List of Papers Browse by Subject Areas Author List
Abstract
Ultrasound bone surface segmentation is a critical prerequisite for orthopedic applications such as quantitative analysis and intraoperative navigation. However, reliable segmentation remains challenging due to ambiguous bone boundaries, strong appearance variability, and interference from bone-like soft tissues, often resulting in unstable predictions in complex clinical scenarios. Existing approaches often rely on network designs tailored to specific ultrasound physical priors, which may limit their adaptability across different imaging conditions. In this work, we propose a Probability-Anchored Network (PANet) for robust ultrasound bone segmentation. Built upon a U-Net backbone, PANet explicitly introduces a probability space as an intermediate representation between image features and segmentation results, enabling controllable activation, discrimination, and refinement of bone boundaries. Specifically, global linear bias, local probability exchange, and global attention mechanisms are jointly employed to enhance bone-related responses while suppressing globally inconsistent false positives. Moreover, a unified segmentation head provides globally consistent supervision across all network stages, enabling probability-aligned learning and segmentation-guided attention. Experiments on freehand-scanned long-bone ultrasound datasets demonstrate that PANet achieves more accurate and stable segmentation performance compared with state-of-the-art methods.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1654_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
Link to the Dataset(s)
N/A
BibTex
@InProceedings{YanWen_PANet_MICCAI2026,
author = { Yan, Wenqing AND Li, Haowei AND Ding, Hui AND Zhao, Zhe AND Wang, Guangzhi},
title = { { PANet: Probability-Anchored Network for Ultrasound Bone Segmentation } },
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
This paper proposes PANet (Probability-Anchored Network) for robust ultrasound bone segmentation. The core contribution is an explicit probability space serving as an intermediate representation between image features and segmentation results, which decomposes segmentation into controllable activation, discrimination, and refinement stages. PANet integrates three key modules: Probability Exchange Module (PEM) for local boundary refinement, Probability Bias Module (PBM) for early bone activation, and Probability Reasoning Module (PRM) with global attention to suppress false positives from bone-like soft tissues. A unified multi-level supervision head aligns probability learning across network stages. Experiments on a freehand long-bone ultrasound dataset demonstrate that PANet achieves superior segmentation accuracy, boundary precision, and stability compared with state-of-the-art methods, while maintaining real-time inference speed for clinical usage.
- 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 introduces an explicit probability space as an intermediate representation between visual features and segmentation outputs, which is conceptually original for ultrasound bone segmentation. Unlike prior methods that rely on ultrasound‑specific physical priors or end‑to‑end feature mapping, PANet decomposes segmentation into activation, discrimination, and refinement within a structured probability domain. This design provides better control over false positives from bone‑like soft tissues and improves boundary stability under ambiguous imaging conditions, which is a meaningful and clinically motivated novelty. 2.The three complementary components—PEM for local boundary refinement, PBM for early structure activation, and PRM for global false‑positive suppression—are tightly coupled and theoretically consistent. The local–global modeling hierarchy effectively addresses weak boundaries and background interference, two long‑standing challenges in ultrasound bone segmentation. The use of rotary positional embedding within probability‑space attention is technically sound and well justified. 3.The UPE‑Head provides consistent supervision across stages to align probability learning and guide attention. This design not only boosts performance but also improves the interpretability and traceability of the model’s decision process, which is valuable for clinical translation. 4.The evaluation is thorough: quantitative comparison with a wide range of state‑of‑the‑art models (U‑Net, TransUNet, BCA‑Net, etc. ), complete ablation studies verifying the contribution of each core module, and clinically relevant metrics including IoU, Dice, HD95, recall, precision, and inference speed. PANet achieves top performance with balanced recall and precision, low HD95, and real‑time inference speed (194 FPS), clearly demonstrating clinical practicality.
- 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 entire dataset is collected from one single ultrasound scanner (Mindray Resona 6) at one institution, with no cross-device, cross-probe, or multi-center experiments. Clinical ultrasound systems vary significantly in gain, frequency, resolution, and artifact patterns, so the generalizability of PANet to real clinical scenarios remains unproven. This is a critical limitation for translation to orthopedic navigation. 2.The idea of using an explicit probabilistic intermediate space and multi-level supervision is conceptually similar to several existing works in medical image segmentation. While the combination for ultrasound bone is new, the core paradigm is not sufficiently innovative from a methodological perspective. 3.The ablation removes individual components but does not compare against probability space vs. deep supervision alone in a controlled manner. It also lacks ablation of key hyperparameters (e. g. , probability dimension K, block size in PRM, weight of loss terms) to validate robustness. The relative importance of Gaussian initialization is not rigorously quantified. 4.The paper claims the probability space improves interpretability but provides no quantitative interpretability analysis (e. g. , uncertainty quantification, failure-mode analysis, or visualization of probability evolution across challenging cases). The link between probability anchors and clinical decision-making is underdeveloped. 5.Ultrasound bone segmentation is a typical thin-structure segmentation task with extreme foreground–background imbalance. The paper does not discuss whether the loss function, sampling strategy, or post-processing is adapted to this challenge. No results are shown on fracture, implant, or pathological cases, which are clinically more important than healthy long bones.
- 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
This paper presents a technically solid, well‑motivated, and clinically relevant contribution to ultrasound bone segmentation for orthopedic applications. The proposed PANet with a probability‑anchored intermediate representation is carefully designed to suppress false positives and stabilize boundary prediction under challenging freehand scanning conditions, and it is thoroughly validated through comprehensive comparisons and ablation studies, achieving superior performance and real‑time inference speed suitable for clinical usage.
- 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 core contributions show clear value for ultrasound bone segmentation and achieve strong quantitative results. However, several critical limitations prevent it from meeting full acceptance criteria: (1)insufficient clinical validation and generalization evidence, as experiments are restricted to a single‑center, single‑device dataset on healthy volunteers with no evaluation on pathological, fracture, or multi‑scanner data; (2) incomplete ablation and interpretability analysis, lacking rigorous hyperparameter studies and concrete evidence to support the claimed clinical interpretability; and (3) reproducibility gaps in preprocessing, annotation quality control, and runtime optimization details.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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
Review #2
- Please describe the contribution of the paper
This paper addresses the challenging task of ultrasound bone segmentation. The main contribution is the introduction of a parallel probability space to the feature space of the Unet at different resolutions that can better tackle false positives in comparison to surrounding soft structures.
- 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.
Novel methodology which introduces a probability space along a unet backbone. The features of the unet are propagated into this space and further updated and propagated based on neighboring probabilities. Finally the attention-based module predicts confidence of bone in one pixel relative to surrounding structures. This is interesting because it models and updates the uncertainty explicitly. The intermediate maps that this representation creates can also be visualized to further understand the output of the method.
- 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 claim that the method outperforms STOA in challenging imaging conditions would benefit from a more precise definition of what constitutes a challenging case, along with a dedicated evaluation in these settings. It is also worth clarifying whether expert sonographers agree on ground truth segmentations for such cases.
- Existing works such as UltraBones100k are publicly available and should have been used for benchmarking. Moreover, the therein proposed bone segmentation method would have been a recent stoa baseline to compare to.
- Despite the generic title, the paper focuses exclusively on long bones with no discussion of whether the method would extend to other structures such as the spine.
- Overview figure is unintuitive and some parts of the methodology are unclear. Here are some notes and suggestions for improvement:
- on the upper part of the figure PRM is shown in yellow separate from UPE head, but further below PRM module contains the UPE head, this is rather confusing.
- instead of calling the lower left corner PBM & PEM I suggest calling them probability update module. This contains both PBM & PEM modules. PBM should be marked on the arrow from CNN Feature F to PBM Bias B. PEM should be on the arrow from CNN feature F to PEM weight. Figure 1 should have an alone standing description with the pipeline steps. Currently it is not understandable without reading the main text.
- 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
- Would be good to shortly explain “qualitative analysis” you are mentioning in the abstract.
- in the main figure it would be more intuitive to keep visual encoder decoder structure
- potential title clash with Path Aggregation Network (PANet) from computer vision
- 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?
I think that the proposed parallel probability space is a novel approach, particularly interesting because of its explainability characteristic. However, the unclarities in the method and main figure require improvement. Moreover, a short discussion on wide-range applicability of the method as well as quantification and explanation of clinically challenging cases to ensure the claim is appropriate would be needed.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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
Review #3
- Please describe the contribution of the paper
The paper introduces a novel framework named Probability Anchoring Network (PANet), which outperforms baseline models like U-Net on a self-built large-scale ultrasound bone segmentation dataset, while maintaining an inference speed suitable for clinical real-time applications.
- 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 constructs a large-scale free-hand scanning long bone ultrasound dataset containing 246471 images, conducts five-fold cross-validation, and has a rigorous experimental design. The model has a fast inference speed and high potential for clinical application. Compared to mainstream models such as U-Net, Attention U-Net, and TransUNet, PANet achieves stable improvements in both IoU (81.39%) and Dice (89.43%), and performs well with HD95 (0.97mm). Ablation experiments demonstrate that the global inference module effectively suppresses false positives caused by soft tissue similar to bone texture.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The paper claims to have solved the “boundary blurring” problem and uses HD95 as a distance metric. However, HD95 is sensitive to outliers and cannot reflect the overall boundary fit. In bone segmentation tasks, ASSD (Average Symmetric Surface Distance) is the gold standard for measuring boundary accuracy. The absence of this metric in the paper completely undermines the ability to prove whether the segmentation results truly achieve smooth boundary fit at non-outlier points. 2.The interpretability of the method description is relatively poor. The paper proposes a “probability space”, but fails to clearly explain its physical significance. The definition and explanation of dimension K are missing, and it is unclear what K=64 channels specifically represent. 3.The comparison method is outdated. Although TransUNet and BCA-Net are compared, the current state-of-the-art medical segmentation models (such as SwinUNETR, nnU-Net) are not included in the comparison. Especially, the citation year of BCA-Net is relatively recent (2025). 4.The data solely originates from a single device (Mindray Resona 6) of a single institution, lacking verification from multi-center and multi-vendor devices, thus casting doubt on its generalization ability. The experiment solely targets a private dataset. Currently, there are numerous datasets related to ultrasound bone imaging, yet none of them are cited in this paper.
- 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?
Despite the flaws in the paper, such as missing indicators and insufficient comparison, it has constructed a massive dataset and the model’s inference speed holds great clinical practical value.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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 the reviewers for their constructive comments. We have revised the manuscript to improve the writing and figures, clarify key points, and expand the discussion of limitations and future work.
Code and Data Availability The code and trained models will be publicly released, while the dataset will be available upon reasonable request.
Dataset 1.Future work will include multi-center and cross-device ultrasound data, clinically challenging cases such as fractures, implants, and pathological changes, and other anatomical bone structures to further assess the robustness, clinical applicability, and generalizability of PANet. 2.Regarding UltraBones100K, although it is an important public dataset, our inspection showed that its labels are mainly derived from CT-based automatic annotation and registration rather than frame-wise expert-refined annotations. The resulting label shifts and boundary inaccuracies may compromise the reliability of fine-grained bone surface evaluation; therefore, it was not included in the current experiments.
Experimental limitations We appreciate the reviewers’ comments on the ablation study, evaluation metrics, comparison baselines, and interpretability analysis. We agree that these aspects are important for a more comprehensive evaluation. Due to the constraints of the MICCAI revision stage, extensive new experiments could not be added. Instead, we have explicitly discussed these limitations in the revised manuscript as important future directions. We will continue to evaluate PANet from more perspectives and present more complete results in subsequent work.
Novelty We acknowledge that probabilistic representations and multi-level supervision have been explored in medical image segmentation. The novelty of PANet lies not in simply using probability maps or deep supervision, but in treating the probability space as a core intermediate state for network reasoning. Around this state, PANet designs dedicated probability initialization, activation, local exchange, global discrimination, and unified supervision mechanisms to address artifacts, bone-like interference, and boundary discontinuities in ultrasound bone segmentation. Future work will include more detailed comparisons and controlled validations with related probability-driven networks.
Writing and Figures 1.We have optimized Fig. 1 and its caption in response to R2.The figure is intended to emphasize the design of probability space, the roles of the submodules, and the visual evolution of probability maps. Therefore, it is not presented as a conventional U-Net-like architecture. 2.In response to R3, we added explanations of the physical meaning of the probability space and the role of dimension K. 3.In response to R2, we clarified that the “challenging imaging conditions” mainly arise from ultrasound imaging characteristics. With improved soft-tissue imaging, multiple bone-like soft-tissue structures become visible and may be spatially connected with bone tissues, making bone surface segmentation more difficult. This point has been emphasized in the qualitative comparison. We also clarified that the ground-truth annotations were confirmed by consensus between the data labeler and the surgeon. To avoid ambiguity, the last sentence of the Abstract was removed. 4.In response to R1, the ethics information, which was previously hidden for anonymous review, has now been supplemented.
Scope and Naming We have further clarified the current experimental scope and stated that future work will evaluate PANet on different bone structures. Regarding the naming conflict, PANet in this work explicitly refers to Probability-Anchored Network, which differs from Path Aggregation Network in both task and methodology. We have written the full name at its first appearance to reduce potential confusion.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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.
Despite several significant concerns raised during the initial review phase, this work demonstrates sufficient methodological novelty and practical utility to warrant early acceptance. The primary strengths of the submission lie in its real-time implementation and its consistent performance gains over established baselines. While these improvements are incremental, their feasibility in time-sensitive environments provides clear value to the field.
To strengthen the final camera-ready version, the authors are encouraged to:
Open-Source Contribution: Release the source code and curated datasets to ensure reproducibility and foster community adoption.
Benchmarking: If feasible, include a comparative analysis against nnU-Net. Given its status as a robust benchmark in medical imaging, such a comparison would provide a more comprehensive validation of the proposed method’s efficacy.