Paper Info Reviews Author Feedback Meta-Review Back to top

List of Papers Browse by Subject Areas Author List

Abstract

Echocardiography is critical for diagnosing cardiovascular diseases, yet the shortage of skilled sonographers hinders timely patient care, due to high operational difficulties. Consequently, research on automated probe navigation has significant clinical potential. To achieve robust navigation, it is essential to leverage historical scanning information, mimicking how experts rely on past feedback to adjust subsequent maneuvers. Practical scanning is an exploratory trial-and-error process that inherently generates noisy trajectories. However, existing methods typically model this history as a sequential chain, forcing models to overfit these noisy paths, leading to performance degradation on long sequences. In this paper, we propose UltraStar, which reformulates probe navigation from path regression to anchor-based global localization. By establishing a Star Graph, UltraStar treats historical keyframes as spatial anchors connected directly to the current view, explicitly modeling geometric constraints for precise positioning. We further enhance the Star Graph with a semantic-aware sampling strategy that actively selects the representative landmarks from massive history logs, reducing redundancy for accurate anchoring. Extensive experiments on a dataset with over 1.31 million samples demonstrate that UltraStar outperforms baselines and scales better with longer input lengths, revealing a more effective topology for history modeling under noisy exploration. Code is available at https://github.com/LeapLabTHU/UltraStar.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1157_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/LeapLabTHU/UltraStar

Link to the Dataset(s)

N/A

BibTex

@InProceedings{WanTen_UltraStar_MICCAI2026,
        author = { Wang, Teng AND Jiang, Haojun AND Li, Chenxi AND Wang, Diwen AND Tang, Yihang AND Sun, Zhenguo AND Deng, Yujiao AND Song, Shiji AND Huang, Gao},
        title = { { UltraStar: Semantic-Aware Star Graph Modeling for Echocardiography Navigation } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16893},
        month = {September},
        page = {pending}
}

Reviews

Review #1

Please describe the contribution of the paper

This paper reformulates echocardiography probe navigation from sequential path modeling to anchor-based global localization. It proposes a Star Graph framework that treats historical keyframes as spatial anchors connected directly to the current view, and further introduces a semantic-aware sampling strategy to select representative landmarks from long and noisy scanning trajectories. Experiments on a large-scale dataset suggest improved accuracy and scalability over prior history-modeling baselines.
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 is overall well motivated and easy to follow. I found the main idea interesting, and the work has several appealing aspects in both formulation and evaluation. -The Star Graph formulation is quite interesting and feels like a fresh way to model the exploratory, trial-and-error nature of echocardiography scanning. -Using historical keyframes as anchors for localization is an intuitive and meaningful idea for probe navigation. -The evaluation on a relatively large robotic echocardiography navigation dataset also makes the empirical results more convincing.
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 robustness of the semantic-aware sampling strategy remains unclear, as it relies heavily on the quality of the view classifier output. If historical candidate anchors are of low quality, their predicted semantic distributions may be inaccurate, leading to misleading similarity estimation and noisy anchor selection. The paper does not sufficiently explain how unreliable anchors are prevented from contaminating the constructed Star Graph. -The semantic-aware sampling strategy maximizes semantic diversity, but semantic diversity does not necessarily imply geometric usefulness for localization. It remains unclear whether the selected anchors are always the most informative ones for geometric reasoning. -The claim of geometric localization would be more convincing if the paper more clearly showed that the relative-action formulation captures meaningful geometry, rather than simply adding pose-related auxiliary features. -It would be helpful if the paper could clarify whether the dataset or source code will be publicly released.
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.

(6) Strong Accept — must be accepted due to excellence
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

I recommend acceptance because the paper presents an interesting and well-motivated idea, with a novel Star Graph formulation that fits the exploratory nature of echocardiography scanning well. The method is intuitive, the paper is clearly written, and the experiments are fairly convincing. Although I still have some concerns about robustness and methodological interpretation, I do not think they undermine the overall contribution of the work.
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.

N/A

Review #2

Please describe the contribution of the paper

The main contribution is 1.modelling the robotic ultrasound-based scanning process and standard plane identification in a Star Graph, instead of in a historic chain of frames. In this manner, most informative key frames from past transducer poses are always kept in connection to current frame by a relative action described by translation and rotation. 2.the semantic aware key frame selection module that ensures informative and complementary keyframes are selected
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 motivation of modeling the trajectory as a star instead of as a chain is clear, well justified and follows naturally from the intuition that most informative historical frames are helpful towards global localization. 2.Large scale evaluation of proposed method both with regard to number of patients as well as with regard to different types of baselines.
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 paper claims but does not demonstrate better performance on longer acquisitions directly. Instead, it varies the size of the graph and shows that the proposed method scales favorably, but it does not report or analyze trajectory length. Whether longer trajectories are implicitly involved as L grows, or whether the same scans are used throughout, is not clear.
- Qualitative Results of the other methods and comparisons are necessary to gain intuition of improvements compared to prior work
- Result Figure 6 needs to be bigger. It should be possible differentiate between the results when printed on A4 pages.
- Discussion about generalizability to other devices which will yield different keyframe appearances is needed.
- it is not clear how the data was split. Results are as well averaged over type of views and not per patient.
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 novel formulation of the scanning process as a graph is conceptually simple, yet well-motivated. However, the clinical scope of the contribution is somewhat unclear. Since the ultimate goal is standard plane identification and cardiac structure assessment, it is not discussed whether prior methods already achieve clinically acceptable performance or whether there is a meaningful gap that this work addresses. On the experimental side, the data splitting strategy is not described in sufficient detail, making it difficult to assess the true generalizability of the method. Furthermore, all experiments are conducted on a single device from a single manufacturer, and there is no discussion of how the approach would transfer to different ultrasound machines, probe types, or acquisition protocols.
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 most important aspects about data splitting and generalization have been explained and clarified in the author’s answer.
- An evaluation on the usefulness of the achieved plane for a downstream diagnosis task compared to baselines would also be interesting in future publications, since for now the effective clinical advantage is rather small.
- It would be very interesting in the future to see the close-loop performance of the system
- The authors should ensure that the main results figure are scaled and sufficiently visible.

Review #3

Please describe the contribution of the paper
This paper studies echocardiography probe navigation and argues that historical scan context should be used for global localization rather than sequential trajectory reconstruction. The proposed UltraStar framework builds a Star Graph in which sampled historical keyframes act as spatial anchors connected directly to the current view, and a semantic-aware sampling strategy selects diverse anchors from noisy scan histories. Main contributions include:
- Replaces chain-based history modeling with an anchor-based Star Graph in which historical keyframes are connected directly to the current view for global localization.
- Evaluated on an in-house robotic echocardiography dataset with 178 adult subjects, 356 expert trajectories, and 1.31M image-pose samples.
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.
I think the major strengths include:
- The paper addresses a clinically relevant problem. Robust probe guidance could meaningfully reduce the operational barrier of echocardiography and is important for both sonographer assistance and autonomous robotic scanning.
- The main modeling insight is clear and well motivated: replacing chain-style history modeling with anchor-based global localization is a sensible response to noisy exploratory trajectories.
- The study is supported by a substantial real-data collection effort, with 178 subjects, 356 trajectories, 10 standard target planes, and 1.31M image-pose samples.
- The experimental section is reasonably strong for an initial paper in this space: it compares against several families of baselines, includes a scalability analysis with longer histories, and examines the effect of sampling strategy.
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.

Overall the paper is well-written and the targeted clinical problem is also very relevant in practice. Here are my comments:

1.The evaluation protocol is weaker than the paper’s claims require. Results are reported on a train/validation split, but I got confused that there seems no held-out test set, no external cohort, and no clear statement about whether the I-JEPA pretraining on “our dataset” is restricted to the training portion only. This makes it difficult to judge true generalization.

2.The paper evaluates only offline action prediction MAE, which is an indirect proxy for navigation quality. It does not report actual target-view acquisition success, final view quality, closed-loop rollout performance, or any clinically interpretable endpoint showing that the predicted actions lead to useful scans.

3.Reproducibility is only partial. The architecture is described reasonably well, but the dataset is in-house, code availability is not clearly stated. It will be very beneficial if the authors can attach some example data in the supplementary (additional view/trajectory demo etc.).
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.

(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 clear and plausible modeling idea on a valuable real dataset, and the anchor-based localization perspective is stronger than a generic sequence-modeling reimplementation. However, the current evidence is still too indirect for a stronger recommendation, primarily because missing test set, generalization, leakage control, and actual navigation success are not established clearly enough.
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

Thanks for reviewers’ valuable comments. <R2,R4:Data split and generalization> Our dataset is strictly patient-level split (Sec. 3.1). The I-JEPA encoder and view classifier were trained only on training scans, ensuring zero data leakage. The 72-scan held-out set evaluates demographic generalization on entirely unseen subjects with diverse ages (18–69), BMIs (19.0–34.6), and cardiac pathologies; thus, view-averaged metrics reflect true per-patient generalization. For cross-device robustness, ultrasound devices mainly affect low-level appearance of the images. In data collection, sonographers dynamically adjusted imaging parameters per patient, naturally inducing broad low-level appearance variation. Moreover, recent work[1] showed that a navigation model trained on 247,000 ultrasound images from a single device successfully generalizes to different ultrasound devices, whereas our model is trained on over one million images. We’ll add multi-center external validation to future work.
Our view classifier achieves >95% validation accuracy, providing reliable semantics for sampling. We do not claim the selected anchors are always geometrically optimal. Instead, semantic diversity is a practical proxy for geometric usefulness: semantically similar images typically have small spatial variation and high redundancy, while distinct semantics imply larger spatial spans and richer geometric complementarity. Moreover, our architecture mitigates anchor noise(Fig.3): Cross-Attn assigns higher weights to anchors useful for the current view, preventing unreliable ones from dominating the prediction. Evaluating navigation performance via offline proxy metrics is a well-recognized and prevalent practice within this field[2-4]. As a practical metric, offline MAE strongly correlates with navigation quality. For a study focused on methodological innovation, it directly and rigorously quantifies the model's intrinsic spatial reasoning capabilities at the algorithmic level. The closed-loop validation is an excellent suggestion to further establish clinical impact. We are actively overcoming hardware and ethical barriers to deploy our model on a robotic platform to evaluate real-world closed-loop performance for future journal extension. Visualizing key cardiac landmarks requires millimeter-level adjustment precision, making even minor mm/° error reductions clinically meaningful for preserving high-quality diagnostic views. We'll add qualitative baseline comparisons to illustrate this gain. As a clinical reference, the average expert repositioning error is reported to be 5.68°[3]. While the best baseline achieves 6.79°, our method reduces this to 6.40°, effectively shrinking the gap to expert-level performance by 35%. We clarify that the scaling experiment varies the number of sampled keyframes from a fixed trajectory, not scan duration. We chose this setting because, in practice, the past scan sequence is a fixed given. Since raw trajectories are too long for full-sequence input, we evaluate whether the model can exploit denser samples from all available history; artificially truncating the trajectory contradicts the practical goal of maximizing context utilization. Intuitively, we believe UltraStar is superior on longer trajectories, as prolonged scans introduce more trial-and-error noise. While baselines overfit this noise, our global anchoring bypasses it, ensuring robustness. Table 1 shows EchoWorld uses more pose features but performs worse, proving that raw pose quantity is insufficient and our gains stem from clean anchor-based geometric reasoning. We'll enlarge and relayout Fig.6.<R1,R4:Reproducibility> We'll release the source code and example data. [1] UltraBot, Nat. Commun.’25 [2] Pose-GuideNet, MICCAI’24 [3] Real-world Visual Navigation for Cardiac Ultrasound View Planning, MICCAI’24 [4] EchoWorld, CVPR’25

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 reviewers agree that the paper addresses an important problem and presents a clear, interesting idea. The proposed Star Graph formulation reframes echocardiography probe navigation from sequential path modeling to anchor-based global localization, using historical keyframes as spatial anchors for the current view. This is a well-motivated approach for noisy exploratory scan trajectories. A relatively large robotic echocardiography dataset also supports the paper and compares against several relevant single-frame, sequential, and graph-based baselines. At the same time, the reviews raise several important concerns. The evaluation is mainly based on offline translation/rotation MAE, which is only an indirect proxy for navigation performance. The paper does not show closed-loop navigation, target-view acquisition success, final image quality, or other clinically interpretable outcomes. Generalization is also not fully established, as the experiments use an in-house dataset collected with a single ultrasound system/probe and lack an external test cohort. The data split and pretraining protocol should be clarified, especially whether the frozen I-JEPA encoder and the view classifier were trained only on the training scans or also used validation subjects. The semantic-aware sampling strategy is interesting, but its reliance on the view classifier and the relationship between semantic diversity and geometric usefulness requires clearer justification. The authors should specifically address the data/pretraining split, the robustness of semantic-aware sampling, and the gap between offline MAE and actual navigation success.
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.

I recommend acceptance. The paper proposes UltraStar, a Star Graph formulation for echocardiography probe navigation that uses historical keyframes as spatial anchors directly connected to the current view, together with a semantic-aware sampling strategy. All three reviewers recommend accept post rebuttal, and the rebuttal clarifies the key concerns raised in the reviews: the dataset is strictly patient level split with the I-JEPA encoder and view classifier trained only on training scans, the held-out set evaluates demographic generalization on unseen subjects, the semantic sampling relies on a view classifier with over 95% validation accuracy with Cross-Attn down weighting unreliable anchors, and offline MAE is justified as a standard proxy in this field with closed-loop validation acknowledged as future work. Some concerns remain, including the absence of closed-loop navigation, single-device evaluation, and limited qualitative comparison with baselines. Still, these are limitations of scope and depth rather than fatal flaws, and the authors have committed to multicenter validation and closed-loop deployment in future work. Given the novelty of the anchor-based formulation, the substantial real data collection of 178 subjects and 1.31M image pose samples, and unanimous reviewer support, the paper meets the bar for acceptance.

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.

The authors have addressed the key concerns regarding data splitting and generalization with an explanation of their patient-level split and the demographic diversity of their test set. The explanation of using offline MAE is reasonable, and commitment to releasing the code and sample data would help address concerns with reproducibility. Addition of qualitative baseline comparisons as well as a clinical reference on repositioning error (provided in rebuttal) are helpful. AC recommends paper for acceptance; however, authors must ensure with all the changes/actions mentioned in the rebuttal are addressed in the final manuscript.

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 evaluation is based on the pre-rebuttal evaluations and the two post-rebuttal ones that were available. Sadly, reviewer 4 did not follow my requests to submit a post-rebuttal evaluation.

The introduced work is valuable as it contains an intuitive and novel way to approach the task at hand. Moreover, evaluation is very sound. Reading the rebuttal and the reaction of reviewers 1 and 2, I think that the major aspects raised by the reviewers were taken into account, I ask the authors to address them in the camera-ready version. I also point to the post-rebuttal recommendations of reviewer 2, and in case they cannot be implemented at least include them as limitations/future work accordingly.

back to top

UltraStar: Semantic-Aware Star Graph Modeling for Echocardiography Navigation

Author(s):