List of Papers Browse by Subject Areas Author List
Abstract
Accurate segmentation of the fetal head and pubic symphysis in ultrasound is critical for quantitative obstetric assessment and delivery decision-making. This task is severely hampered by inherent speckle noise and low signal-to-noise ratio, which obscure anatomical boundaries. Furthermore, semantic gaps in skip connections of existing networks degrade structural fidelity, leading to clinically significant segmentation errors. To overcome these limitations, we propose the Boundary-Aware Dual-Domain Feature Reconstruction Network (BDFR-Net), which synergistically models frequency-spatial complementarity to enhance boundary perception and structural fidelity. Specifically, the Dual-Domain Feature Decoupling and Boundary Enhancement (DDBE) is designed to tackle boundary ambiguity by decomposing features into frequency-decoupled edge responses and spatial contextual cues, which are then consolidated via channel-attentive cross-domain fusion. Furthermore, an Adaptive Context Alignment Module (ACAM) is embedded in skip connections to bridge semantic gaps. ACAM recalibrates cross-layer features through dual attention mechanisms, ensuring hierarchical consistency and mitigating feature drift. Extensive experiments on multiple fetal ultrasound datasets show that BDFR-Net achieves superior segmentation accuracy, attaining 86.83\% mean Dice on the MICCAI IUGC 2024 benchmark and consistent improvements across cross-dataset evaluation, demonstrating strong generalization under distribution shift.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3612_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/CodeCQ789/BDFR-Net
Link to the Dataset(s)
Miccai2024 dataset: https://www.kaggle.com/datasets/aspirexxx/iugc-ultrasound-video-dataset-miccai-2024
JNU-IFM dataset: https://figshare.com/articles/dataset/JNU-IFM/14371652
PSFHS dataset: https://zenodo.org/records/10969427
FH-PS-AoP dataset: https://www.kaggle.com/datasets/aspirexxx/psfhs-ultrasound-dataset-miccai-2023
BibTex
@InProceedings{ZhaWen_BDFRNet_MICCAI2026,
author = { Zhang, Wenfeng AND Liu, Nana AND Qin, Qibing AND Huang, Xin AND Tan, Cong AND Hu, Wei AND Bai, Jieyun},
title = { { BDFR-Net: Boundary-Aware Dual-Domain Feature Reconstruction Network for Fetal Ultrasound Image Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16883},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes Boundary-Aware Dual-Domain Feature Reconstruction Network for Fetal Ultrasound Image Segmentation, and conducted Extensive experiments on the MICCAI IUGC 2024 benchmark and cross-dataset for evaluation.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The paper proposes Boundary-Aware Dual-Domain Feature Reconstruction Network for Fetal Ultrasound Image Segmentation, and Extensive experiments on the MICCAI IUGC 2024 benchmark and cross-dataset are conducted in the paper. 2.The low-frequency and high-frequency decompositions are separately designed for multi-scale feature enhancement on boundary segmentation.
- 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.Paper mentioned speckled noise and low signal-to-noise ratio obscure anatomical boundaries. What’s the difference? 2.The proposed modules added in the UNET are more like a combination of different feature enhancement modules, although a performance improvement is observed. Because the spatial attention and different pooling operations are fundamental in unet-based segmentation. 3.The difference between BDFR-B1 and BDFR-B2
- 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 does not provide sufficient information for 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?
The paper is well-organized, but the novelty is limited.
- Reviewer confidence
Somewhat confident (2)
- [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 response only partially addresses my concern. The authors mainly restate the intended functions of DDBE and ACAM, but do not sufficiently demonstrate why the proposed decoder is not a combination of existing frequency-domain, boundary-enhancement, and attention-based skip-fusion modules. The claim that BDFR-Net is “not a generic block assembly” needs stronger support. The authors should clarify the concrete technical novelty of DDBE and ACAM, explain their difference from closely related decoder and skip-refinement modules, and provide additional evidence showing their complementary effect. Boundary-specific metrics, visualizations, or ablation studies under speckle/noisy boundary conditions would make the argument more convincing.
Review #2
- Please describe the contribution of the paper
The paper’s main contribution is BDFR-Net, a fetal ultrasound segmentation network designed to better handle unclear anatomical boundaries and semantic mismatch in skip connections. Its key innovations are a Dual-Domain Feature Decoupling and Boundary Enhancement (DDBE) module, which combines frequency-domain edge cues with spatial context, and an Adaptive Context Alignment Module (ACAM), which aligns encoder and decoder features before fusion.
- 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 focuses on segmenting the fetal head and pubic symphysis in transperineal ultrasound, which matters for obstetric measurements used in delivery decisions. The main novelty is that the proposed BDFR-Net combines frequency-domain boundary cues with spatial-context modeling instead of relying on only one type of feature. The results show that BDFR-Net appears especially helpful on segmenting the pubic symphysis, which is anatomically challenging and clinically relevant.
- 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 clinical task of this study is automated AoP measurement, but the results mainly focus on segmentation accuracy. The paper does not clearly demonstrate the impact on downstream AoP measurement or real clinical workflow. For instance, what is the accuracy of AoP estimation by BDFR-Net? 2.The architecture includes multiple specialized components. While this may improve accuracy, it also increases implementation and deployment complexity. Metrics such as the number of parameters and inference speed should be reported to assess efficiency. 3.The cross-dataset results show a clear performance drop compared to the FHPS-AoP results. This raises my concerns about the generalizability of the proposed BDFR-Net. 4.In addition, the results on the IUGC2024 dataset show performance drops across all evaluation metrics compared to FHPS-AoP. This further raises my concerns about the robustness of BDFR-Net. 5.The training epoch is set to 50 for the IUGC2024 dataset, but the reason for using a different setting is not explained. For fair comparison, training settings should be consistent across datasets. 6.How are Dice loss and cross-entropy loss balanced and combined? Do they have equal weight?
- 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 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?
The paper presents a novel method that shows improved segmentation performance in segmenting FH and PS. However, my main concerns are the limited clinical impact and insufficient evaluation of practical applicability. The paper mainly focuses on segmentation results without clearly showing benefits in downstream clinical use. In addition, I’m concerned about generalizability across datasets, and model efficiency. Overall, despite these limitations, the methodological contribution and relevance of the problem support a borderline acceptance.
- 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 rebuttal addressed my main questions and concerns satisfactorily. I encourage the authors to incorporate these clarifications and responses into the camera-ready version to further improve the clarity and completeness of the paper.
Review #3
- Please describe the contribution of the paper
This paper proposes a Boundary-Aware Dual-Domain Feature Reconstruction Network to segment fetal head and pubic symphysis in ultrasound. The main contributions are as follows: First, this paper proposes a Dual-Domain Feature Decoupling and Boundary Enhancement (DDBE) module from frequency-domain to yield boundary aware features preserving both edge fidelity and regional coherence. Second, the Adaptive Context Alignment Module (ACAM) refines cross-layer representations with spatial and channel attention before decoder fusion, which helps reduce the distribution difference between encoder and decoder feature spaces. Experiments on multiple fetal ultrasound datasets show state-of-the-art segmentation performance and good cross-dataset generalization.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The main strength of this paper is to propose a DDBE module which synergistically models frequency-spatial complementarity to enhance boundary perception and structural fidelity. Specifically, DDBE addresses the issue of boundary blurring by decomposing features into frequency-decoupled edge responses and spatial contextual cues, and then integrating them through channel-attentive cross-domain fusion.
- 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 high-frequency branch is introduced to enhance boundaries, its robustness to ultrasound-specific noise and artifacts is not fully validated. Ultrasound images usually contain considerable noise and various artifacts. As a result, the extracted high-frequency components may include not only true anatomical boundaries, but also speckle noise and false edge responses. The current paper does not provide enough analysis to explain how the proposed design can effectively suppress these unwanted responses. The authors are encouraged to provide more visual evidence, especially on challenging cases with severe speckle noise, blurred boundaries, and obvious artifacts. For example, feature response maps or segmentation comparisons would help support the analysis. (2) Some parts of the method description are not sufficiently clear or rigorous. The definition of the Efficient Channel Attention (ECA) seems closer to a CBAM-style channel attention module than to a standard ECA module. In addition, several equations and notations should be presented in a clearer and more consistent way. (3) The computational cost of the proposed method is not discussed in enough detail. Since wavelet decomposition is integrated into the network pipeline, it seems to be performed in every forward pass during both training and inference. This raises concerns about the extra computational burden, as well as the training and inference time, compared with conventional convolution-based designs. The authors are encouraged to provide a clearer analysis of model complexity, including parameter count, FLOPs, and runtime overhead.
- 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 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?
While the proposed method is well motivated and shows promising results, it still has several limitations, including insufficient analysis of ultrasound-specific noise and artifacts, unclear description, and limited discussion of computational cost.
- 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 rebuttal sufficiently addresses my main concerns. The authors clarify the role of the HF branch and agree to revise the attention-module description and notation. Although additional visual evidence would further strengthen the analysis, the current rebuttal provides a reasonable clarification. Therefore, I recommend Accept.
Author Feedback
We thank all reviewers (R) and the Meta-Reviewer (MetaR) for reviewing and recognizing our clinical motivation (R1,R2),SOTA performance (R1,R2,R3) and clear ablation (R1,R3).The code will be released upon acceptance. Q1.Speckle vs.low SNR(R1).Both are major ultrasound degradations noted in [21],but they affect boundaries differently.Speckle introduces granular false responses that obscure existing boundary cues,whereas low SNR,often caused by attenuation or shadowing,weakens echoes and reduces boundary visibility. Q2.Module-combination concern(MetaR-W1,R1).BDFR-Net is a boundary-driven decoder, not a generic block assembly. Prior decoders [Attention U-Net,SegNet] enhance spatial feature decoding, and frequency-based models [Wave-ViT] introduce spectral representation, but they are not tailored to the speckle-induced ambiguity between weak anatomical boundaries and texture-like responses in FH/PS ultrasound. DDBE addresses this ambiguity through frequency–spatial decoupling and cross-domain coordination. Prior skip refinements [UNet++,UCTransNet] improve feature reuse, whereas ACAM aligns encoder–decoder semantics before fusion. Q3.B1 vs.B2(R1).Only the encoder backbone differs (PVTv2-B1 vs.PVTv2-B2[27]); all other components are identical.This will be clarified in the camera-ready. Q4.AoP clinical relevance (R2). AoP is a key clinical indicator of labor progression [15] and, as reported in the PSFHS challenge report [MIA 99(2025), 103353], is derived from segmented FH and PS via deterministic geometric fitting. AoP is not learned as an independent target, where its accuracy depends directly on segmentation boundary fidelity. A standalone AoP-error table was omitted due to space limits, but the most pronounced PS gain (the AoP baseline structure) in Table 1 already supports clinical relevance. A quantitative AoP-error comparison will be presented in future work. Q5.Computational cost(MetaR-W2,R2,R3).Due to space limitations,parameter count,FLOPs and runtime analyses are not included; a systematic evaluation will be presented in future work. Q6.Cross-dataset drop,IUGC 2024 drop and 50-epoch setting(MetaR-W3,R2).The drops are mainly attributable to distributional differences across datasets rather than a failure specific to BDFR-Net.The cross-dataset protocol trains on JNU-IFM[17] and tests on PSFHS[5] with no overlap in centers,scanners or patients; IUGC 2024[1] tests on 300 standard-plane images from disjoint videos.Both introduce higher variability,so absolute scores decrease for all methods,yet BDFR-Net still achieves the best overall mean metrics (Table 1),evidencing stronger generalization.The 50-epoch setting on IUGC 2024 follows its smaller training scale and prior practice[1],and is applied uniformly to all compared methods for fairness. Q7.Dice+CE loss(R2).Inspired by [TransU-Net,6;I²U-Net,8],we adopt an equally weighted combination of Dice and cross-entropy losses,where Dice provides robustness to foreground-background imbalance and cross-entropy provides dense pixel-level supervision for boundary discrimination. Q8.HF robustness to speckle/artifacts(MetaR-W3,R3).We agree that speckle may enter the high-frequency response,but HF is not used in isolation: under segmentation supervision,DWT-decomposed subbands are projected and refined to bias toward anatomical edge responses,and then re-weighted with the LF and CTX branches via channel-spatial gating (Eq.6) to suppress branch-specific artifacts.Prior work has also established the robustness of wavelet-based high-frequency representations to speckle [Wavelet-CNN,CVPR’20;Wave-ViT,ECCV’22]; This will be clarified in the camera-ready.Due to space limitations,finer-grained feature response maps will be included in future work. Q9.”ECA” naming(R3).This module is a CBAM-style channel attention in which the shared MLP is replaced by two independent 1×1 convolutional branches.We will refine the description and unify the notation in the camera-ready version.
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.
Please address the concerns about limited novelty, missing efficiency analysis, and insufficient validation under ultrasound-specific conditions.
- 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.
two accept and one reject.
i read through and would accept subject to minor
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.
This paper addresses a clinically relevant fetal ultrasound segmentation problem and reports competitive results on several benchmarks. However, after considering the reviews and rebuttal, I recommend rejection because the current submission remains below the acceptance threshold for a methodological contribution.
1) The methodological novelty is limited. Frequency-domain decomposition, high-/low-frequency feature modeling, boundary enhancement, and attention-based skip-feature refinement have already been widely explored in medical image segmentation, including ultrasound-related settings. The proposed DDBE and ACAM modules appear to be a task-oriented combination of these existing ideas rather than a clearly new technical formulation. The rebuttal mainly restates the intended roles of the modules, but does not sufficiently distinguish the concrete technical contribution from prior frequency-domain, boundary-aware, and attention-based decoder designs.
2) The baseline comparison is not sufficiently targeted. Since the paper’s central claim is boundary-aware dual-domain reconstruction, the comparison should include closely related frequency-domain segmentation methods, boundary-enhancement methods, and ultrasound-specific segmentation approaches. In addition, the baseline set does not adequately cover recent architecture families such as Mamba- or RWKV-style medical segmentation models. Without these stronger and more relevant comparisons, the claimed superiority over existing methods is not fully convincing.
3) The evidence for the claimed boundary and ultrasound robustness is incomplete. The method is motivated by speckle noise, low SNR, and ambiguous anatomical boundaries, but the paper does not provide sufficient boundary-specific evaluation, feature visualizations, or stress analysis under challenging ultrasound artifacts. The downstream clinical relevance to AoP measurement is also not quantitatively demonstrated, even though this is a key motivation for segmenting the fetal head and pubic symphysis.
4) The practical value of the added architectural complexity is not well established. The framework introduces multiple components, including wavelet-based frequency decomposition, frequency-spatial reconstruction, contextual aggregation, and adaptive feature alignment. However, the ablation study shows relatively limited gains from several individual components, and the paper does not provide complexity or runtime analysis to justify whether these added modules are worthwhile compared with simpler alternatives.
Overall, although the paper is well organized and addresses an important application, its core technical contribution is not sufficiently novel or distinctive for a methodological paper at MICCAI.
Meta-review #3
- 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.
The rebuttal clarified several presentation and implementation details and parts of the experimental setup. However, the main issue remains that the proposed method does not provide sufficient methodological novelty. In particular, the overall design still appears to rely on relatively standard frequency-feature fusion and attention-based refinement ideas, and the rebuttal does not sufficiently establish these components as a more distinct technical contribution. Considering the remaining post-rebuttal reviewer concerns, I do not support acceptance.
