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

Abstract

Source-Free Domain Adaptation (SFDA) has emerged as a crucial technique for deploying medical image segmentation models across domains, particularly when source data is inaccessible due to privacy regulations. While standard SFDA methods based on entropy minimization self-supervised learning (EMSSL) have shown promise in natural images, they can be unreliable in echocardiography due to the modality’s inherent speckle noise and artifacts. In this work, we observe that blindly minimizing entropy on noisy ultrasound data can induce optimization instability and increase the risk of catastrophic drift, in which the network becomes overconfident in artifacts rather than anatomical structures. To address this, we propose the Anatomy-Texture Aware Safe Gradient Guidance Source-Free Domain Adaptation (ATASGG-SFDA) framework tailored for robust echocardiography video segmentation. Our method incorporates two clinically grounded mechanisms: (1) an Anatomy-Texture Decomposition strategy that leverages the frozen source model as a coarse anatomical anchor while adapting to target-specific textures; and (2) a Safe Gradient Guidance mechanism that acts as a safety gate, dynamically suppressing EMSSL gradients when they conflict with anatomy-texture pseudo-label supervision learning (ATPSL). This encourages adaptation to be guided by structural fidelity rather than noise overfitting. Experiments on three multi-center echocardiography datasets with significant domain shifts (cross-device and cross-population) demonstrate that ATASGG-SFDA reduces this instability and improves over compared SFDA baselines.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1509_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/lvmarch/ATASGG-SFDA

Link to the Dataset(s)

CAMUS dataset: https://www.creatis.insa-lyon.fr/Challenge/camus/ EchoNet-Dynamic dataset: https://echonet.github.io/dynamic/ EchoNet-Pediatric dataset: https://echonet.github.io/pediatric/

BibTex

@InProceedings{LvJin_AnatomyTexture_MICCAI2026,
        author = { Lv, Jinrong AND Gong, Xun AND Cheng, Ning AND Li, Zhaohuan AND Jiang, Weili},
        title = { { Anatomy-Texture Aware Safe Gradient Guidance for Source-Free Domain Adaptive Echocardiography Video 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

    The paper proposes a novel source-free domain adaptation (SFDA) framework for echocardiography video segmentation that explicitly addresses model collapse caused by noisy self-supervised signals. Its main contribution is the introduction of a safe gradient guidance mechanism, which dynamically modulates entropy minimization based on the geometric relationship between gradients, ensuring that adaptation remains consistent with reliable anatomical priors. In addition, the paper presents an anatomy-texture decomposition strategy that separates domain-invariant structural information from domain-specific appearance, enabling more robust and clinically meaningful adaptation. These components are integrated into a unified framework that improves stability and performance under challenging cross-domain ultrasound conditions.

  • 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.A key strength of this work is the introduction of a safe gradient guidance mechanism, which explicitly models the interaction between optimization signals. By leveraging the geometric relationship (e. g. , cosine similarity) between gradients from entropy minimization and anatomy-guided supervision, the method dynamically suppresses harmful updates that may lead to model collapse. This gradient-level control is a relatively novel perspective in SFDA and provides a more principled way to stabilize training compared to standard self-training or entropy minimization approaches. 2.The proposed anatomy-texture decomposition separates domain-invariant structural priors from domain-specific appearance variations. This design provides a clear conceptual framework for what should be preserved versus adapted during domain shift, which is particularly meaningful in medical imaging and could be extended to other modalities or tasks. 3.The method is evaluated across multiple datasets and domain shifts, including cross-vendor and cross-population scenarios, which are highly relevant in clinical deployment. This strengthens the empirical validity of the approach compared to studies limited to a single dataset or synthetic domain shifts.

  • 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.While the proposed safe gradient guidance mechanism is interesting, its conceptual novelty is somewhat unclear. The idea of resolving conflicting optimization signals based on gradient relationships is related to existing directions such as gradient conflict mitigation in multi-task learning (e. g. , PCGrad) and constrained optimization strategies. In addition, the use of entropy minimization as a self-supervised signal follows prior SFDA and test-time adaptation methods such as Tent and SOTTA. The paper would benefit from a clearer positioning of how the proposed mechanism differs fundamentally from these prior approaches beyond the specific application scenario. 2.Although experiments are conducted on multiple datasets, all are within the echocardiography domain. The method is designed with strong modality-specific assumptions (e. g. , anatomy-texture decomposition tailored to ultrasound characteristics), and its applicability to other imaging modalities such as CT or MRI remains unclear. Additional validation or discussion on generalizability would strengthen the contribution. 3.While the method achieves state-of-the-art performance, the improvements over recent baselines are relatively modest in some metrics. Given the added methodological complexity, it would be helpful to further analyze whether the gains are consistent across all cases (e. g. , statistical analysis).

  • 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • 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?

    Please see the strength

  • 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

    • Anatomy-Texture Decomposition (ATDS): Proposes a pseudo-labeling strategy that explicitly separates anatomical structure and texture, enabling more reliable supervision

    • Safe Gradient Guidance Mechanism (SGGM): Identifies the risk of entropy minimization in ultrasound domains and introduces a gradient-level filtering mechanism that suppresses entropy gradients when they conflict with anatomical priors.

    • State-of-the-Art Performance: Demonstrates consistent performance improvements over existing methos under both cross-vendor and cross-population settings.

  • 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.

    • Clear problem formulation: The paper clearly identifies the challenges of data accessibility and domain shift in medical imaging. In particular, it provides a well-motivated analysis of why directly applying conventional SFDA methods can cause problems, supported by intuitive gradient direction schematics.

    • Intuitive and well-aligned solution: The proposed gradient alignment-based filtering mechanism is effective. It is well-motivated by the problem setting and provides a principled way to prevent harmful updates caused by noisy signals.

    • Strong validation with SOTA performance: The method is evaluated under both equipment heterogeneity and population variation scenarios, demonstrating consistent improvements over existing methods without using target labels. Furthermore, comprehensive ablation studies support the effectiveness of the proposed components.

  • 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.

    None

  • 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • 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?

    This paper presents a well-motivated and practically effective framework for source-free domain adaptation in echocardiography segmentation. It addresses the model collapse issue of entropy minimization-based SFDA in ultrasound domains. The main reasons for assigning a score of 5 are as follows:

    • Clear and intuitive problem analysis: The limitations of conventional SSL-based methods are convincingly illustrated through gradient update schematics.
    • Clinically grounded design: The decomposition of invariant anatomical structure and variant texture is well incorporated into ATDS and SGGM, leading to effective adaptation.
    • Strong empirical performance and practical relevance: The method achieves state-of-the-art results under both cross-vendor and cross-population settings, demonstrating robustness in realistic domain shift scenarios.
    • Comprehensive ablation study: The contributions of each component are systematically validated.
    • Reproducibility: The planned code release further enhances the reproducibility of the work.
  • 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 paper presents a clinically relevant framework for SFDA in echocardiography segmentation, with effective results under domain variations.

    Although some components are based on existing techniques, I believe the paper integrates and adapts these ideas appropriately to the echocardiography SFDA problem with gradient-level safety control and demonstrates practical improvements for important medical imaging challenges such as limited data accessibility and domain shifts. The planned code release also improves the reproducibility and potential impact of the work. Therefore, I maintain my recommendation of Accept.



Review #3

  • Please describe the contribution of the paper

    This paper proposes ATASGG-SFDA, a safety-aware source-free domain adaptation framework for echocardiography segmentation that addresses the model collapse problem caused by blindly applying entropy minimization on noisy ultrasound data. The framework consists of two key components: (1) the Anatomy-Texture Decomposition Strategy (ATDS), which separates pseudo-label supervision into a frozen anatomical anchor preserving cardiac shape priors and a momentum-updated texture adapter capturing target-domain appearance via K-Means clustering; and (2) the Safe Gradient Guidance Mechanism (SGGM), which dynamically suppresses entropy minimization gradients when they geometrically conflict with the anatomical supervision gradient, ensuring optimization remains within a structurally safe subspace. Evaluated on two challenging cross-domain scenarios—cross-vendor and cross-population—the method consistently outperforms existing SFDA baselines, with particularly strong gains in boundary precision metrics (ASSD and HD95).

  • 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 Safe Gradient Guidance Mechanism (SGGM) is well-motivated and theoretically grounded. The use of cosine similarity between the anatomical prior gradient and the entropy minimization gradient to detect geometric conflicts, combined with a dynamic scalar gate α, provides an elegant safety constraint. Proposition 1 offers a formal guarantee via first-order Taylor expansion that the anatomical loss is non-increasing under the gating mechanism. The loss dynamics visualized in Fig. 4 provide convincing empirical evidence of SGGM’s stabilizing effect on training convergence.

    2.The experimental scenarios carry clear clinical relevance. The two adaptation pathways—cross-vendor (CAMUS→EchoNet-Dynamic) and cross-population (CAMUS→EchoNet-Pediatric)—represent the most common domain shift types encountered in clinical deployment. The substantial improvements in boundary-sensitive metrics (HD95, ASSD) are directly linked to the accuracy of functional indices such as ejection fraction, lending the work tangible translational 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 paper posits entropy minimization-induced model collapse as a core motivation, yet this claim lacks sufficient empirical substantiation. In Fig. 4, the curve for EMSSL (entropy minimization only) exhibits greater fluctuation but still follows an overall downward trend, which does not convincingly demonstrate model collapse. Moreover, the ablation study (Table 3) does not include an entropy-minimization-only baseline. The paper provides neither qualitative visualization of degraded segmentation outputs nor quantitative characterization of when and how frequently collapse occurs, rendering the foundational motivation insufficiently supported.

    2.The core assumptions underlying the method lack sufficient empirical validation. The Anatomical Anchor relies on the premise that the frozen source model retains structurally reliable predictions on the target domain despite significant domain shift; however, only source-domain performance is reported (Table 1, DSC 94.03%), with no quantitative evidence (e.g., topological correctness rate, shape error analysis) of structural prediction reliability on the target domains. This anchor may itself introduce systematic bias. Furthermore, the Texture Adapter generates pseudo-labels via K-Means (K=2) clustering on deep feature maps, yet the sensitivity to the choice of K is not analyzed, nor are the clustering results visualized to verify whether the method truly captures texture-level information rather than high-level semantics. Additionally, SGGM operates as a single global scalar gate applied uniformly across all network layers, which is overly coarse—shallow layers encoding texture features and deep layers encoding semantic features may have fundamentally different requirements for entropy-driven updates. The computational overhead of computing full-parameter gradient inner products is also left undiscussed.

    3.The distinction between ATDS components and existing techniques is insufficiently articulated. Both the frozen source model as an anchor and the Mean-Teacher framework as an adapter are well-established techniques. The paper does not adequately clarify how its pseudo-label generation strategy fundamentally differs from prior methods such as POEM, leaving the incremental novelty unclear.

    4.The comparative evaluation is incomplete. First, the paper lacks brief descriptions of the core ideas behind all compared methods (UPL, IAPC, AIF, POEM), making it difficult for readers to understand the sources of performance differences. Second, the comparison scope is limited to source-free adaptation methods, omitting essential reference baselines—such as supervised fine-tuning on the target domain as an oracle upper bound, and domain randomization via data augmentation as an alternative adaptation strategy—needed to comprehensively assess the relative merits of the proposed approach.

    5.The characterization of the two target domains is overly vague. The paper summarizes the domain shifts merely as “Cross-Vendor” and “Cross-Population” without specifying concrete device models, age ranges for the pediatric cohort, or other critical clinical metadata. Moreover, no per-domain image examples or feature distribution analyses (e.g., pixel intensity histograms, t-SNE visualizations of feature spaces) are provided, hindering the reader’s understanding of the specific nature and severity of the domain shifts.

  • 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • 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 addresses source-free domain adaptation for echocardiography segmentation and proposes a framework integrating Anatomy-Texture Decomposition (ATDS) with Safe Gradient Guidance (SGGM). The research problem carries clear clinical significance, the theoretical derivation of SGGM is solid, and the dual cross-vendor/cross-population evaluation is well-designed. However, several concerns collectively place this work marginally below the acceptance threshold. The paper invokes entropy minimization-induced model collapse as a core motivation, yet the experimental evidence does not convincingly substantiate this claim. The key assumption that the frozen source model retains structurally reliable predictions on target domains lacks quantitative validation under domain shift. The incremental novelty over existing techniques is not sufficiently delineated. The comparative evaluation is narrow in scope, with baseline methods inadequately described, and the target domain characterization remains superficial. Beyond methodological concerns, the writing quality requires improvement: justifications for several critical design choices are insufficient, and the logical coherence between method description and experimental analysis could be strengthened, which diminishes overall rigor and persuasiveness. That said, the framework is complete, the research direction is sound, and the work demonstrates a reasonable degree of novelty.

  • 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.

    N/A



Author Feedback

We appreciate the reviewers’ constructive comments. They recognized our method as “a relatively novel perspective”(R1), “a reasonable degree of novelty”(R3), “well-motivated and practically effective framework”(R2), and “clear clinical relevance”(R3). Here we address their main concerns:

1.Clarifications of novelty (R1,R3). SGGM differs from PCGrad-style multi-task gradient surgery because it does not symmetrically modify task gradients; it treats ATDS as an anatomical safety reference and gates only the self-supervised entropy branch. It also differs from Tent/SOTTA-type entropy minimization in that it uses entropy only when its gradient is geometrically compatible with anatomical supervision (Sec. 2.3). Compared with POEM and other SFDA baselines, our novelty lies in combining anatomy-texture pseudo-supervision with gradient-level safety control for LV segmentation. We will further clarify these distinctions in the final version.

2.Domain characterization and generalization (R1,R3).

  • Our method is specifically tailored for echocardiography. While theoretically extensible to CT and MRI, where anatomy is relatively fixed, this remains unvalidated and is left for future work. (R1)
  • We appreciate the R3’s suggestion. Due to page constraints, Sec. 3.1 provides only essential domain information alongside the respective citations. Detailed clinical metadata and image examples are available in the cited literature, and we will expand these descriptions in the final version. (R3)

3.Comparative evaluation and baselines (R1,R3).

  • In the cross-population setting, our improvement over POEM is moderate but consistent across all four metrics. In the cross-vendor setting, the gains are larger, particularly for boundary accuracy, which is important for LV function assessment. (R1)
  • We agree that a brief description of the baseline methods enhances readability and will add this to the manuscript. Since our scope is strictly confined to SFDA, evaluating supervised target fine-tuning and domain randomization—while beneficial for completeness—is not strictly necessary. We will include a more comprehensive comparison in the extended version. (R2)

4.Motivation and evidence for collapse risk (R3). We agree that “model collapse” in Fig. 4 may be too strong. Our point is not that EMSSL always degrades the final loss, but that unconstrained entropy updates in echocardiography can conflict with anatomical supervision, leading to instability and potential forgetting of anatomical features. Since entropy minimization blindly increases confidence, it may fit target-domain noise rather than domain-invariant anatomy. In Fig. 4, EMSSL decreases while ATPSL fluctuates sharply, indicating such instability; Sec. 2.1 and Fig. 1 explain this conflict. SGGM stabilizes ATPSL, whereas unconstrained embedded-mask self-supervision does not. Table 3 further shows that fixed-weight entropy brings limited improvement over ATDS and underperforms ATDS + SGGM (DSC: 79.87 vs. 82.74; HD95: 6.30 vs. 4.38). We will revise this as “collapse/catastrophic drift risk” and clarify that Fig. 4 indicates instability, not inevitable collapse.

5.Validity of the ATDS assumption (R3). Table 2 shows that the frozen source model achieves non-random performance on two target datasets (DSC: 73.77/75.63), indicating useful LV structural priors, though it is not directly deployable. ATDS further includes an EMA-based texture adapter (Sec. 2.2); Table 3 shows that removing the texture loss substantially worsens HD95 (10.10 vs. 6.02), confirming the need for target-domain texture adaptation. K=2 follows the binary LV/background setting rather than arbitrary tuning. The global scalar gate is intentional, as Proposition 1 guarantees that the anatomical loss will not increase; hierarchical gating would require additional validation beyond this submission. Since the gradient dot product is first-order, SGGM avoids second-order optimization, keeping the computational cost acceptable.



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.

    This paper proposes a source-free domain adaptation framework for echocardiography segmentation combining anatomy-texture decomposition with a safe gradient guidance mechanism. Reviewers find the problem clinically relevant and the gradient-level safety design well motivated. They also appreciate the consistent gains under cross-vendor and cross-population settings. However, concerns are raised about insufficient evidence for the claimed model collapse, unclear novelty, and a limited comparative evaluation. Given these mixed assessments, the paper is invited to rebuttal phase for the authors to address the concerns.

  • 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 accepted the paper independent of rebuttal, praising the principled safe gradient guidance mechanism, clinically relevant evaluation, and strong empirical results. The third reviewer raised concerns about insufficient motivation of model collapse and incremental novelty, but moved to accept after the rebuttal adequately addressed these issues. The overall consensus across reviewers supports 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.

    I recommend accept: post-rebuttal unanimous Accept, with anatomy-texture decomposition + safe gradient guidance delivering consistent gains on echocardiography SFDA across three multi-center datasets.



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.

    All three reviewers acknowledged the paper’s clear motivation, problem definition, method design, and validation. The major concerns centered on the evidence for entropy minimization‑induced model collapse, the lack of validation for the STDS assumption, and the unclear incremental novelty relative to existing methods. The authors have largely addressed these issues in their rebuttal.



back to top