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

Abstract

Deep learning-based echocardiography segmentation models often suffer performance degradation when deployed across different centers and ultrasound vendors. This gap is especially pronounced between public datasets and real-world clinical data, mainly due to differences in imaging physics, speckle statistics, and vendor-specific post-processing. Existing domain generalization methods typically aim to make models passively tolerate unseen appearance shifts, while generative adaptation may unintentionally alter anatomical structures essential for clinical diagnosis. We propose Proactive Domain Unification (PDU), an inference-time framework that maps heterogeneous target images into a fixed source-aligned domain without requiring target labels or test-time model updating. PDU employs a structure-conditioned generative unifier guided by boundary cues to normalize appearance while limiting geometric drift. To further preserve anatomical integrity, we introduce a reliability-guided frequency fusion module that retains low-frequency structural components from the raw image and injects unified high-frequency style residuals under reliability-controlled strength. On EchoNet→CAMUS, PDU consistently improves six strong segmentation architectures, achieving average gains of +6.0% Dice and -4.7 mm HD95. Across three representative architectures, it recovers 67.2–83.5% of the cross-domain performance gap relative to the in-domain upper bound. Validation on a private clinical dataset further confirms its robustness in real-world scenarios. These results position proactive input-domain unification as a promising route for medical image domain generalization. Code is available at https://github.com/PXinTao/PDU

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3642_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/PXinTao/PDU

Link to the Dataset(s)

N/A

BibTex

@InProceedings{PanXin_Proactive_MICCAI2026,
        author = { Pang, Xintao AND Yang, Jinlin AND Sun, Yue AND Gao, Zhifan AND Li, Wei AND Tan, Tao},
        title = { { Proactive Domain Unification for Robust Echocardiography Segmentation } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16881},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This paper introduces Proactive Domain Unification (PDU), an inference-time framework designed to mitigate the performance degradation of echocardiography segmentation models across different centers and vendors. The core contributions include: a boundary-constrained generative domain unifier to standardize the appearance of heterogeneous images; a reliability-guided frequency fusion module that injects unified texture details while preserving anatomical structures via wavelet transform. As a plug-and-play preprocessor, the proposed method enables robust segmentation without requiring target-domain labels or model fine-tuning.

  • 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 framework benefits in its focus on anatomical integrity, utilizing a boundary-conditioned diffusion model and strict wavelet-based low-frequency inheritance to prevent geometric drift. The experimental design is exceptionally thorough, validating the method across six diverse architectures ranging from classic CNNs to recent Mamba and KAN models. Furthermore, the model achieves significant performance gains, which highlights its immense practical value for real-world clinical deployment.

  • 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 specific mathematical formulations for the loss functions used in Stage I and Stage III are not explicitly listed in the manuscript. However, this is understandable given the strict page limitations of the conference format, especially since the authors provided appropriate citations to the original works.

  • 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.

    (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?

    The method is logically sound and the four stages are highly complementary, which demonstrate clear innovation by shifting the focus from passive domain generalization to proactive domain unification at the input level. The experimental design is exceptionally thorough, validating the method across different architectures and UDA methods. Furthermore, the paper provides rich visualization results, which significantly strengthen the interpretability of the model and confirm its ability to preserve anatomical integrity.

  • 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 #2

  • Please describe the contribution of the paper

    This paper addresses the problem of cross-center / cross-vendor domain shift in echocardiographic left ventricle (LV) segmentation. The authors argue that passive domain generalization is often insufficient under large texture and speckle-statistics shifts, while direct generative adaptation may distort clinically important anatomical structures. To address this issue, the paper proposes Proactive Domain Unification (PDU): 1.a structure-constrained generative unifier that maps target-domain images into a source-aligned appearance space; 2.HED-based boundary conditioning to reduce geometric drift during generation; 3.a BYOL-style hypersphere reliability estimator to assess the trustworthiness of the unified image; 4.a reliability-guided wavelet fusion mechanism that preserves low-frequency anatomical structure from the raw image while injecting unified high-frequency texture residuals according to the estimated reliability. Experiments are conducted on EchoNet→CAMUS and a private clinical dataset (PrivateEcho). The authors report that PDU improves performance consistently across six segmentation backbones, with average gains of +6.0% Dice and −4.7 mm HD95, and recovers 67.2%–83.5% of the cross-domain performance gap for several representative models.

  • 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 addresses a practically important problem. Cross-vendor and cross-center domain shift is a major obstacle for deploying echocardiography segmentation systems in real clinical settings. Focusing on LV segmentation is well motivated and clinically meaningful. 2.The overall method is well structured and reasonably coherent. Instead of adapting features directly through standard DG/UDA techniques, the paper proposes to “safely unify” the input domain before feeding images into a frozen segmenter. This inference-time preprocessing perspective is clear. In particular, the combination of structure-constrained generation, reliability estimation, and wavelet-domain fusion forms a coherent pipeline. 3.The experiments cover multiple segmentation backbones, supporting model-agnostic behavior. The method is evaluated with U-Net, UNext, Swin-UNet, VMUNet, UKAN, and SegMamba, which suggests that PDU can function as a plug-and-play preprocessing module across diverse architectures. 4.The empirical performance is generally strong. On both CAMUS and PrivateEcho, most models benefit noticeably from PDU. For example, on CAMUS, SegMamba improves from 85.28 to 88.42 Dice, while on PrivateEcho, U-Net improves from 76.67 to 82.25 Dice. The paper also includes comparisons with pure generation, in-domain stability tests, and several UDA baselines. 5.The safety motivation is well taken. Rather than fully trusting the generated image, the method explicitly preserves the raw image’s low-frequency structure and only injects unified detail bands based on reliability. This is a thoughtful design choice for medical imaging scenarios where anatomical fidelity is essential.

  • 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 comparison with existing methods is not fully fair In Table 2, the paper compares PDU with several state-of-the-art UDA methods, but the manuscript explicitly states that these UDA numbers are directly cited from [29], rather than reproduced under the same protocol. This raises fairness concerns because:

    • different papers may use different data splits, preprocessing pipelines, and evaluation protocols;
    • the proposed method reports the best results among six different segmenters, whereas each prior UDA method is typically tied to a single model;
    • PDU is trained on a “PDU-unified source domain” and evaluated on “PDU-unified target images,” which is not directly equivalent to the standard UDA setting. Therefore, the current comparison should be presented more cautiously as a reference-level comparison rather than a strict like-for-like superiority claim. 2.“No target labels” does not mean “no target-domain information,” and the presentation should be more precise The paper repeatedly emphasizes that PDU is an inference-time framework requiring no target labels or fine-tuning of the segmentation network. However, at inference time the framework still processes each target image through HED-based structure extraction, generative unification, reliability estimation, and wavelet fusion. So it is not “free” test-time generalization in the usual sense. More importantly, in the experiments the segmentation models are trained on the PDU-unified source domain, rather than simply taking a raw-source-trained model and attaching PDU at test time. This means the method actually changes the training input distribution as well. In other words, this is not merely a post hoc test-time enhancement for an already trained baseline, but closer to a joint train/test input-domain redesign. This should be stated more clearly, otherwise readers may misunderstand PDU as a pure plug-in module that can be added to any already-trained model without retraining. 3.The evidence from the private dataset is still limited The private clinical dataset contains 203 annotated images from 20 patients. This is useful, but still relatively small if it is intended to support strong claims about “real-world robustness.” The paper does not provide sufficient detail on:
    • whether the dataset includes multiple centers and/or multiple vendors;
    • the annotation protocol and annotator consistency;
    • the distribution of cardiac phases or views;
    • the exact sources of domain shift relative to EchoNet and CAMUS. If the authors want to make strong claims about real-world clinical robustness, the current evidence is still somewhat limited. 4.More detailed ablations and failure-case analysis are needed The paper compares pure generation vs. PDU, and also presents in-domain stability and frequency-domain verification. However, a more complete ablation of the key design choices is still missing. For example:
    • what happens if HED-based structure conditioning is removed, while generation and fusion remain?
    • what happens if the reliability score is removed and a fixed fusion coefficient is used?
    • what happens if low-frequency locking is kept but high-frequency reliability control is removed?
    • how much do the two reliability components srgs_{rg}srg and sdists_{dist}sdist contribute individually, and how sensitive is the method to α,β,η\alpha, \beta, \etaα,β,η? Currently, the experiments mainly show that the full system works, but not yet that each design choice has been rigorously isolated and validated. 5.The evidence for “structural safety” is still indirect The paper emphasizes that PDU preserves anatomy and limits geometric drift, and supports this with spectral analysis and difference visualizations. However, these are still indirect indicators: reduced high-frequency energy, small low-frequency deviation, and residual maps concentrated in background regions. For medical imaging, more direct safety validation would be more convincing, such as:
    • deformation statistics between contours before and after processing;
    • landmark displacement analysis;
    • whether downstream clinical metrics such as EF or volume estimation improve or remain stable;
    • quantitative boundary-consistency measures before and after unification. The current evidence is promising, but not yet sufficient to fully support a strong “clinically safe” 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 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?

    This paper studies an important and practical problem: the degradation of echocardiography segmentation performance under domain shift. The proposed PDU framework is reasonably well engineered, conceptually clear, and empirically effective. The consistent improvements across multiple backbones are particularly encouraging. However, the main limitations are that:

    • the fairness of comparisons with prior methods is not fully established;
    • the evidence for anatomical safety and real-world clinical robustness could be stronger;
    • the ablation studies are still not sufficiently fine-grained.
  • 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 proposed a transfer learning framework for LV segmentation based on ultrasound. The main idea is to design a domain unification pipeline that maps the downstream inputs into a fixed source-aligned domain. The anatomical integrity is also considered. Experiments on EchoNet->CAMUS demonstrated the effectiveness of the proposed method.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
    • The clinical motivation is strong. Although we are facing the quick development of fundation models, the cross-centre gap is still severe in clinical application which requires good solutions.
    • Th overall framework is easy to follow and it is believed that the proposed method can be easily reproduced.
    • Experiment on current settings claims a strong improvement compared with previous works.
  • 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.
    • This is a very interesting topic. My major concern is that this is very related to test-time adaptation or test-time training. Although the proposed method requires no fine-tuning on pretrained models, it is sitll easy to implement the propmt-based TTA methods which perform closely to the proprosed method. I think more discussions and analysis should be presented.
    • Another concern is the experiment settings, especially only one dataset is adopted. According to my understanding, this work is not specifically designed for LA segmentation. Therefore, it is better to see more ultrasound adaptation settings.
    • It is also expected to see better models adopted in the experiments including recent foundation models.
  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

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

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    n/a

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (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?

    My major concerns are limited experiments and lack of literature reviews.

  • 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

N/A



Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    The reviewers agreed that this research had strong clinical motivation and well-structured method designs, which were validated with detailed experiments. Also, the reviewers raised several concerns about the manuscript. (1) The lack of explicit mathematical formulations for loss functions in Stage I and Stage III; (2) potential misrepresentation of no target labels as equating to no target-domain information; (3) limited evidence from a small private clinical dataset, which undermines claims of real-world robustness. Additionally, reviewers noted concerns about the limited scope of datasets and the lack of comparisons with recent advanced models like foundation models. In general, this work is high quality and easy to follow.



back to top