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

Abstract

Accurate thoracoabdominal multi-organ segmentation of the fetal from mid-trimester volumetric ultrasound is crucial for prenatal screening and assessment of fetal development. Over half of severe structural anomalies can be detected through such examinations. However, this task is particularly challenging due to the inherent low contrast and ill-defined boundaries in volumetric ultrasound, which are compounded by significant feature distribution shifts caused by variations in fetal pose. To address this, we propose PoseBridgeNet, a novel framework for segmenting six key organs in this challenging setting. Our core innovation is the integration of a Pose-Evolving Schrödinger Bridge (PSB) within the bottleneck of the segmentation model. This module formulates the alignment of feature distributions across different poses as an entropy-regularized optimal transport problem. By explicitly simulating the dynamic evolution of features across pose variations via a learnable stochastic differential equation, the model gains robustness against pose-induced feature shifts. Furthermore, we design a Mixture-of-Experts (MoE) decoder that integrates the stochastic features from the Pose-Evolving Schrödinger Bridge: it consists of one shared expert capturing cross-pose anatomical commonalities and multiple specialized experts modeling organ- and pose-specific variations, enabling joint optimization of pose-aware feature alignment and segmentation.Evaluated on an in-house dataset of fetal volumetric ultrasound, PoseBridgeNet outperforms existing methods, achieving an average DSC of 85.52% , an IoU of 75.44% and a HD of 3.10 mm. Ablation studies further validate the effectiveness of each proposed component. The code and data are available at https://anonymous.4open.science/r/PoseBridgeNet.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2461_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/GanSusan/PoseBridgeNet.git.

Link to the Dataset(s)

https://github.com/GanSusan/PoseBridgeNet.git.

BibTex

@InProceedings{GanShu_PoseBridgeNet_MICCAI2026,
        author = { Gan, Shushen AND Wu, Yongqian AND Duan, JinQiao AND Liang, Jiahong AND Deng, Chengdong AND Zhang, Wanqin AND Zhou, Xiaojing AND Zhu, Yixin AND Shi, Yu AND Gao, Yi},
        title = { { PoseBridgeNet: Learning a Pose-Evolving Schrödinger Bridge for Multi-Organs Segmentation in Prenatal Volumetric Ultrasound } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16884},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The paper proposes PoseBridgeNet, a new segmentation framework for fetal thoracoabdominal organ segmentation in 3D ultrasound that explicitly addresses the large appearance changes caused by fetal pose variation. The main contribution is the introduction of a Pose-Evolving Schrödinger Bridge in the bottleneck of a U-Net-like architecture. This module models the feature distributions of different fetal poses as a stochastic transport process and learns to smoothly transform features between poses, thereby making the segmentation more robust to pose-induced feature shifts.

  • 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 addresses an important and underexplored clinical problem: multi-organ segmentation in mid-trimester fetal thoracoabdominal volumetric ultrasound. Most prior fetal ultrasound work has focused on the head or femur, whereas this paper tackles simultaneous segmentation of six clinically relevant organs in a challenging anatomical region 2.The main methodological idea is original and technically interesting. 3.The manuscript is clearly written and well organized.

  • 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 experimental validation is limited by the small dataset size. Although 453 volumes are available, the authors ultimately select only one volume per subject, resulting in a final dataset of just 80 volumes. This substantially weakens the claim that the model is robust to pose variation, since the discarded scans likely contain additional pose diversity. The paper does not justify why the remaining 373 volumes are not used, nor whether alternative train/test splits would lead to similar conclusions. 2.The absolute performance gains are relatively modest. Compared with the strongest baseline, the overall Dice only improves from 84.92% to 85.52% (+0.6%), while IoU improves by less than 1%. The larger improvement in Hausdorff distance is encouraging, but the overall gains may not be sufficient to justify the additional complexity of the proposed Schrödinger bridge and Mixture-of-Experts framework. The paper reports mean results across the 5 folds, and Table 3 includes mean ± standard deviation for the ablation study, but there is no formal statistical significance analysis (e. g. , paired t-test, Wilcoxon signed-rank test, confidence intervals, or p-values) comparing the proposed method against the baselines. Therefore, it is unclear whether the relatively small improvements in Dice and IoU are statistically significant 3.The paper lacks external validation. All experiments are conducted on a single in-house dataset acquired with one scanner and one probe. 4.The comparison with baselines is not fully convincing because it is limited to conventional segmentation architectures such as 3D U-Net, V-Net, and Swin UNETR. Given the recent rapid development of medical foundation models, the paper should also compare against stronger pretrained approaches, for example variants of Segment Anything Model adapted to 3D or ultrasound, or recent medical foundation models such as MedSAM, SAM-Med3D, or pretrained volumetric transformer models. Since the proposed contribution is intended to improve robustness to pose variation, it is important to understand whether the same robustness could already be achieved by large pretrained models without the added complexity of the Schrödinger bridge formulation.

  • 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?

    The paper presents a genuinely novel and well-motivated methodological idea, and the application to fetal ultrasound is interesting and underexplored. Although the experimental gains are modest and the validation is limited to a small in-house dataset without comparison to foundation models or statistical significance testing, the proposed Schrödinger bridge formulation is sufficiently original to place the paper slightly above the acceptance threshold.

  • 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 has satisfactorily addressed my main concerns: statistical significance is now supported by Wilcoxon tests, and the addition of nnUNet and MedSAMv2 comparisons strengthens the empirical evaluation. I maintain my score.



Review #2

  • Please describe the contribution of the paper

    This paper proposes a new network architecture to segment mid-trimester fetal thoracoabdominal volumetric ultrasound. The proposed method is based on the UNet architecture and adds two contributions aiming at making the network more robust to appearance variations due to fetal pose variation:

    • First, a Schrodinger bridge transporting the bottle-neck feature maps towards the distributions of such features under varied fetal pose is learned
    • Second, the decoder is equipped with several mixture of expert blocks, different experts supposedly covering different pose ranges. A general shared expert is also added. The model is trained end-to-end, randomly sampling target feature distributions from within the mini-batch to train the Schrodinger bridge. This approach is evaluated on an in-house dataset with manual expert annotations. An ablation study is also provided. Both of these experiment tend to indicate a positive contribution of these 2 additions to the UNet in this context.
  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    The paper tackles a clinically relevant challenge (i.e. mid-trimester fetal thoracoabdominal volumetric ultrasound segmentation) and clearly exposes the technical difficulties associated to it (i.e. pose-dependent organ appearances).

    The proposed modifications to the UNet are significant and would constitute a clear methodological contribution.

    The experiments, while partially incomplete in my opinion (see following section of the review), are informative and sound. The data seems qualitative and well curated. The choice of models considered as baseline is reasonable and the ablation study is informative. A fair amount of details are provided, which should allow to reproduce the experiments. The code should be released later.

  • 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.

    In my opinion, the main weakness of the paper is that the justification for the methodological contribution/architecture modifications are not clear or fully convincing to tackle this issue. I list in the following, in no particular order (even though I number them to ease the potential rebuttal), a series of questions I have that would need to be answered for me to be fully convinced by this contribution. 1.The pose-dependent organ appearance impacts all stages of the encoder, why only trying to deal with this issue at the bottle neck? Especially as the model is a UNet, the skip-connections might significantly reduce the impact of the Schrodinger bridge. 2.It is not clear how the “pose-appearance-augmented bottle-neck features” (as I recon you can consider the z_i s as such) are fed into the decoder. 3.The fetal-pose not only impacts the organ appearance but also the spatial distribution of the features. This does not seem to be taken into account. In particular, the second term of the loss L_PSB aims at reconstructing the target feature map in L2 norm. Hence, if the second element sampled from the batch for the end-point of the bridge has a very different pose, then the bridge is not only taking care of the appearance shift but also of the pose-shift. How is that informative for a segmentation task? It seems to me like this would damage the model rather than help it. 4.There is no systematic direction in this learned Schrodinger bridge, z_0 and z_1 can be swapped from one training batch to the next. Hence, I don t understand what functionality this can effectively learn. For instance, if one pose is in feature distribution space in between 2 other poses, then in which direction will the drift function push the intermediate pose? 5.The addition of the mixture of expert mechanism and the Schrodinger bridge seems somehow antagonistic. The mixture of expert aims at dealing with pose-dependent features by rooting them to the relevant expert while the Schrodinger bridge aims at augmenting the bottle-neck representation to achieve pose-agnostic feature representation. In a way, if the Schrodinger bridge succeed, the mixture of expert is not needed. It would be interesting to clarify this.

    A more minor weakness in my opinion is that the experiment section could be easily improved and made more convincing. I would suggest two things: 1.Regarding the main comparative experiment, while the choice of baseline models is reasonable, I feel like it is always a good practice to include nnUNet results as a point of reference. Even if the preprocessing pipeline is different and if might be complex to fully reproduce with custom architectures, this allows to make sure the scores reported are not too far from an accessible open-source solution. 2.Regarding the ablation study, it seems like the MoE is the main source of improvement (which tend to support some of my interrogations above). Hence, a deeper investigation would have been insightful, e.g. more smaller experts vs fewer larger experts.

    Finally, another minor issue is that the link to the anonymized repo points to an empty repo but I assume the code will be released later.

  • 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 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?

    While the paper has some clear merits, the methodological section raised too many unanswered important question to be accepted.

  • 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.

    Authors provided a sufficiently convincing answer to my points in their rebuttal. I therefore suggest approval for this paper.



Review #3

  • Please describe the contribution of the paper

    This paper introduced a U‑Net–based architecture for multi‑organ segmentation in mid‑trimester fetal thoracoabdominal volumetric ultrasound, to tackle a major challenge, i.e. pose‑induced feature distribution shifts. Two main contributions are: (1) A Pose‑Evolving Schrödinger Bridge (PSB) inserted into the bottleneck, modeling pose‑driven feature evolution via a learnable SDE. (2) A Mixture‑of‑Experts (MoE) decoder combining a shared expert with multiple specialized experts to fuse stochastic features and capture organ‑ and pose‑specific variations.

  • 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 mid‑trimester thoracoabdominal ultrasound is a domain with limited prior segmentation work. The authors achieved detection over half of severe structural anomalies. (2) Consistently outperformed baselines across all organs. Improvements are especially notable for the liver, where DSC increases from 85.61% to 87.56%—the organ most affected by pose variation. (3) Clear ablation studies. (4) The SDE formulation, Euler–Maruyama discretization, and consistency loss are clearly described.

  • 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 method assumes pose variation but does not measure or report pose differences. Without pose annotations, it is difficult to verify whether PSB truly aligns pose‑conditioned distributions or simply acts as a stochastic augmentation. (2) The paper does not quantify: training times, inference latency, memory overhead, which is important for clinical deployment. (3) The paper does not discuss competitive methods other than SB, such as diffusion and flow matching models.

  • 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?

    The PSB approach is grounded in SDE theory, and amply supported by experiments with improvements over several strong benchmarks. The paper lacks discussion of other generative models such as diffusion and flow matching models in lieu of SB for multi‑organ segmentation in mid‑trimester fetal thoracoabdominal volumetric ultrasound datasets.

  • Reviewer confidence

    Somewhat confident (2)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Author Feedback

We sincerely thank all reviewers for their valuable comments and insightful suggestions. Due to the space limitation, we consolidate and respond to several major concerns here. We first clarify the working mechanism of the proposed Pose-evolving Schrödinger Bridge (PSB), followed by the underlying design considerations and trade-offs, which also help address several reviewer concerns. Finally, we provide additional explanations regarding the experimental results. Regarding the working mechanism, the objective of the bridge is to model the trajectory process of fetal image variations under pose changes. PSB can be understood as a stochastic semantic interpolation process on a high-level semantic feature manifold. Sampling along such trajectories effectively enriches feature diversity. Therefore, concerning Reviewer2 (Q5) on whether PSB conflicts with the Mixture-of-Experts (MoE) module, we argue that the pose-evolving Schrödinger bridge and the MoE decoder are complementary rather than contradictory. Specifically, PSB generates richer and more diverse latent features, while the MoE decoder is designed to enhance the model’s capability to adaptively model heterogeneous data distributions. Since the sampling direction inherently contains stochasticity, the specific evolution direction is determined by the current state of the drift function. Nevertheless, regardless of whether a “lateral” pose sample is gradually evolved toward a “supine” or “prone” pose, the sampled intermediate features still contribute to improving feature diversity. For Reviewer2 (Q1 and Q3), we first clarify that the second L2 norm term in the L_PSB​ loss serves as a directional regularization term. It measures the consistency between the target sample z1 and the terminal state reached through the recursive formulation in Eq. (3), thereby constraining the bridge to evolve toward the direction of z1.This design is inspired by the classical Iterative Proportional Fitting (IPF) strategy in Schrödinger bridge methods. The reviewer insightfully pointed out that fetal pose variation simultaneously involves both spatial distribution changes and appearance-shift. We intentionally place the pose-evolving Schrödinger bridge at the bottleneck layer of the segmentation network so that it mainly models appearance-shift in the deep semantic feature space while alleviating the burden of modeling spatial transformations. If introduced into shallow low-level features, the bridge would inevitably need to heavily model spatial distribution variations. We deliberately weaken this aspect because segmentation networks themselves already possess a certain degree of robustness to spatial transformations, and conventional augmentations such as random rotations and flipping have been widely validated as effective means to improve segmentation performance. Therefore, although the two types of shifts cannot be completely disentangled at the bottleneck level, our design encourages the model to focus more on learning appearance variations. We appreciate the reviewer’s suggestions regarding the experiments, and to comply with the rebuttal regulations, we only provide a concise summary of the conclusions here.For the comparative experiments, we conducted Wilcoxon signed-rank tests using the second-best performing VNet as the baseline. The results indicate that, except for the HD95 metrics of the Stomach and Left Kidney, all performance improvements are statistically significant. In Table 3, both PSB and MoE significantly improve UNet across all metrics, while the final model further achieves highly significant HD95 improvements over the MoE-only variant. nnUNet achieved an mDSC of 86.09%, while integrating the proposed method into the ramework further improved the mDSC to 87.19%. In addition, we also evaluated MedSAMv2 using box prompts for segmentation. However, without task-specific fine-tuning, its performance in this task remained limited.



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.

    This paper proposes PoseBridgeNet, a novel framework for multi-organ segmentation in mid-trimester fetal thoracoabdominal 3D US. It addresses pose-induced appearance variation via a Pose-Evolving Schrödinger Bridge in the bottleneck of a Unet and a Mixture-of-Experts decode.

    Reviewers agree that the problem is important and underwxplored and that the main idea is original and technically interesting. Experiments show consistent improvements across organs and supporting ablation results.

    At the same time, the reviewers raise some concerns which should be addressed and discussed in a rebuttal. The main concerns include the limited dataset size (selection of only one volume per subject), only modest performance gains (without statistical significance tests), no validation on external data, comparison to stronger baselines is missing, and, finally, conceptual questions about the design and placement of the Schrödinger bridge and combination with the MoE components.

  • 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 authors have sufficiently addressed the reviewers’ concerns, and all reviewers agree on acceptance. Notably, the reviewers requested additional results in the rebuttal and responded positively to them. However, I did not consider these new results (nnU-Net, statistical testing) in my evaluation.



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.

    PoseBridgeNet introduces an innovative, mathematically rigorous Pose-Evolving Schrödinger Bridge integrated into a U-Net bottleneck alongside a Mixture-of-Experts decoder to successfully counter complex pose-induced appearance variations in mid-trimester prenatal 3D ultrasound multi-organ segmentation. The authors’ rebuttal successfully clarified the complementary mechanics between the stochastic semantic interpolation of the bridge and the distribution modeling of the decoder, while well-justifying the bottleneck placement as a deliberate design choice to isolate appearance shifts from low-level spatial transformations. Evaluated strictly on the merits, architectural novelty, and empirical validation presented in the original manuscript, this work constitutes a strong and complete methodological contribution suitable for publication.



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 proposes PoseBridgeNet, addressing a clinically significant but relatively understudied problem, i.e., the fetal thoracoabdominal multi-organ segmentation in prenatal volumetric ultrasound images. The reviewers reached a consensus regarding the clinical relevance and technical novelty of the proposed method. After the rebuttal, the authors successfully clarified issues regarding the underlying working mechanisms and clarified some missing details. As a result, all three reviewers expressed positive opinions on this paper. While some limitations, such as the small dataset scale and some missing competing baselines (which are not allowed to be added after rebuttal), remain, the paper’s methodological innovation and its contribution to an understudied domain still outweigh the limitation and make it a valuable addition to the conference. Therefore, I recommend accept.



back to top