List of Papers Browse by Subject Areas Author List
Abstract
Maternal-fetal US is the primary imaging modality for monitoring fetal development, yet accurate automated segmentation remains challenging due to the scarcity of pixel-level annotations. To address this issue, we propose DACL, a semi-supervised framework for robust fetal US image segmentation. DACL jointly trains a deployment-oriented lightweight convolutional network (1.47M parameters) and a Transformer-based network, leveraging labeled data for supervised learning and unlabeled data via CPS. To enhance prediction stability, we introduce a dual-agreement consistency loss that couples pixel-wise probabilistic divergence with entropy-guided confidence alignment. Unlike conventional CPS methods that enforce agreement only at the prediction level, DACL explicitly regularizes both distributional alignment and uncertainty, thereby suppressing unreliable pseudo-labels and enabling stable cross-architecture pseudo-label learning under extreme annotation scarcity. Furthermore, an interpolation-based consistency strategy using mixup is applied to unlabeled samples to enhance robustness. Under 5% labeled data, DACL improves Dice by up to 2.77% and reduces HD95 by up to 14.69 mm compared with the strongest recent semi-supervised methods, demonstrating significant improvements in boundary accuracy on both fetal head and abdomen datasets. These results demonstrate the effectiveness of agreement-based consistency learning for annotation-efficient fetal US segmentation. Our code is on GitHub.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0556_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/13204942/DACL-for-Semi-supervised-Ultrasound-Segmentation
Link to the Dataset(s)
https://figshare.com/articles/dataset/Automated_measurement_of_fetal_head_circumference_using_2D_ultrasound_images/7002086
https://zenodo.org/records/12697994
BibTex
@InProceedings{WanFan_Dual_MICCAI2026,
author = { Wang, Fangyijie AND Silvestre, Guénolé AND Wang, Ziyang AND Curran, Kathleen M.},
title = { { Dual Agreement Consistency Learning for Semi-supervised Fetal Ultrasound Segmentation } },
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 identifies a practical bottleneck, which is deploying lightweight models under scarce annotation settings, and addresses it with a well-designed teacher-student framework.
- 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 framework is designed to output a highly efficient lightweight convolutional neural network that requires only 1.47 million parameters and 7.03 GFLOPs. This makes the resulting model highly practical for deployment on portable or low-cost ultrasound devices. The implementation of the dual-agreement consistency loss is a skillful approach. By enforcing both probabilistic alignment and confidence-aware agreement at the pixel level, the framework elegantly suppresses the unreliable pseudo-labels that often plague semi-supervised learning in low-data regimes.
- 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 massive gap in capacity between the 1.47M parameter UNeXt student and the 27.15M parameter Swin-Unet teacher might establish a rigid ceiling on what the student can actually learn. Additionally, the very small labeled batch size needs more justification to fully ensure the reproducibility of the training process.
- 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 identifies a practical clinical need (deploying lightweight models under scarce annotation settings) and addresses it with a well-designed, technically sound teacher-student framework. The ablation studies are comprehensive and justify the integration of the proposed loss components. However, the manuscript requires clarification on training stability (due to the batch size of 1). The paper claims DACL outperforms the fully supervised UNeXt and other SSL methods, but it lacks statistical significance testing. Include tests (e.g., paired t-tests or Wilcoxon signed-rank tests) to confirm if the modest gains, such as the 0.36% DSC improvement over LMCT, are statistically significant. Figure 2 presents visual comparisons, but the text does not mention the criteria used to select these specific samples. Ensure and state that these are representative samples matching the average metric performance, rather than outlier best-case scenarios.
- 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.
The detailed rebuttal is greatly appreciated. The responses provided by the authors are highly constructive. 1.It makes logical sense that this design is driven by the practical clinical need for low-resource deployment, especially since the high-capacity Swin-Unet teacher network (27.15M parameters ) is entirely discarded after training. This practical justification is accepted. However, the theoretical ceiling effect of transferring knowledge from a massive model to a highly constrained 1.47M parameter UNeXt student model remains a valid academic consideration. Adding a brief sentence in the discussion section to acknowledge this distillation bottleneck as a potential limitation or a direction for future research would be better? 2.The clarification that a labeled batch size of 1 is a standard practice in extreme low-data settings within medical image segmentation is understandable. Relying on EMA guidance, interpolation consistency, and the proposed dual-agreement consistency regularization to maintain training stability is an understandable approach. But, understandable doesn’t mean accepatable, at least to me. This point may be rendered by my experience. 3.The commitment to reporting the mean and standard deviation across repeated runs, alongside formal statistical significance tests, resolves the primary quantitative concern. 4.The assurance that the selected visual samples represent overall metric trends rather than cherry-picked outliers is fine.
In the end, the rebuttal is mostly satisfactory. Clear, logically sound answers have been provided, along with concrete commitments to improve the manuscript’s overall rigor.
Review #2
- Please describe the contribution of the paper
1.the paper proposed a semi-supervised framework of Dual Agreement ConsistencyLearning (DACL) for robust fetal Ultrasound image segmentation. DACL jointly trains a deployment-oriented lightweightconvolutional network (1.47M parameters) and a Transformer-based network,leveraging labeled data for supervised learning and unlabeled datavia cross pseudo supervision (CPS). 2.To enhance prediction stability, the authors introduced a dual-agreement consistency loss that couples pixel-wise probabilistic divergence with entropy-guided confidence alignment. 3..The ablation study showed the methodd contributes meaningful metrics.
- 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.It focuses on semi-supervised fetal ultrasound segmentation when labeled data is limited. 2 The main strength is the dual-agreement consistency loss, which does more than standard CPS by encouraging the CNN and Transformer to agree not only on predictions but also on confidence. 3.The ablation results also suggest that this part of the method improving the performance.
- 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 overall framework is still built on several existing ideas. It combines known components like CPS, CNN–Transformer co-training, EMA teacher supervision, and mixup-based consistency. So the novelty feels more like an incremental improvement than a really new framework.
2 Also, it is not very clear what the main focus of the method is. Is it mainly about improving accuracy, making the model lightweight, or increasing speed? At the moment, it seems to cover all of these, and they are all presented as strengths. I think the paper would be stronger if the authors picked one main theme and made that the central message, while discussing the other benefits, like lightweight design or faster inference, in the discussion section.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
1.The method innovation is not that strong. It mainly builds on the main framework from other papers and adds some extra components. 2 the uploaded code does not include complete settings for the comparison experiments, so it is hard to properly verify the method’s generalization ability.
- 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 reviewing the authors’ rebuttal, I find that the main concerns raised in the initial review have been adequately addressed.
The authors clarified the task definition, better explained the novelty and motivation of the proposed framework. While some minor presentation and clarification issues should still be addressed in the final manuscript, I believe the work is technically interesting and clinically relevant. Therefore, I support acceptance.
Review #3
- Please describe the contribution of the paper
This paper proposes the DACL Framework, a semi-supervised learning framework for fetal ultrasound segmentation that jointly trains a lightweight CNN (1.47M params) and a Transformer using Cross Pseudo Supervision (CPS).
- 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 addresses a challenge: pixel-level annotation scarcity in fetal ultrasound, while maintaining practical deployability. 2) Using a heterogeneous teacher-student (CNN + Transformer) architecture leverages complementary representations, which is a reasonable strategy for semi-supervised segmentation. 3) Promises open-source code, enhancing reproducibility.
- 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 core novelty is ambiguous. It remains unclear whether the two agreements refer to (a) CNN vs. Transformer predictions, (b) two terms inside the loss, or (c) CPS plus interpolation consistency. This risks being an overstated claim. 2) Comparisons are made only against LMCT and a few others. Recent SOTA methods are missing. The claimed improvement may not hold against true state-of-the-art. 3) The entropy-guided confidence alignment is described vaguely. How are unreliable pseudo-labels suppressed? No thresholding, weighting, or filtering scheme is specified. 4) The 1.47M parameter claim applies only to inference. Training cost (memory/time) is not reported. 5) All improvements/losses are presented as point estimates without p-values or confidence intervals, making it impossible to judge whether differences are meaningful.
- 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 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?
All improvements/losses are presented as point estimates without p-values or confidence intervals, making it impossible to judge whether differences are meaningful.
- 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.
The authors’ response has not resolved my concerns, and it appears that the manuscript requires major revisions.
Author Feedback
R2-Q1, R3-Q1, R3-Q3: We thank the reviewers for highlighting the need to better clarify the core novelty and the role of the proposed dual-agreement mechanism. The core novelty of this work is the proposed Dual-Agreement Consistency (DAC) loss, which introduces two complementary cross-architecture agreement constraints for semi-supervised segmentation: (1) probabilistic agreement via KL divergence and (2) uncertainty-aware confidence agreement via entropy regularization. Unlike standard CPS-style supervision that mainly enforces hard pseudo-label agreement, DAC jointly regularizes both prediction distributions and prediction uncertainty across heterogeneous CNN/Transformer architectures. Rather than explicit confidence thresholding, the entropy-guided agreement term penalizes inconsistent high-entropy predictions while favoring mutually confident low-entropy agreement. The heterogeneous CNN/Transformer pairing further provides complementary local/global inductive biases, making agreement regularization more informative than homogeneous supervision. Consequently, DAC improves pseudo-label stability under extreme annotation scarcity, particularly for lightweight models. This is supported by the ablation study (Table 3), where adding L_dac produces the largest gains in boundary-sensitive metrics, especially on the challenging F-Abd dataset. We agree these mechanisms can be clarified further and will revise the manuscript accordingly.
R1-Q1, R2-Q2, R3-Q4: We appreciate the reviewers’ comments regarding lightweight deployment. The primary goal of this work is to develop a robust semi-supervised framework for lightweight fetal ultrasound segmentation suitable for low-resource clinical deployment. Swin-Unet is used only during training as a high-capacity teacher to guide the lightweight UNeXt model (1.47M parameters, 7.03 GFLOPs). During inference, only UNeXt is deployed, while the teacher is discarded and introduces no inference overhead. The large teacher–student capacity gap is therefore intentional and clinically motivated. Existing SSL frameworks still struggle to effectively train lightweight models under limited annotations, whereas DACL consistently improves both overlap and boundary accuracy.
R2-Q1, R3-Q2: We thank the reviewers for raising concerns regarding fair comparison. All cross-supervision methods were implemented using the same Swin-Unet/UNeXt pair, while mean-teacher methods used UNeXt/UNeXt. All methods were trained using identical data splits, augmentations, optimization settings, and evaluation protocols. Therefore, the observed improvements are attributable to DACL rather than stronger backbones. We additionally included several recent SSL methods, including PCPCS (2024), DSTCT (2024), and LMCT (2025). DACL achieves the best overall HC18 performance and the strongest boundary accuracy on the challenging F-Abd dataset.
R1-Q1, R2-Q2: We appreciate the reviewers’ concerns regarding reproducibility and training stability. The labeled batch size of 1 follows common SSL medical image segmentation practice under extreme low-label settings. Stability is further improved through EMA guidance, interpolation consistency, and DAC regularization. We will release the complete training configuration, hyperparameters, implementation details, and data split settings to ensure reproducibility.
R1-Q1, R3-Q5: We thank the reviewers for these valuable suggestions. In the camera-ready version, we will report mean±std across repeated runs together with statistical significance testing for the primary quantitative comparisons. Fig. 2 samples were selected as representative examples reflecting overall metric trends rather than best-case outliers, and we will clarify this explicitly in the manuscript. We additionally agree that future validation across additional datasets and lightweight backbones would further strengthen the generality of the framework.
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 reviews are mixed. The lightweight semi-supervised concept addresses a real clinical need. A rebuttal can resolve key concerns: clarify novelty, add statistical significance tests, and report training cost.
- 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 proposed Dual Agreement Consistency Learning (DACL) framework is recognized as a clinically meaningful and practically valuable solution for resource-constrained fetal ultrasound segmentation under limited annotations. While concerns were initially raised regarding the incremental nature of the framework, potential distillation bottlenecks, and the lack of statistical significance testing, consensus post-rebuttal highlights that the authors have successfully clarified these queries through robust justifications and commitments.
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 rebuttal, two reviewers have recommended acceptance. The other reviewer did not fully explain his/her motivation for rejecting the paper, mentioning only that concerns (presumably, the lack of p-value measuring the significance of improvements) where not resolved. Therefore, the AC recommends acceptance but strongly urges authors to update their paper accordingly,
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.
After carefully reviewing the rebuttal and the post-rebuttal reviewer comments, my recommendation is to accept the paper. The paper addresses a clinically relevant and practical problem: semi-supervised fetal ultrasound segmentation with a deployment-oriented lightweight model. The proposed dual-agreement consistency learning framework is technically sound, and the use of a lightweight CNN student with a stronger Transformer-based teacher is well motivated for low-resource clinical deployment. The rebuttal adequately addressed most of the major concerns, including the motivation of the lightweight design, training stability under limited labeled data, the representativeness of qualitative examples, and the need for repeated runs and statistical significance testing. Although one reviewer still maintained concerns about the need for major revisions, the majority of post-rebuttal opinions support acceptance, and the authors provided concrete commitments to improve the final manuscript. Therefore, I recommend acceptance, with the expectation that the authors will clarify the novelty, report statistical significance and mean/std results, and better explain the training/inference cost distinction in the final version.