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

Abstract

Cross-site generalization in rectal cancer MRI segmentation is challenged by severe domain shifts across different scanners and imaging protocols. While Test-Time Adaptation (TTA) mitigates this by dynamically updating models during inference, generic TTA methods often cause semantic drift and disrupt the optimized architectures of 3D medical networks. To address this under strict patient privacy constraints, we propose Tri-KA, a Tri-level Knowledge Anchoring TTA framework. Tri-KA safely distills and transmits privacy-preserving source priors across three dimensions without altering the base network structure or requiring raw source data. Specifically, (1) at the input level, an energy-calibrated Fourier frequency module aligns global image contrast to prevent initial gradient instability; (2) at the feature level, a deep-to-shallow prototype guidance mechanism freezes deep layers to anchor invariant semantic topologies while adapting shallow representations to target textures; and (3) at the output level, an anatomical KL-divergence regularization suppresses extreme macroscopic hallucinations. Extensive validation on a challenging cross-site cohort of highly anisotropic sagittal MRI demonstrates that Tri-KA effectively balances target-domain adaptation with the robust preservation of essential anatomical structures, achieving significant improvements over state-of-the-art TTA approaches. Code is available at https://github.com/WayneBo98/Tri-KA.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4192_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/WayneBo98/Tri-KA

Link to the Dataset(s)

N/A

BibTex

@InProceedings{BoWan_TriKA_MICCAI2026,
        author = { Bo, Wang AND Xue, Ting AND Pan, Leyang AND Huang, Dingfu AND Liu, Zaiyi AND Liu, Shiyuan AND Zhou, Shaohua Kevin},
        title = { { Tri-KA: Tri-level knowledge anchoring test-time adaptation for source-free cross-site MRI rectal cancer 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 proposes to use source data via three-level knowledge extraction to guide the adaptation during test-time. It utilises FFT amplitute to retain anatomical structure information. It utilises class distribution to avoid macroscopic hallucinations. It utilises deep-level features to retain isotropic features while utilises shallow-level features to learn anisotropic texture information. Each level of knowledge extraction was well supported by physical meanings.

  • 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 main strength of this paper is: It provides an original way to use source data which balances privacy preserving and utility for modal training. Each level knowledge extraction of the source data is backed up by clinical rationale.

  • 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.

    From my understanding, the class distribution is for 3 labels, i. e. , background, rectum, tumor. As the distribution is highly skewed towards background, I suspect how this impact the adaption to the target domain. More qualitative analysis would be great.

    Another curiosity would be how the deep feature anchors are like, and what features they are representing.

  • 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.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

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

    Overall, the paper develops a novel way to utilise source data for target domain adaptation task. Each derivation of source data is backed by clinical relevance. the experimental results also demonstrate the superiority of the proposed methods. It would be better if more detailed qualitative analysis are provided for each level knowledge.

  • 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

    The paper proposes Tri-KA that performs source-free domain adaptation for rectal cancer segmentation through integrating privacy-preserving source priors across three dimensions: input-level, a feature-level, and output-level.

  • 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 source-free domain adaptation for rectal cancer segmentation of MRI images that might be an underexplored segmentation target in domain adaptation. 2.The module of Energy-Calibrated Frequency Stylization proposes an energy-calibrated mixing strategy to safely blend styles.

  • 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 has limited novelty. The paper proposes three modules, which are the Fourier module to reduce input style discrepancy, the feature-level prototype module to anchor invariant semantics, and the KL divergence regularization to regularize the class ratio. However, each of the three modules is similar to existing works. 2.In the results for comparison with SOTA, each SOTA method exhibit performance very close to the baseline, which is questionable. 3.In the introduction section, the authors do not establish how Tri-KA addresses the limitations of existing TTA methods.

  • 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 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.

    (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?

    The proposed method does not have sufficient novelty for acceptance. Each compared SOTA method exhibits performance very close to the baseline, which makes the experiment results questionable.

  • 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.

    I thank the authors for their feedback, but my concerns remain largely unaddressed in two key aspects: 1.Unusually uniform results for the comparison methods and the baseline. The 6 comparison methods (published in 2020-2025) represent distinct, optimized architectures. However, the paper reports surprisingly similar Dice scores for most of them. For instance, the Dices of Rectum for 5 comparison methods and the baseline all fall within 58.5%-59.5%, and the Dices of tumor for 5 methods and the baseline all fall within 45.2%-46.2%. Consequently, most comparison methods show negligible improvements over the unadapted baseline. Accurate benchmarking is crucial for demonstrating superiority. Yet the rebuttal only discusses SHOT, failing to explain this uniformity. Moreover, Tri-KA exceeds these methods with minimal improvements (0.13% on Rectum and 1.4% on tumor), which undermines the practical significance. 2.The novelty of Tri-KA is not fully clarified. The rebuttal claims the “deep-to-shallow routing” as a non-trivial design; however, it is essentially freezing deep layers and training shallow layers. The general observation that shallow layers capture style-related features while deeper layers encode semantic features is well established, and does not constitute a substantial novelty in domain adaptation. Regarding the three proposed modules (the Fourier module to reduce input-style discrepancy, the feature-level prototype module to anchor invariant semantics and match prototypes, and the KL divergence loss to regularize class ratios), their ideas appear closely related to the following works, respectively: [1] Wang et al. “Fvp: Fourier visual prompting for source-free unsupervised domain adaptation of medical image segmentation.” IEEE Transactions on Medical Imaging, 2023.[2] Yu et al. “Source-free domain adaptation for medical image segmentation via prototype-anchored feature alignment and contrastive learning.” MICCAI, 2023.[3] Bateson, et al. “Source-relaxed domain adaptation for image segmentation.” MICCAI, 2020.For me, Tri-KA seems more like an aggregation of existing techniques. While Energy-Calibrated Frequency Stylization is the author’s contribution, its advantage over existing Fourier-based style alignment methods remains unclear, especially given that the overall difference over the baseline (no adaptation) is merely around 3%.



Review #3

  • Please describe the contribution of the paper

    This paper proposes Tri-KA, a source-free test-time adaptation (TTA) framework for cross-site rectal MRI segmentation. The key idea is to incorporate privacy-preserving source priors at three levels: input (Fourier-based frequency alignment), feature (deep-to-shallow prototype guidance), and output (distribution-level KL regularization). The framework aims to mitigate domain shift while preventing semantic drift and preserving anatomical structure. Experiments on cross-site MRI datasets demonstrate improved segmentation performance compared to existing TTA methods.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
    • The paper addresses a clinically important and realistic problem: cross-site generalization under strict privacy constraints, which is highly relevant in medical imaging.
    • The tri-level design (input, feature, output) is well-structured and provides a clear conceptual framework for stabilizing TTA.
    • The idea of freezing deep semantic layers while adapting shallow layers is well-motivated and aligns with known properties of medical CNN feature hierarchies.
    • The use of privacy-preserving priors (frequency statistics, prototypes, distribution constraints) is practically meaningful and avoids reliance on source data.
    • Experimental results show consistent improvements over multiple TTA baselines (Table 1, page 6), particularly in tumor Dice, which is clinically important.
  • 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.
    • While the tri-level framework is well-organized, each component (Fourier alignment, prototype guidance, KL regularization) builds on relatively standard ideas, making the overall novelty somewhat incremental.
    • The method involves multiple carefully designed modules and hyperparameters, which may limit robustness and reproducibility in different datasets or settings.
    • The improvement margins, although consistent, are relatively modest, especially considering the added complexity of the framework.
    • The evaluation is limited to two in-house datasets, and broader validation on public benchmarks would strengthen the generalizability claims.
    • The dependence on precomputed source priors (e.g., prototypes, distributions) may still introduce implicit reliance on source data characteristics, which could limit applicability under more extreme 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

    The paper presents a practical and well-motivated framework for source-free test-time adaptation in medical imaging. The tri-level design is intuitive and supported by experimental results. However, the contribution is somewhat incremental, and further validation on diverse datasets or analysis of robustness would strengthen the work.

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

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

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

    The paper presents a practical and well-motivated framework for source-free test-time adaptation in medical imaging. The tri-level design is intuitive and supported by experimental results. However, the contribution is somewhat incremental, and further validation on diverse datasets or analysis of robustness would strengthen the work.

  • 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.

    After carefully considering the authors’ rebuttal, I maintain my recommendation of Accept, although I consider the paper to be a weak/marginal accept rather than a strong accept.

    The rebuttal clarifies the authors’ intended contribution as a system-level source-free test-time adaptation framework rather than a collection of individually novel modules. I appreciate this clarification, as the proposed tri-level design is coherent and clinically motivated: input-level frequency alignment addresses appearance/style shift, feature-level prototype anchoring aims to reduce semantic drift, and output-level distribution regularization is intended to suppress macroscopic hallucinations. The authors also provide a clearer rationale for the deep-to-shallow adaptation strategy in anisotropic rectal MRI and explain why freezing deep layers while adapting shallow layers may help preserve semantic topology while accommodating scanner- or protocol-related texture shifts.

    I still find the application setting important. Source-free cross-site adaptation for rectal cancer MRI segmentation is clinically relevant under privacy constraints, and the method is practically meaningful. The reported improvements, particularly for tumor segmentation, support the usefulness of the proposed framework. The rebuttal also provides helpful clarification regarding the behavior of baseline TTA methods, the interpretation of the KL anatomical prior, and the clinical trade-off between rectum ASD and tumor ASD.

    However, several concerns remain. First, the novelty of the individual components remains somewhat incremental, since Fourier alignment, prototype-based guidance, and KL-based output regularization all have related precedents. The main contribution is therefore best viewed as a coherent integration and task-specific adaptation strategy rather than a fundamentally new methodological component. Second, the validation is still limited to a single in-house cross-site setting, without an external public benchmark or broader multi-center evaluation. I understand the difficulty of obtaining high-quality annotated sagittal rectal MRI data, but this still limits the strength of the generalizability claim. Third, the rebuttal provides qualitative justification for the chosen hyperparameters, but does not fully resolve the concern that several interacting parameters may affect robustness. A systematic sensitivity analysis would be important to establish that the method is not overly dependent on specific settings. Finally, the explanation of deep feature anchors is plausible but remains somewhat abstract; more direct qualitative or quantitative analysis would strengthen the interpretation.

    Overall, I believe the paper remains slightly above the acceptance threshold. Its clinical motivation, coherent tri-level design, and empirical improvements justify acceptance, but the remaining concerns regarding incremental novelty, limited validation scope, and robustness analysis should be taken into account in the final decision.



Review #4

  • Please describe the contribution of the paper

    The paper proposes Tri-KA, a tri-level knowledge anchoring framework for source-free test-time adaptation in cross-site rectal cancer MRI segmentation. Its main contribution is to inject privacy-preserving source priors at the input, feature, and output levels to improve target-domain adaptation while preserving essential anatomical structures.

  • 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.Cross-site rectal cancer MRI segmentation under source-free test-time adaptation is a highly relevant yet challenging setting, especially under privacy constraints and strong scanner/protocol shifts. The application scenario is clinically important and well motivated. 2.Rather than applying a generic TTA recipe, the paper explicitly designs adaptation at three complementary levels—input, feature, and output—to address appearance shift, semantic drift, and macroscopic hallucination, respectively. This gives the method a coherent and task-aware formulation.

  • 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 feature-level strategy relies on the assumption that deeper layers mainly encode domain-invariant semantic topology, while shallow layers primarily capture domain-specific texture variations. This is a reasonable intuition for anisotropic MRI, but it may not always hold under severe cross-site shifts. The paper would be stronger with clearer evidence showing that freezing deep layers and updating only the shallow encoder is consistently the right trade-off. 2.The input-level Fourier stylization, confidence thresholding, prototype construction, and prior regularization all depend on important design choices and hyperparameters (e.g., stylization strength, prototype clustering setting, confidence threshold, and prior weight). Since these modules interact with each other, more systematic sensitivity analysis would help establish that the observed gains are not overly dependent on specific parameter settings. 3.The KL-based anatomical prior is well motivated for suppressing extreme false positives, but it also constrains the target prediction distribution toward a source-derived population baseline. In a cross-site clinical setting, however, the true target distribution may legitimately differ from the source due to cohort composition or lesion burden. The paper would benefit from stronger analysis showing that this prior suppresses hallucinations without also suppressing genuine patient-level heterogeneity.

  • 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.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

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

    The paper addresses a practically important and challenging setting, and the proposed tri-level anchoring framework is well tailored to cross-site rectal MRI segmentation under source-free test-time adaptation. My score is slightly above the acceptance threshold because the method is coherent and empirically promising, although several key design assumptions and the robustness of some components would benefit from stronger validation.

  • Reviewer confidence

    Very confident (4)

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

    Accept

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

    The author has addressed my concerns, and if the author revises the paper as promised, I believe it will meet the acceptance criteria.



Author Feedback

We sincerely thank the AC and all reviewers for their constructive feedback and recognition of our clinical motivation. 1.System-Level Novelty & Anisotropic Adaptation (AC, R2, R3, R4) While tools like Fourier alignment and KL divergence have precedents, we clarify that Tri-KA’s contribution is a system-level paradigm tailored for 3D medical networks, rather than a simple module aggregation. Generic TTAs using unconstrained monolithic updates often suffer from semantic drift and structural collapse in low-contrast MRIs. Tri-KA mitigates this via decoupled adaptation. Crucially, our deep-to-shallow routing represents a non-trivial architectural design explicitly aligned with the physical properties of highly anisotropic rectal MRIs and asymmetric downsampling mechanisms. Addressing R4’s concern regarding this trade-off: freezing deep layers stably anchors domain-invariant isotropic semantic topologies; exclusively updating shallow layers adapts to cross-site anisotropic texture variations (e.g., slice thickness shifts) without compromising global anatomical integrity. 2.Dataset Validation & SOTA Benchmarking (AC, R2, R3) Regarding the absence of public benchmarks (R3), high-quality sagittal rectal cancer MRIs with voxel-wise annotations are exceptionally rare, particularly for multi-center TTA tasks. Thus, our cross-site cohort of 157 3D anisotropic volumes (80 source → 77 target) provides a realistic and substantial validation scale. We will publicly release our code and model weights upon acceptance. Addressing R2: generic TTAs relying on unconstrained entropy minimization frequently suffer from gradient collapse in low-contrast MRIs. Furthermore, SOTA methods exhibit varying behaviors, not just baseline matching. For instance, SHOT achieves competitive Rectum Dice (61.40%) but severely degrades Tumor Dice to 45.21% (below baseline). Tri-KA balances both. Our improvements’ reliability is corroborated by rigorous Wilcoxon tests (most p<0.001). 3.Hyperparameter Robustness & Feature Anchors (AC, R1, R4) Regarding hyperparameter sensitivity (β, τ, λ, K), we clarify that their values were derived from physical intuitions and validated for broad stability during the initial design phase, rather than being heavily tuned. The strict confidence threshold τ=0.95 enforces reliable anatomical filtering, making it robust across conservative ranges. FFT blending β=0.5 neutrally balances source style and target integrity, while the soft prior λ=2.0 functions stably without demanding precise calibration. The cluster number K=5 provides sufficient capacity to model intra-class feature heterogeneity (e.g., varying tumor core/margin textures). Regarding R1’s query, these deep prototypes represent high-dimensional, domain-invariant semantic descriptors (e.g., spatial topologies between the rectum and pelvic floor). Capturing abstract structural relationships over local intensities makes them highly reliable semantic anchors even under severe domain shifts (R4). 4.Clinical Trade-offs & Anatomical Priors (AC, R1, R4) The rectum ASD relaxation (2.98 to 3.70 mm) noted by the AC is a sub-voxel variance, falling below even the high in-plane resolution (e.g., ~0.75-0.88 mm). Clinically, accurate malignant tumor delineation is of paramount importance for cancer staging. Accepting a sub-millimeter boundary variance on the stable rectal wall to achieve significant improvements in the challenging tumor ASD (reduced to 7.88 mm) is a highly pragmatic clinical trade-off. Regarding the KL prior (R1, R4): the background-skewed distribution (R1) makes unconstrained models prone to massive false-positive predictions. The KL prior leverages this baseline to explicitly suppress organ-scale macroscopic hallucinations. Crucially, acting as a soft constraint (λ=2.0), it mitigates structural collapse while allowing prototype guidance to accommodate genuine patient-level lesion heterogeneity and natural volume variations (R4).



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.

    I recommend rebuttal, consistent with the reviewer comments. Mixed scores with two confident accepts and one Reject naturally fit the rebuttal track, and the paper has paper-grounded merit that justifies reconsideration. Tri-KA targets a realistic and underexplored clinical setting (privacy-preserving source-free TTA for cross-site rectal cancer MRI segmentation) and integrates three privacy-preserving source priors at distinct levels: an energy-calibrated Fourier input alignment, a deep-frozen / shallow-adapted prototype guidance, and a population-level KL prior on output class distribution. The deep-to-shallow gradient routing in Sec. 2.3 is a non-trivial design and the paper includes Wilcoxon significance testing throughout Tables 1–2 with most baseline comparisons reaching p<0.001.However, three concerns need clarification: per-module novelty is incremental since Fourier domain alignment, prototype-cosine supervision, and KL anatomical priors all have direct precedents (R3, R4); validation is restricted to a single 80→77 volume in-house cross-site pair with no public benchmark (R3); and four interacting hyperparameters (β=0.5, τ=0.95, λ=2.0, K=5) lack any sensitivity analysis (R4). Two paper-level issues also surface: rectum ASD actually degrades from 2.98 to 3.70 mm relative to the baseline

  • 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: Tri-KA’s tri-level source-free TTA design for cross-site rectal MRI segmentation (Fourier alignment + deep-frozen/shallow-adapted prototype guidance + KL output prior) delivers statistically significant gains (Wilcoxon p<0.001) with a non-trivial D/S layer-routing result in Table 3, supported by two Weak Accepts; camera-ready should clarify integration novelty vs FDA/prototype/KL precedents, justify the rectum ASD degradation (3.70 vs 2.98 mm), and provide β/λ sensitivity.



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.

    After considering the reviews and the authors’ rebuttal, I recommend acceptance.

    The paper addresses an important clinical problem and presents a coherent source-free adaptation framework for rectal MRI segmentation. Although the novelty of the individual components is incremental, the overall design is well motivated and the empirical results support the effectiveness of the proposed approach.

    The rebuttal adequately clarifies the main design choices and addresses most of the concerns raised during review. While questions remain regarding external validation, robustness analysis, and the degree of novelty of individual modules, I view these as limitations of scope rather than fundamental weaknesses of the work.

    Overall, I believe the paper makes a useful contribution and that its strengths outweigh its remaining limitations.

    Recommendation: Accept.



back to top