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

Abstract

Cross-domain diagnosis remains a major challenge in cervical cell pathology due to pronounced domain shifts across institutions and the subtle visual differences among disease stages, which jointly impair model generalization. To address these issues, this paper proposes a two-stage framework for cross-domain cervical cell detection. In the first stage, we propose the Spatially-Continuous Unpaired Neural Schrödinger Bridge (SC-UNSB), which constructs a synthetic intermediate domain to mitigate cross-domain distribution shifts by modeling image translation as an entropy-regularized optimal transport process. In the second stage, we propose a dual-level feature alignment strategy within a knowledge distillation, which progressively aligns shallow structural features and deep semantic representations to facilitate the transfer of domain-invariant knowledge from the source to the target model. Experimental results demonstrate that the proposed method effectively mitigates domain shift and category ambiguity, improving the cross-domain detection performance.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2238_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/ZhanYiHui06/MICCAI2026

Link to the Dataset(s)

N/A

BibTex

@InProceedings{LiJin_TwoStage_MICCAI2026,
        author = { Li, Jincheng AND He, Yuzhi AND Zhan, Yihui AND Zhang, Xinmei AND Sun, Yifei AND Liu, Zelin AND Zhang, Lichi AND Shao, Minye AND Zhao, Lili},
        title = { { Two-Stage Cross-Domain Cervical Abnormality Screening with Cytopathological Image Synthesis and Knowledge Distillation } },
        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 authors proposed an approach that combines spatially coherent domain bridging with dual-level feature alignment to effectively transfer domain-invariant knowledge across institutions, advancing cross-domain diagnosis in cervical cytopathology. They introduces a spatially continuous image translation framework designed to construct an intermediate domain through dense statistical normalization, with the stated goal of mitigating boundary-induced artifacts in ultra-high-resolution cytopathological images. Furthermore, the authors present a dual-level feature alignment strategy embedded within a knowledge distillation framework, which jointly operates on shallow structural representations and high-level semantic embeddings. This is relevant for data-driven learning methods to perform consistently, regardless of the data source, particularly regarding fine-grained cell boundary details, which are especially critical in the context of cytological imaging.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
    • The proposed SC-UNSB framework introduces a principled extension to UNSB by re-parameterizing statistical moments as spatially continuous functions of pixel coordinates, rather than scalar quantities. This is a non-trivial and well-motivated design choice that directly addresses boundary-induced artifacts inherent to ultra-high-resolution cytopathological image translation.
    • The knowledge distillation framework is thoughtfully designed, with LFA and CFA targeting distinct yet complementary aspects of feature alignment — shallow structural fidelity and high-level semantic consistency, respectively. This hierarchical approach reflects a strong understanding of the challenges in cross-domain knowledge transfer.
    • SC-UNSB mitigates tiling artifacts by producing smooth, spatially coherent statistical moment fields that preserve cellular morphology and staining consistency, yielding consistent improvements across multiple image quality metrics and benefiting downstream detection performance.
    • The evaluation on two large-scale cervical cytology datasets is a notable strength, as it grounds the contribution in a practical and impactful medical imaging context. The reported performance gains of up to 26.9% mAP and 45.8% mAP50 are substantial and suggest that the proposed approach meaningfully advances the state of the art in cross-domain cytopathological detection.
  • 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 paper lacks a computational cost analysis, omitting data on training duration, inference speed, and memory requirements, which makes it difficult to assess the practical viability of the SC-UNSB framework for clinical deployment.
    • This work is also limited dataset scope and fixed evaluation direction insufficiently establish the framework’s robustness across diverse domain shifts, multiple targets, or varying tissue preparation protocols.
    • This reliance on RetinaNet as the sole detector of cervical cytopathology restricts the generalizability of the findings, as the lack of testing with alternative backbones leaves it unclear whether the observed improvements are a universal benefit of the framework or merely architecture-specific.
  • 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.

    (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?

    If the authors include a brief analysis of the computational cost of the proposed method compared with other methods, as well as integrating other cervical cytology datasets, and even datasets from other cell types, it would be relevant to highlight the article’s contribution and improve its score.

  • 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 proposes a two stage framework for cross domain cervical cell detection aimed at mitigating inter institutional domain shift and disease category ambiguity. The key methodological contributions include a SC UNSB (Spatially Continuous Unpaired Neural Schrödinger Bridge), which constructs a synthetic intermediate domain via entropy regularized optimal transport to reduce cross domain distribution gaps; and a dual level feature alignment strategy within a knowledge distillation framework, aligning both shallow structural features and deep semantic representations to facilitate domain invariant knowledge transfer.

  • 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 tackles a clinically relevant and technically challenging problem, namely cross domain cervical cytology detection under pronounced domain shifts.
    • The proposed SC UNSB formulation is conceptually interesting, leveraging optimal transport and Schrödinger bridges to create an intermediate domain rather than directly translating between source and target domains.
    • Integrating unsupervised / unpaired domain modeling with a regularization based detection framework is a promising direction for scenarios with limited labeled data.
    • The use of dual level feature alignment within knowledge distillation is well motivated, as it explicitly addresses both low level structural discrepancies and high level semantic gaps across domains.
    • The paper includes ablation studies and several experimental components intended to isolate the impact of the proposed modules.
  • 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 dataset description lacks critical detail. While approximately 7000 bounding boxes are mentioned for both domains, the number of subjects/patients, samples per subject on each domain are not clearly specified.
    • Given that cell labeling is also provided in the Cric dataset it would have been of interest to also perform classification better and not only detection.
    • The training, validation, and test splits are insufficiently described for each stage of the framework. For reproducibility, the number of subjects, samples, and patches used in each subset should be explicitly stated.
    • It is unclear whether results in Table 1 are reported on the test set.
    • Table 1 is confusing, as it combines generative model metrics with detector performance. When reporting detection results, it should be made explicit which column corresponds to RetinaNet trained with Ds only (Ds?), which uses Ds + Di (all the rest).
    • From Table 2 (ablation study), it is not clearly stated which configuration is selected as the final proposed method, complicating interpretation of the reported results.
    • Table 3 lacks clarity on whether results are evaluated on the test set; hyperparameter analysis should not be conducted on test data, and this distinction should be clearly addressed.
    • The paper does not report statistical significance testing, making it unclear whether the observed improvements in Tables 2 and 3 are statistically meaningful.
    • Overall, the lack of clarity in experimental protocol, dataset usage, and result reporting weakens the strength of the empirical claims despite the methodological novelty. . Minor, some typos should be corrected. Some phrases not starting with capital letter “we…”, or some formulas not introduced with “:”
  • 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.

    (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?

    This paper addresses an important and challenging problem in cervical cytology, namely cross domain detection under strong inter institutional domain shifts and subtle inter class visual differences. The proposed two stage framework combines a conceptually interesting Spatially Continuous Unpaired Neural Schrödinger Bridge (SC UNSB) for intermediate domain construction with a dual level feature alignment strategy via knowledge distillation, which together represent a novel and well motivated approach. The use of optimal transport to explicitly model domain transitions, coupled with regularization and feature alignment in the detection stage, is technically sound and aligned with current research directions in domain adaptation. However, the paper suffers from insufficient clarity in experimental design and result reporting, including unclear dataset composition (number of subjects vs. cells), ambiguous train/validation/test splits, and confusing presentation of results across multiple tables. The lack of statistical significance analysis and unclear identification of the final selected configuration further weaken the empirical claims. Despite these shortcomings, the methodology is original, the problem is relevant, and the reported results suggest potential

  • 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

    Authors propose a two-stage framework for cross-domain cervical cell detection based on cytopathological Image Synthesis (stage 1) and knowledge distillation (stage 2). In stage 1, building upon UNSB, authors make targeted modifications to address the specific challenges of ultra-high-resolution cytopathological images that require patch-wise translation (Authors propose the SC-UNSB module to overcome the boundary drift problem). In stage 2, they introduce a dual-level feature alignment strategy within a knowledge distillation framework, jointly aligning shallow structural representations and high-level semantic embeddings to reduce semantic misalignment under cross-domain settings. Authors conduct extensive experiments on two large-scale public datasets (one as source domain, one as target domain), demonstrating that their framework simultaneously enables high-quality image style transfer and effective model generalization from the source domain to the target domain.

  • 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.
    • Well written: This paper is logically clear, with well-justified motivation, high-quality writing, and elegant figures. Numerous details reflect that the authors have put great care and effort into their work.
    • Methodological soundness: Authors demonstrate deep insight into the problems and each part of the proposed method is appropriate and well-reasoned.
    • Relatively sufficient experimental validation: Authors conduct experimental evaluations from multiple perspectives, including image generation quality, model generalization performance, and hyperparameter selection, etc.
    • Good contributions: The proposed framework meets practical medical needs through its model generalization capability. Moreover, the first stage alone can also function as an effective image style translator.
    • Reproducible: Authors have made the code publicly available, and their experiments are conducted on public datasets.
  • 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.
    • Limited generalization validation: Datasets from different sources can vary significantly, and the difficulty of generalization also differs. However, this paper selects only one dataset (the ComparisonDetector dataset) as the target domain, and the dataset is collected from a single center.
    • Lack of discussion on different disease stages: Authors mention that “due to pronounced domain shifts across institutions and the subtle visual differences among disease stages”. However, they only evaluate their framework on different institutions, with no evaluation across different disease stages. (No additional experiments are required. )
  • 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

    Questions and suggestions to authors:

    • Section 3 (line7): You mention that “Pseudo-labels with confidence above 0.9 supervise the student model”. Do we need to worry about the risk that the teacher model may have high confidence but low accuracy, thereby misleading the student model?
    • Two minor errors: 1.(Introduction, contribution 2) There is an extra period at the end of the last sentence; 2.(Subsection 2.2, para 1) Missing capitalization at the beginning of the first sentence.
  • 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?

    Overall, this is an excellent piece of work that does a good job on multiple levels.

  • 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 paper makes a novel, well-motivated, and technically sound contribution to an important clinical problem. The SC-UNSB module is a genuine methodological advance, the dual-level knowledge distillation is thoughtfully designed, and the reported performance gains are substantial. While experimental reporting could be clearer and validation scope broader, these are addressable weaknesses that do not undermine the core contribution. All three reviewers support acceptance (two Weak Accept, one Accept), and the overall quality is consistent with MICCAI standards.



back to top