neural function that maps local coordinates to geometric structure. Section 2.6 explains the process, and we will refine the explanation further in the revision.

R1W7: Our method addresses both UDA and SSL. We conduct three experiments in total—one SSL setting and two modality transfer tasks in UDA.

R2W2: We apologize for the mistake. Due to a typo, we mistakenly referred to the PS layers as LPFN layers. Thank you for pointing this out. You are correct in assessing that it should be “the combined PS layer and DepthConv outputs.” In Equation~(6), we provide ( z_d’ ) (the output of the LPFN-DB block) to the \texttt{BlockUp} operation, and the resulting output is ( z_e’ ), which is then passed into Equation~(7) to obtain the final output. Therefore, the final output is a combination of the LPFN-DB and LPFN-UB blocks. The input to the LayerNorm (LN) layer remains the same as in Equation~(5), since the output of the \texttt{PatchEmbed} operation ( z_d ) is added to the combined output of the PS layer and DepthConv, i.e., ( z_d’ ), which is what we indicated in Equation~(5).

R2W5: We apologize for the oversight. Initially, we included experiments on the ACDC dataset; however, due to the page limit, we had to remove those results from the manuscript. Unfortunately, we missed updating this particular sentence accordingly.

R2W6: We have provided the model parameter details in the supplementary material.

R3W1: Motivation: We apologize for the confusion. However, we have clearly stated the motivation in Section 1 of the Introduction (paragraphs 2, 3, 4, and 5). Additionally, the motivation for using LPFN is clearly explained in the Introduction. Table 3 clearly demonstrates the improvement achieved by using LPFN blocks compared to standard convolutional layers.

R3W2: In SSL, overfitting to limited labeled data is the primary challenge, while domain shift is the main challenge in UDA. SSL methods typically perform poorly on UDA tasks, and vice versa. When combining SSL and UDA in a unified framework, these challenges become even more critical, as a single approach must effectively address both. Thus, designing a unified SSL-UDA framework is itself a significant challenge.

R3W3: Novelty: While prior work has explored separating features into high- and low-frequency components, we are the first to apply this in a unified SSL+UDA setting. Moreover, our Learnable Parametric Feature Network (LPFN), which incorporates downsampling and upsampling blocks with PS layers, is a novel component. The combination of these two elements together leads to significant performance gains.

R1,R2,R3:Typos: Thank you for pointing this out. We will correct it in the revised version.



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”.

    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. 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’

    This paper receives an initial review of 2WA (R1, R2) and 1WR (R3). After rebuttal, R3 remains Reject while final decisions from R1 and R2 are not explicitly provided. The main concerns include: 1) Unclear motivation for wavelet transformation choice over alternative frequency-based approaches, with R3 citing existing methods like FDA, fourier-based frameworks, and XNet that perform similar frequency manipulation, 2) Technical presentation issues including inconsistent notation (K representing different concepts), formula numbering problems, and discrepancies between text descriptions and figures (R1, R2), 3) Limited experimental scope with UDA evaluation only covering MR-CT modality transfer and missing visual results for SSL experiments (R1, R2), 4) Insufficient research background discussion and unclear justification for design choices such as model diversity metrics and LPFN complexity analysis (R1, R2). Despite R3’s continued concerns about fundamental methodology questions, R1 and R2 recognized the work’s innovation in combining SSL and UDA through wavelet frequency exchange, the novel LPFN module with learnable spline functions, and comprehensive experimental validation across multiple datasets. The authors adequately addressed most technical concerns in their rebuttal, clarifying notation issues and providing additional experimental details. I suggest a recommendation of Accept.



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’

    Please address the concerns raised by the reviewers in the camera ready, especially by R3.



back to top