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

Abstract

Intrapartum ultrasound enables objective monitoring of fetal head progression during labor, yet clinical practice relies predominantly on single-frame analysis, which is sensitive to noise and fails to capture continuous anatomical dynamics. Leveraging video clips offers richer information but introduces a critical challenge in multi-clinic settings: acquisition variability arising from differences in probe motion, imaging hardware, and clinical protocols distorts latent temporal representations in ways that static harmonization methods cannot address. We collected a novel intrapartum ultrasound dataset from three clinics with heterogeneous acquisition protocols and propose a self-supervised representation learning framework that explicitly disentangles clinic-specific acquisition distortions from physiologically meaningful motion dynamics. The framework integrates temporal regularization, adversarial disentanglement, and hierarchical cross-attention to effectively separate clinic-invariant and clinic-specific representations. The proposed Hierarchical Cross-Attention encoder achieves 99.70% temporal order prediction accuracy, substantially reduces clinic-induced variability, and generalizes to fetal head segmentation on the HC18 benchmark with a Dice score of 87.23%, outperforming both an in-domain autoencoder baseline and random initialization despite its encoder being pre-trained exclusively on intrapartum video. These results establish temporal disentanglement as a principled strategy for separating acquisition variability from true motion dynamics, enabling robust representation learning across heterogeneous clinical conditions.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4921_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/miccai-intrapartum/miccai_intrapartum

Link to the Dataset(s)

N/A

BibTex

@InProceedings{AytIlh_Disentangling_MICCAI2026,
        author = { Aytutuldu, Ilhan AND Yassa, Murat AND Birol Ilter, Pinar AND Ilhan, Sevval AND Erdem, Rumeysa Elif AND Unlu, Umur Goktug AND Akgul, Yusuf Sinan},
        title = { { Disentangling Clinic-Specific Acquisition Variability from Motion Dynamics in Intrapartum Ultrasound } },
        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 makes several key contributions to intrapartum ultrasound analysis and multi-center medical imaging harmonization: it identifies that clinic-specific acquisition variability in intrapartum ultrasound primarily distorts latent temporal trajectories rather than just static appearance, showing that conventional static harmonization methods cannot resolve such temporal distortions; it contributes a new multi-center intrapartum ultrasound video dataset with heterogeneous acquisition protocols from three clinics to support research in temporal representation learning; it proposes a novel self-supervised framework that explicitly disentangles clinic-specific acquisition distortions from clinically meaningful fetal motion dynamics by integrating temporal regularization, adversarial disentanglement, and hierarchical cross-attention; it introduces a Hierarchical Cross-Attention (HCA) encoder that achieves 99.70% temporal order prediction accuracy, strong generalization in leave-one-center-out validation, and 87.23% Dice score on cross-domain fetal head segmentation in the HC18 benchmark, outperforming multiple baseline methods; and it establishes temporal disentanglement as an effective and principled strategy for robust representation learning across multi-clinic ultrasound environments with heterogeneous acquisition 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.The work correctly identifies that multi-clinic variability in intrapartum ultrasound mainly distorts temporal trajectories rather than static appearance, a gap overlooked by traditional harmonization methods (e. g. , ComBat, DANN). This insight is clinically important because fetal head progression is a dynamic process, and preserving motion dynamics directly supports obstetric decision-making. 2.The paper proposes an original self-supervised framework that explicitly disentangles clinic-invariant physiological motion from clinic-specific acquisition artifacts by integrating three complementary components: temporal regularization (scale and spectral) to stabilize latent dynamics, adversarial dual-stream encoding to separate invariant and specific representations, and hierarchical cross‑attention (HCA) for multi-scale temporal modeling. This design is novel because it unifies disentanglement and temporal modeling for ultrasound video, whereas prior work either addressed static domain shift or modeled time without disentanglement. 3.The evaluation is thorough and convincing: it uses temporal order prediction (TOP) to validate temporal structure preservation, leave‑one‑center‑out cross‑validation to test multi-center generalization, ablation studies to verify each component’s contribution, and cross‑domain transfer to the HC18 benchmark to demonstrate anatomical generalizability. Quantitative results are strong (99.70% TOP accuracy, 87.23% Dice score) and statistically significant. 4.The authors contribute a new multi-center intrapartum ultrasound video dataset from three clinics with heterogeneous protocols, addressing the data scarcity issue in this field. The dataset supports reproducible research on temporal dynamics and multi-center harmonization.

  • 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 conducts its most critical cross-domain generalization evaluation on HC18, which is antenatal (prenatal) fetal head ultrasound data, not true intrapartum ultrasound. Although HC18 is a standard benchmark, it lacks the severe motion artifacts, probe instability, and in vivo labor dynamics of real intrapartum imaging. As a result, the reported 87.23% Dice score does not fully validate clinical performance for the target task, and the claimed clinical translational value is not sufficiently verified on the actual target data distribution. 2.The paper compares mainly with traditional static harmonization methods (ComBat, DANN) and a simple temporal autoencoder, but does not benchmark against state-of-the-art temporal self-supervised or video disentanglement models tailored for medical imaging, making it difficult to assess the absolute advancement of the HCA framework. 3.While the model quantitatively disentangles clinic-invariant and clinic-specific representations, the paper does not provide clinically interpretable analysis, leaving unclear how the representations support real obstetric decision-making. 4.The paper does not report inference speed, model complexity, or practical deployment feasibility for real-time intrapartum monitoring, which is essential for clinical translation in busy labor ward environments.

  • 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?

    No

  • 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

    No

  • 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?

    1.The most prominent cross-domain validation uses HC18 (antenatal fetal head ultrasound), not real intrapartum ultrasound, which has distinct imaging conditions, artifacts, and clinical targets. This weakens the evidence for real-world clinical utility in the target labor-monitoring scenario. 2.While the framework is well-designed, the core idea of temporal regularization + adversarial disentanglement is incremental rather than conceptually groundbreaking. The paper lacks comparison with state-of-the-art temporal self-supervised or video disentanglement models in medical imaging, so the absolute technical advance is unclear. 3.The paper does not link the disentangled invariant representations to clinically meaningful motion dynamics (e.g., fetal head descent, rotation, or progression speed), so the clinical benefit of the method remains abstract rather than concrete.

  • 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 authors have addressed all confused issues.



Review #2

  • Please describe the contribution of the paper

    This paper describes the disentanglement of acquisition factor from the latent embeddings of a temporal sequence of image (i.e. a video). The author presented a collection of methods developed for the said purpose, which is evaluated using an intrapartum ultrasound video dataset collected in-house. Code is publicly available while there is no mention of plan to publish the dataset in the paper.

  • 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.

    This paper attempts to solve an important problem, which is the disentanglement of the latent space based on some confounding factors, and specifically on the latent space of a temporal sequence. The author’s careful approach to data splitting is commendable, which is evident from the subject-level and center-level split.

  • 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.

    These are the major issues with the paper in my opinion:

    • clinical relevance: this paper is motivated by the observation that “current practice rely predominantly on single-frame analysis, but fetal head progression is inherently dynamic”. while i agree with the notion where “meaningful clinical info is encoded in motion patterns across time”, i believe this should refer to multiple timepoints across the labor process. the dataset used in this paper, however, appears to be short video clip at a single timepoint during the labor process, likely acquired with some sweep-like protocol, which is really an attempt to capture wider anatomical region instead of motion pattern. I would suggest rephrasing the motivation to explain how such video is beneficial over single-frame approach, as well as to provide more details about the acquisition protocol.
    • basic pre-processing steps: based on Fig.1D, it appears that videos collected from center A and B follow a protocol where the video recording starts either before the probe was placed on the patient, or the initial position of the probe is away from any visible tissue, resulting in the near-zero intensity of the first ~10 frames (and a similar pattern close to the end of the videos). It is intuitive to pre-process these videos such that the initial/final dark frames are cropped away, followed by interpolating the remaining frames to a constant total number of frames (i.e. 100) if necessary. doing so would also make the comparison between center in Fig.1E more meaningful and interpretable.
    • better ablation study design: based on table 3, the ablation study was performed by adding one loss component at a time. it would be more thorough to explore more combination of the loss components. L_scale and L_spec, for example, attempt to tackle very similar issue (both trying to minimize fluctuation, with L_scale doing it at a more global scale while L_spec focus at neighboring points at various distance apart). however, i also understand that this may be out of the scope of the paper.
    • better explanation on evaluation strategy: as far as I understand, the goal of this paper is to disentangle the latent factor based on the acquisition site. the temporal order prediction (TOP) task, however, is designed to evaluate the temporal continuity and consistency of the latent factors. I can see that somewhere down the line the author has also make use of the ASW metric as part of the evaluation process, which presumably is related to the disentanglement aspect. The paper would benefit significantly from an explanation of how this works. Furthermore, circling back to the main motivation of the paper, it would also be interesting to connect this to some clinically relevant downstream task and to provide an evaluation on how the single-frame, temporal (vanilla), and temporal (ours) approaches compare to each other, although again this might be out of scope of the paper.
  • 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?

    Overall I appreciate the effort in addressing the problem of disentangling the temporal latent space by auxiliary factor, but I am concerned about the weaknesses mentioned above.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [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 appreciate the authors’ effort in addressing my concerns. My final recommendation is based on the following comments:

    • single vs multiple timepoint: while the authors had clarified that each subject contributes to multiple clips at distinct timepoints throughout labor, each prediction from the model still come from clip acquired at single timepoint, which, in my opinion, does not leverage on “meaningful clinical info encoded in motion patterns across time”.
    • clinical evaluation: thank you for the explanation on how ASW works
    • removing dark frames: while the dark frames could be resultant of protocol-driven variability, my opinion remains that they should be removed as part of the pre-processing steps of the deployed code. It is likely that these dark frames do not contain any meaningful anatomy, and trimming them off should be straightforward. Otherwise, it is hard to conclude whether the proposed method can disentangle acquisition factor from the embeddings based on how the anatomies are encoded in the embeddings (which is meaningful), or merely based on the number of dark frames in the beginning/end, or simple setup difference e.g. gain level (which is less meaningful). “Center A rises from ~60 to >100 while Center C remains ~60”: assuming that the clips were acquired properly at Center C, potential pre-processing steps can be a combination of (1) intensity normalization across all temporal frames and (2) some mechanism to clip only when there is a huge change in intensity.
    • ablation: the authors confirmed that L_spec is added on top of L_scale during the ablation study. however, to investigate whether L_scale and L_spec are overlapped, they should be activated individually.



Review #3

  • Please describe the contribution of the paper

    This paper presents a self-supervised representation learning framework for multi-center intrapartum ultrasound videos, which explicitly disentangles clinic-specific acquisition variability from real physiological motion dynamics using temporal regularization, adversarial disentanglement, and hierarchical cross-attention.

  • 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) This paper clearly identifies that clinic-specific acquisition variability distorts temporal motion trajectories rather than only static appearance, which is rarely addressed in existing harmonization methods. This new formulation is clinically meaningful because fetal head progression is inherently dynamic, and static domain adaptation cannot correct temporal distortion. (2) Extensive experiments are conducted on a multi-center dataset using cross-validation, ablation studies, and external validation on HC18.This comprehensive evaluation strongly supports the effectiveness and generalization of the proposed framework. (3) This study constructed a multi-center intrapartum ultrasound dataset with heterogeneous protocols, which is valuable for the community and supports reproducible research in this field.

  • 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 proposed model involves multiple parallel encoders, several loss functions. While effective, the overall architecture is somewhat complex, which may slightly limit its reproducibility. (2) The paper relies on numerous abbreviations (e.g., HCA, ASW, FE, …). throughout the text, many of which are not fully expanded at their first occurrence. This may reduce readability.

  • 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?

    The paper addresses a clinically meaningful topic and provides valuable multi-center dataset and solid experiments. However, the overall readability of the manuscript can be further improved.

  • 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 authors have made efforts to address the concerns. While some minor issues remain, I maintain my original decision.



Author Feedback

We thank the editor and reviewers for their time and evaluations. R1/R2/R3 = Reviewer 1/2/3, M = Meta-Reviewer, W = Weakness. CLINICAL SIGNIFICANCE Our paper addresses a fundamental clinical challenge: robust multi-center assessment of labor progression across heterogeneous clinical sites. The partograph — tracking cervical dilation and fetal head station — relies on manual assessment introducing operator subjectivity; intrapartum ultrasound faces similar challenges. A critical barrier for AI-driven labor monitoring is acquisition variability across centers, distorting temporal representations and hindering generalization. Our work represents a principled first step by learning center-invariant temporal representations that generalize across unseen centers (Table 2) and encode anatomically transferable structure (Table 5), establishing a foundation for future clinically robust multi-center AI systems. [M.1/R1.W1] HC18 RATIONALE R1 questions the validity of HC18 as cross-domain evidence. HC18 is used as a representation quality probe, not a clinical performance claim. Both datasets involve 2D B-mode fetal head ultrasound imaging with notable variability (Voluson E8/730, 6-fold pixel range, all trimesters). The domain gap strengthens our interpretation: an encoder trained on intrapartum video achieves 87.23% Dice on antenatal images, surpassing an in-domain autoencoder (83.12%) and random initialization (80.64%), suggesting anatomically transferable representations. HC18 provides pixel-level masks, making it a suitable benchmark. [M.2/R1.W2] SOTA COMPARISON R1 notes absence of CPC/MoCo comparison. We acknowledge this would strengthen the evaluation; however, HCA already achieves 99.70% TOP (CI [99.10, 100.0]), approaching the practical ceiling. Moreover, CPC and MoCo target single-domain SSL and are not designed for multi-center temporal variability. Table 4: raw 47.3%, DANN 92.3%, HCA 99.3%. [M.3+M.4/R1.W3/R2.W1/R2.W4] CLINICAL MEANING AND EVALUATION R2 questions whether clips are single-timepoint sweeps; R1 and R2 question the clinical relevance of evaluation metrics. Regarding R2.W1: each subject contributes multiple clips at distinct timepoints throughout labor, capturing longitudinal physiological variation rather than isolated sweeps. Regarding R1.W3 and R2.W4: ASW measures center-label separability in FE latent space — lower values confirm disentanglement. FE ASW 0.24 vs. RE 0.725 shows center-specific variability isolated in RE while FE captures acquisition-invariant structure; Table 3 shows ASW improves from 0.41 to 0.24 as components are added. TOP evaluates temporal coherence by distinguishing ordered from shuffled sequences. Together these validate temporally structured, center-invariant representations. HC18 transfer (87.23% Dice) provides anatomical evidence. [R1.W4] INFERENCE SPEED R1 notes absence of inference speed data. Deployment uses only the encoder (~2.8M params., ~5.7 GMACs/seq.); Tesla T4 latency is ~2.9s/sequence. Clips are acquired periodically, making periodic inference clinically feasible. [R2.W2] DARK FRAMES R2 suggests trimming boundary frames. Figure 1D shows center-specific variability persists in valid frames (15–80): Center A rises from ~60 to >100 while Center C remains ~60, reflecting protocol-driven variability. Such frames occur in realistic deployment scenarios; robustness is therefore desirable. [R2.W3] ABLATION R2 notes that L_scale and L_spec may overlap. Table 3 shows complementary effects: L_scale provides the largest gain (+33.1pp, 52.8%→85.9%), while L_spec adds further improvement (→91.4%). This suggests distinct roles: L_scale regularizes transition magnitude, whereas L_spec suppresses high-frequency instability. [R3.W1+W2] COMPLEXITY AND ABBREVIATIONS We appreciate R3’s positive assessment of the work. All architectures share an identical backbone; complexity is incremental. Code available in anonymized repository.

Thank you for your consideration.



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.

    The clinical importance is agreed by the reviewers. However, some concerns need to be addressed, including but not limited to:

    1.Clarify the rationale for using HC18.2.Clarify the technical advance, as lack of SOTA comparison. 3.Provide clear justification of the clinical meaning and interpretability of disentangled representations. 4.Explanation of evaluation strategy.

  • 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.

    The rebuttal addressed many of the main concerns and resulted in an overall positive post-rebuttal balance. One reviewer still raises a substantive concern that the current formulation may not yet fully support the claim of capturing clinically meaningful temporal motion information, as opposed to clip-level temporal variation or acquisition-related cues. I agree that this remains an important reservation. However, I do not view it as sufficient to outweigh the overall positive post-rebuttal assessment, given the importance of the problem, and the support from the other reviewers.



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Reject

  • Please justify your recommendation.

    Key concerns persisted after rebuttal, including limited technical novelty and weak clinical interpretability. As these methodological and validation concerns leave the central claims insufficiently substantiated, the paper is not yet ready for publication and is recommended for rejection.



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.

    The work addresses an important and clinically meaningful problem in intrapartum ultrasound, and the proposed temporal disentanglement framework is well motivated and technically solid. While some concerns remain regarding clinical validation and comparison to broader SOTA baselines, I believe the rebuttal has addressed many of the major issues, and the paper makes a valuable contribution overall.



back to top