List of Papers Browse by Subject Areas Author List
Abstract
The Angle of Progression (AoP), derived from three anatomical landmarks in intrapartum ultrasound, is a crucial metric for assessing fetal head descent and supporting safe labor management. However, automated AoP measurement is severely impeded by speckle noise, acoustic shadows, heterogeneous acquisition conditions, and scarce expert annotations. To address these challenges, we propose DK-Net, a semi-supervised Wavelet-KAN framework that combines student-teacher self-ensembling with ultrasound-aware spectral-spline modeling. DK-Net introduces two Kolmogorov-Arnold Network (KAN) driven components: (1) a domain-adaptive channel recalibration module that dynamically adjusts feature responses to contrast variations, and (2) a wavelet-domain spectral refiner that learns non-linear gating functions on high-frequency subbands to suppress speckle-like perturbations while preserving landmark-defining structures. The network further enforces prediction consistency between weakly and strongly augmented views using confidence-filtered pseudo-labels to effectively leverage unlabeled data. Experiments on the public Intrapartum Ultrasound Grand Challenge (IUGC) benchmark demonstrate that DK-Net consistently outperforms representative single-model baselines, enabling accurate landmark localization and reliable AoP estimation for clinical labor monitoring. Our code is publicly available at https://github.com/db0725/DKNet, and the datasets can be accessed on Kaggle (https://www.kaggle.com/aspirexxx) and Zenodo (https://zenodo.org/records/18217137).
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4589_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/db0725/DKNet
Link to the Dataset(s)
Intrapartum Ultrasound Grand Challenge (IUGC) datasets: https://www.kaggle.com/aspirexxx
Intrapartum Ultrasound Grand Challenge (IUGC) datasets: https://zenodo.org/records/18217137
BibTex
@InProceedings{DenBo_DKNet_MICCAI2026,
author = { Deng, Bo AND Bai, Jieyun AND Wang, Huijin AND Long, Shun AND Li, Shuo},
title = { { DK-Net: Semi-supervised Wavelet-KAN Landmark Localization for Angle of Progression Measurement in Intrapartum Transperineal Ultrasound } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16880},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper presents a semi-supervised system for landmark detection in obstetric ultrasound. The main technical novelties are two additional blocks based on KAN layers, one for feature normalization at the bottleneck and one for adaptive smoothing of the final feature maps. The model is evaluated on the IUGC 2025 benchmark.
- 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 blocks make sense for high-throughput architectures and could be interesting for other researchers working on real-time deployments.
Semi-supervised training is well described, even if it is not the main innovation of the paper.
- 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 IUGC 2025 report has now been published, and this makes both the paper’s bilbliography and quantitative evaluation very weak: there is no point in implementing baselines when the results from previous challenge participants are available. Incorporating this information into the paper would require a major rewrite.
Bai, Jieyun, et al. “IUGC: A benchmark of landmark detection in end-to-end intrapartum ultrasound biometry. “ Medical Image Analysis (2026): 103960. https://doi. org/10.1016/j. media. 2026.103960
That leaderboard shows that the performance of the proposed method is below the state of the art. The KAN blocks themselves may be useful but they would have to be added to a stronger baseline for the benefits to be convincing on this benchmark.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Integrating the IUGC report into the paper requires a major rewrite, I am not sure that can be done within the rebuttal period.
- 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.
With a more thorough representation of the state of the art, the paper should reach the acceptance threshold.
Review #2
- Please describe the contribution of the paper
This paper proposes DK-Net, a semi-supervised framework for automatic angle of progression (AoP) estimation from intrapartum transperineal ultrasound. The method formulates the task as landmark localization via heatmap regression and integrates a Mean Teacher framework to leverage unlabeled data. To address ultrasound-specific challenges, the authors introduce two modules: a KAN-based channel recalibration module (KCA) for handling contrast variations, and a wavelet-based frequency refinement module (WKG) for suppressing speckle noise. The approach is evaluated on the large-scale multi-center IUGC benchmark and demonstrates improved AoP estimation accuracy over several semi-supervised baselines.
Main Contributions (1) A semi-supervised framework (DK-Net) for AoP estimation via landmark localization in intrapartum ultrasound. (2) A KAN-based channel recalibration module (KCA) designed to improve robustness to contrast and acquisition variability. (3)A wavelet-domain refinement module (WKG) for suppressing high-frequency speckle noise while preserving structural features. (4)Comprehensive evaluation on a multi-center public benchmark, including both landmark localization error and clinically relevant AoP error.
- 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.Clinically meaningful problem: The paper addresses a well-defined and practically important task (AoP estimation), with clear relevance to labor monitoring and decision support. The motivation is strong and aligned with current clinical needs. 2.Realistic evaluation setup: The use of the IUGC multi-center benchmark, with strict separation between training and test sets, strengthens the credibility of the results. The inclusion of both labeled and unlabeled data reflects realistic clinical scenarios. 3.Appropriate metrics: Reporting AoP error (in degrees) in addition to landmark localization error is a significant strength, as it directly reflects clinical utility rather than relying solely on pixel-based metrics. 4.Consistent empirical improvements: The proposed method shows improvements over multiple semi-supervised baselines, and ablation studies indicate that both KCA and WKG contribute to performance gains. 5.Robustness in challenging cases: The qualitative examples suggest that the method is more stable under severe speckle noise and acoustic shadowing, which are common failure modes in ultrasound.
- 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.Limited methodological novelty: While the overall design is well-motivated, the framework combines several existing components (Mean Teacher, wavelet-based processing, channel recalibration) in a relatively straightforward manner. The novelty mainly lies in the integration rather than in fundamentally new formulations. 2.Unclear role of KAN: The use of Kolmogorov–Arnold Networks (KAN) is interesting, but its advantage over simpler alternatives (e.g., MLP-based recalibration) is not strongly justified. The reported improvement over MLP is relatively modest, and it is unclear whether the gains stem from KAN specifically or from general nonlinear modeling capacity. 3.Limited mechanism-level validation: Although ablation studies are provided, the paper does not thoroughly verify whether the proposed modules address the intended challenges. For example, it would be helpful to explicitly demonstrate robustness under controlled contrast shifts or varying levels of speckle noise. 4.Weak analysis of AoP geometry: The observation that AoP error is not directly correlated with landmark localization error is insightful, but this aspect is not explored in depth. Additional analysis could strengthen the clinical interpretation of the results. 5.Potential over-complexity: The framework introduces multiple components (KCA, WKG, phase-guided interaction), which may raise concerns about complexity versus benefit, especially given the relatively incremental performance gains.
- 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 clinical problem and presents a method that is both practically motivated and empirically effective. The evaluation is conducted on a strong multi-center benchmark with appropriate clinical metrics, which adds credibility and relevance to the work.
However, the methodological contribution is somewhat incremental, and the novelty of individual components—particularly the use of KAN—is not fully convincing. The paper would benefit from deeper analysis to better explain why the proposed design works, rather than primarily demonstrating that it works.
Despite these limitations, the combination of a clinically meaningful task, solid experimental validation, and consistent performance improvements makes this work a reasonable contribution to the MICCAI community. Therefore, I lean towards a weak accept.
- 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.
After reading the rebuttal and considering the other reviews, I maintain my weak accept recommendation.
The concerns about comparison with the IUGC 2025 challenge methods are valid. The paper would be stronger if it more explicitly positioned DK-Net against the top-ranked challenge solutions and clarified that it is not necessarily the state-of-the-art method on the benchmark. However, the authors reasonably clarified that their focus is a controlled, single-model, real-time semi-supervised framework rather than an ensemble-based challenge-winning pipeline.
The rebuttal also partially addresses my original concerns about the role of KAN and the AoP geometry. The explanation that KAN is used as a task-motivated calibration function, together with the KAN-vs-MLP ablation and learned function visualization, makes the design more convincing. The clarification that AoP error depends on PS/FH geometry rather than being strictly linear with MRE is also helpful.
I still consider the methodological novelty moderate and the mechanism-level validation somewhat limited. Nevertheless, the paper addresses a clinically meaningful problem, uses a relevant public benchmark, reports clinically meaningful AoP metrics, and shows consistent empirical improvements over controlled baselines. Overall, I consider it a marginal but acceptable contribution.
Review #3
- Please describe the contribution of the paper
This paper presents a semi-supervised framework, DK-Net, for automatic landmark localization and Angle of Progression (AoP) measurement in intrapartum transperineal ultrasound. The method combines a Mean Teacher style training strategy with two ultrasound-oriented modules, namely a KAN-based domain-adaptive channel recalibration module and a wavelet-domain feature refinement module, with the goal of improving robustness to speckle noise, acoustic shadowing, and limited annotations. According to the manuscript, the method is evaluated on the public IUGC benchmark and shows improved performance over several generic baseline models.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
This paper presents a semi-supervised framework, DK-Net, for automatic landmark localization and Angle of Progression (AoP) measurement in intrapartum transperineal ultrasound. The method combines a Mean Teacher style training strategy with two ultrasound-oriented modules, namely a KAN-based domain-adaptive channel recalibration module and a wavelet-domain feature refinement module, with the goal of improving robustness to speckle noise, acoustic shadowing, and limited annotations. According to the manuscript, the method is evaluated on the public IUGC benchmark and shows improved performance over several generic baseline models.
- 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.Although the manuscript reports comparisons with several generic landmark detection and semi-supervised baselines, it does not compare against, cite, or discuss the dedicated methods reported in the IUGC 2025 challenge benchmark, which is the most relevant evaluation setting for this work. In particular, the IUGC benchmark report explicitly documents the top-performing challenge methods and their associated technical papers, including the winning T2 entry. The authors should therefore cite the IUGC benchmark report [1] and include a direct comparison with the winning method [2]. According to that paper, the winning approach is based on a modified TransUNet with a TinyViT backbone, together with MAE-assisted knowledge distillation from an ultrasound foundation model, label perturbation, and semi-supervised pseudo-labeling, and it reports 11.6749 px mean radial error and 3.8061° mean absolute AoP error on the IUGC2025 challenge test set. Since the current manuscript reports 12.83 px MRE and 4.30° AoP MAE on the same benchmark, the omission of this direct comparison makes it difficult to assess the real state-of-the-art competitiveness of DK-Net. At minimum, the authors should explicitly cite these two works and discuss why DK-Net should be preferred over, or considered complementary to, the current challenge-winning solution. 2.The practical value of the semi-supervised setting is not fully characterized. The method is trained with 1,580 labeled images and 5,497 unlabeled images from the same benchmark protocol, but the manuscript does not show how performance changes as the amount of unlabeled data varies, nor whether the benefit remains meaningful when only a smaller in-domain unlabeled pool is available. This matters because the usefulness of a semi-supervised approach in practice depends not only on the final test accuracy, but also on its data-efficiency and deployment assumptions. A simple sensitivity analysis or label-efficiency curve would make the paper much stronger and would clarify whether the observed gain is robust or mainly dependent on access to a relatively large amount of benchmark-matched unlabeled data.
[1] Bai J, Tang Y, Liu X, et al. IUGC: A benchmark of landmark detection in end-to-end intrapartum ultrasound biometry[J]. Medical Image Analysis, 2026: 103960.[2] Ma C, Li Y, Guo B, et al. Unlabeled Data-Driven Fetal Landmark Detection in Intrapartum Ultrasound[M]//Intrapartum Ultrasound Grand Challenge. Cham: Springer Nature Switzerland, 2025: 14-23.
- 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 clinically relevant problem and presents a technically sound semi-supervised framework for automatic landmark localization and AoP measurement in intrapartum ultrasound. The combination of Mean Teacher training with KAN-based channel recalibration and wavelet-domain feature refinement is well motivated for the characteristics of ultrasound imaging, and the manuscript is generally clear and well organized. The reported improvements over several generic landmark detection and semi-supervised baselines also suggest that the proposed method is effective.
My main reservations are related to the completeness of the experimental validation rather than to the overall quality of the work. In particular, since this paper is evaluated on the IUGC benchmark, it should more explicitly position itself against the most relevant prior work in the same challenge setting, especially the IUGC benchmark report and the challenge-winning method. In addition, the practical value of the semi-supervised setting is not fully characterized, as the manuscript does not analyze performance under different amounts of unlabeled data or more limited unlabeled-data regimes.
Overall, I believe this is a useful and relevant contribution, although the paper would be stronger with more complete benchmark positioning and a clearer characterization of the semi-supervised setting.
- 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.
After reading the rebuttal, I fully support acceptance. The authors have clearly addressed my main concern regarding the positioning against the IUGC 2025 benchmark and the challenge-winning method. They clarified that DK-Net is designed as a clean, single-model, real-time framework rather than a leaderboard-oriented competition pipeline, which I find convincing and important.
In my opinion, the comparison with the top IUGC result should be interpreted carefully. The challenge-winning solution benefited from a full competition pipeline, including preprocessing, TTA, ensemble/model selection, and post-processing, all of which can substantially improve the final leaderboard score. According to the details reported in the winning solution, when these competition-oriented strategies are removed and the comparison is made under a fair single-model setting, DK-Net is clearly stronger and more favorable in terms of accuracy, simplicity, efficiency, and deployability.
The rebuttal also provides a clearer explanation of the KAN-based modules and the geometric relationship between landmark localization and AoP measurement. Although annotation-ratio analysis would further strengthen the work, I do not consider this a critical limitation. Overall, I believe this is a clinically meaningful, technically sound, and practically valuable contribution. I fully agree that this paper should be accepted.
Author Feedback
We sincerely thank all reviewers and the AC for their constructive comments. We respond to the main concerns on IUGC 2025, KAN, and the semi-supervised setting.
Comparison with IUGC 2025 and positioning (R2,R4,AC): We thank R2, R4, and the AC for pointing out that the comparison with IUGC 2025 methods was not sufficiently discussed. We also thank R2 for questioning the strength of our baselines. We clarify that these baselines remain useful because they were re-trained under the same input, data split, and semi-supervised protocol as DK-Net, allowing a controlled assessment of the proposed KCA/WKG modules rather than a comparison confounded by different training recipes. After revisiting our work in the context of the IUGC benchmark (Bai et al., 2026) and related challenge methods (Ma et al., 2025), we would like to clarify the practical advantage of DK-Net. Compared with top-ranked IUGC methods, DK-Net uses a single-model inference pipeline, without relying on multi-model ensembles or complex post-processing, while still keeping competitive AoP measurement performance. We believe this simpler design is useful for real-time assisted recognition in intrapartum ultrasound, where timely feedback and deployment cost matter. The IUGC report also notes the trade-off between accuracy and efficiency, and the need to consider latency and model complexity in practical deployment. Therefore, we view DK-Net as a lightweight real-time complement to high-performance IUGC methods, and as a step toward more deployable and methodologically grounded AoP measurement systems.
Methodological contribution and KAN effectiveness (R3,AC): We thank R3 and the AC for raising concerns about novelty and mechanistic validation. We agree that the role of KAN should be explained more clearly. Our intention is not to present KAN as a generic replacement for CNNs or Transformers. Instead, we use KAN in two small but task-motivated calibration modules. Specifically, KCA uses simple channel statistics to adapt feature responses under contrast and acquisition variations. WKG uses local high-frequency statistics in wavelet subbands to learn data-dependent soft-thresholding, suppressing speckle- and shadow-related perturbations while preserving structured landmark edges. In both modules, KAN is used as a nonlinear statistics-to-calibration function learner. Its spline-like form provides local adaptivity, which is useful because ultrasound responses can be non-monotonic: strong responses may correspond to true PS/FH boundaries, but may also come from noise or shadows. Fig. 3 further shows that KAN learns more dynamic calibration/gating curves than the MLP counterpart, supporting that the gain is not merely from adding parameters. Together with the ablation results, this supports that KCA/WKG address the intended ultrasound degradations rather than acting as generic add-ons. In addition, AoP depends on the geometry between the PS axis and the FH tangent point, rather than on independent landmark errors linearly. For example, correlated shifts of the two PS endpoints along the PS axis may increase MRE but cause little angular change, whereas a small FH displacement perpendicular to the PS axis can noticeably change AoP. DK-Net uses PS structural information to guide FH localization, encouraging the model to focus on geometry directly relevant to AoP. This helps explain why MRE and AoP MAE are not strictly linearly correlated.
Semi-supervised setting and annotation-ratio analysis (R4): We thank R4 for the helpful suggestion. We agree that varying labeled/unlabeled ratios would better characterize data efficiency. Here, we focus on a controlled comparison: DK-Net, KCA/WKG, and all baselines are evaluated under the same semi-supervised protocol. The results suggest improved AoP measurement over representative methods and the no-KCA/WKG variant while retaining real-time inference. We regard annotation-ratio analysis as important future work.
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 paper introduces a semi‑supervised framework for measuring the angle of progression in transperineal intrapartum ultrasound by localizing anatomical landmarks. The work addresses a clinically relevant problem and is tested on a public multi‑center benchmark. However, the methodological advances are found to be incremental by the reviewers, which might be defended in a rebuttal. Also, the authors should probably address the missing comparisons to the IUGC benchmark methods and to provide a more thorough mechanistic validation of their modules.
- 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 convincingly clarifies the KAN-based modules and the geometric relationship between landmark localization and AoP measurement, making the work technically sound and clinically meaningful. The reviewers and myself are more keen to accept the work now, and I recommend acceptance. For the camera-ready version, the authors should add an annotation-ratio analysis, and explicitly discuss the issues outlined by the reviewers (e.g., the baseline relies on a full competition pipeline in contrast to a fair single-model setting).
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.
All three active reviewers moved to accept post-rebuttal and the recommendation is therefore Accept.
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.
This paper presents a clinically meaningful and empirically sound framework for automated angle of progression measurement. Following the rebuttal phase, the reviewers reached a consensus recommendation of accept. The authors’ rebuttal successfully addressed the initial concerns regarding comparisons to the challenge benchmark by demonstrating that DK-Net balances competitive accuracy with the superior simplicity and efficiency required for real-time ultrasound deployment, contrasting favorably with over-engineered, competition-optimized pipelines. Although concerns regarding incremental novelty remain, the practical value of the proposed method outweighs this limitation. Therefore, I recommend accept.