List of Papers Browse by Subject Areas Author List
Abstract
Ultrasound video segmentation is clinically valuable yet difficult due to speckle noise, weak boundaries, and rapid anatomical deformation. Recent promptable foundation models enable point-guided segmentation, but their direct deployment in ultrasound remains unreliable: a single point provides insufficient spatial context to resolve scale ambiguity, and greedy memory updates amplify early errors into severe temporal drift. We present EchoPilot, a training-free framework for ultrasound video segmentation under sparse first-frame interaction, requiring only a single point click and an anatomical category name. EchoPilot orchestrates a frozen medical vision-language model (VLM) for semantic localization, a vision foundation model (VFM) for dense geometric feature extraction, and a promptable video segmentor for mask prediction and propagation. To resolve initialization ambiguity, we propose Scale-Space Semantic Prompting, which first selects an optimal contextual view via a parameter-free S.E.E.D. (Semantic Energy-Entropy Density) criterion, and then synthesizes geometrically precise auxiliary point prompts from dense foundation features without additional user interaction. To reduce propagation drift, a Reliability-Gated Memory update is further introduced to selectively freeze the segmentor’s memory bank under uncertain predictions, preventing error accumulation. We also contribute the first dynamic fetal placenta ultrasound video segmentation dataset with 671 annotated frames. Across three ultrasound video datasets, EchoPilot achieves state-of-the-art performance under the sparse-interactive setting, consistently outperforming training-free baselines and finetuned specialists.
Project page: https://keeplearning-again.github.io/EchoPilot/.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1265_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://keeplearning-again.github.io/EchoPilot/
Link to the Dataset(s)
N/A
BibTex
@InProceedings{XiaRui_EchoPilot_MICCAI2026,
author = { Xiao, Ruiqiang AND Xing, Zhaohu AND Yang, Yijun AND Han, Zhenyan AND Wang, Weiming AND Wu, Kaishun AND Zhu, Lei},
title = { { EchoPilot: Training-Free Ultrasound Video Segmentation via Scale-Space Semantic Prompting and Reliability-Gated Memory } },
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 authors propose EchoPilot, a training-free framework for ultrasound video object segmentation that operates with extremely sparse supervision and requires only a single point click and an anatomical category name on the first frame. Its core contribution is to address two key challenges in this setting: (1) initialization ambiguity from sparse prompts, and (2) temporal error accumulation (drift) during video propagation. To solve these, the authors propose a combination of Scale-Space Semantic Prompting, which uses frozen vision-language and vision foundation models to automatically select the optimal spatial context and generate auxiliary prompts, and a Reliability-Gated Memory mechanism, which selectively updates the segmentor’s memory to prevent unreliable predictions from propagating errors. Importantly, the method is fully plug-and-play (no fine-tuning), yet outperforms both training-free baselines and finetuned models across multiple ultrasound datasets, alongside introducing a new placenta video segmentation dataset.
- 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 training-free, sparse-interactive ultrasound video segmentation requiring only a single point click and a category label, which is significantly more practical than prior approaches relying on dense annotations or task-specific fine-tuning. 2.The work clearly identifies two fundamental failure modes initialization ambiguity and temporal drift and proposes targeted solutions for each. 3.The method leverages frozen VLM and VFM priors in a complementary way (semantic grounding + spatial refinement), showcasing a nice way of using pretrained models rather than modifying or retraining them. 4.The generation of auxiliary point prompts from dense foundation features (without additional user input) is an effective way that improves segmentation robustness while maintaining minimal interaction. 5.The proposed Reliability-Gated Memory update is a plug-and-play modification to existing segmentors that avoids error accumulation without additional training or significant computational overhead.
- 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 relies on multiple pretrained foundation models (VLM, VFM, and a video segmentor), which increases system complexity and may limit practicality in resource-constrained clinical settings where deploying and maintaining multiple large models could be challenging. 2.Although the approach is training-free, it implicitly depends on the quality and domain alignment of pretrained models (e. g. , BioMedCLIP, DINO), and the paper does not thoroughly analyze how performance degrades under domain shifts 3.A component ablation isolating Stage II on all datasets is missing. The paper ablates the reliability-gated memory mainly on Breast Lesion, but a broader analysis across CAMUS and Placenta would better establish that the gain is consistent rather than dataset-specific. 4.A comparison against simpler scale-selection heuristics is missing. Since S. E. E. D. is presented as a key novelty, it would be helpful to compare it against alternatives such as fixed crop size, best semantic score only, random scale, or largest/smallest crop. This would show whether the proposed energy-entropy formulation is truly necessary.
- 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?
I recommend a weak accept as the paper presents a novel and practically relevant formulation of training-free, sparse-interactive ultrasound video segmentation using only a single point and a category label, which is highly aligned with real clinical constraints. The method is conceptually strong, with well-motivated components—particularly the Scale-Space Semantic Prompting (S.E.E.D.) for resolving initialization ambiguity and the Reliability-Gated Memory for mitigating temporal drift—along with a thoughtful integration of multiple frozen foundation models. The approach is validated with consistent empirical gains across multiple datasets and backbones, and the introduction of a new placenta dataset further strengthens the contribution. However, the paper is somewhat limited by missing ablations (e.g., scale selection, prompt design, and memory gating variants), lack of runtime analysis, and limited robustness evaluation (e.g., prompt and domain sensitivity), which prevent a stronger recommendation.
- 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
The paper studies a clinically relevant and practically motivated setting for ultrasound video object segmentation, where only a single positive click and an anatomical category name are provided on the first frame, and the model must segment the entire sequence without task-specific fine-tuning. The main technical contribution is a training-free framework, EchoPilot, that addresses two key failure modes in this setting: ambiguous first-frame initialization and temporal drift during propagation. Concretely, the manuscript introduces Scale-Space Semantic Prompting for resolving scale ambiguity at initialization, and a Reliability-Gated Memory Update mechanism for suppressing error accumulation over time. A further contribution is the introduction of a dynamic fetal placenta ultrasound VOS dataset. The empirical study on CAMUS, Breast Lesion, and Placenta suggests that the proposed framework improves over several training-free baselines and remains effective across both SAM2 and MedSAM2 backbones.
- 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.
This work addresses an important and clinically relevant use case for ultrasound video segmentation. The sparse-interactive, training-free setting is well motivated, particularly because dense video annotations are expensive to obtain and the common assumption of a full first-frame mask is often unrealistic in practice. A further strength is that the method is organized around two concrete failure modes in this setting: Stage I addresses the ambiguity of determining what and where to segment from sparse initialization, while Stage II targets temporal drift by reducing contamination of the memory bank from unreliable predictions. This high-level decomposition makes the pipeline easy to follow, and Fig. 2 communicates the design particularly clearly.
The empirical evaluation is also a strength. Table 1 indicates consistently stronger Dice and ASD performance across CAMUS, Breast Lesion, and Placenta relative to the compared training-free baselines, and the qualitative examples in Fig. 3 are consistent with the claim that the proposed approach reduces drift over time. Finally, the introduction of a dynamic placenta ultrasound VOS dataset adds potential value beyond the method itself and could be useful to the community, provided that sufficient documentation and access details are made available.
- 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 methodological novelty appears somewhat limited, and the key scale-selection criterion remains only partially justified. The overall framework is well designed and practically motivated, but it mainly combines existing frozen priors in a careful way rather than introducing a clearly new segmentation principle. In addition, although the proposed S.E.E.D. criterion is intuitive, the manuscript does not yet fully justify why a more spatially dispersed attribution pattern should reliably indicate a better contextual scale in ultrasound, where such dispersion may also reflect uncertainty. 2.The reliability-gating mechanism seems to rely on a rather restrictive notion of correctness. The write decision is based on the similarity between the current prediction feature and an anchor derived from the first frame. This design is elegant and computationally efficient, but it implicitly assumes that correct predictions should remain close to the initial appearance. In ultrasound videos, however, substantial deformation, probe motion, shadowing, and out-of-plane variation are common, and a semantically correct target may legitimately deviate from its first-frame representation. As a result, the proposed gate currently reads more as a conservative stabilization heuristic than as a general reliability estimator. 3.The experimental design does not fully disentangle the source of the reported gains. The quantitative improvements are encouraging, but the comparison protocol makes it difficult to determine how much of the gain should be attributed to the proposed orchestration itself and how much comes from the richer prompting setup. In the main comparison, EchoPilot is evaluated with both point and text inputs, whereas most SAM2/MedSAM2 baselines are evaluated with point prompts only. This does not invalidate the comparison, since the target setting indeed includes a category name, but it does weaken the causal interpretation of the results. The ablation study is helpful, yet the respective effects of text conditioning, auxiliary prompt refinement, and memory gating remain only partially separated. 4.The practical scope of the method would benefit from clearer discussion. The framework depends strongly on the first frame providing a sufficiently informative anchor, since both prompt refinement and memory gating are tied to initialization. However, the manuscript does not clearly discuss what happens when the target is only partially visible at the beginning, is poorly localized in the annotated frame, disappears temporarily, or reappears later in the sequence. These cases seem important for understanding the realistic operating range of the method in clinical ultrasound videos. 5.There also appears to be a bibliographic inconsistency in the current manuscript. The same MedSAM3 work seems to be listed twice in the reference list under two different reference numbers, while the main text uses those indices as if they referred to different papers. This should be corrected for clarity and accuracy.
- 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?
I lean toward weak accept. The paper is not especially strong in terms of methodological novelty, since the main ideas read more as a careful integration of existing frozen priors and lightweight heuristics than as a fundamentally new segmentation principle. That said, it addresses a clinically relevant and underexplored ultrasound VOS setting, and the overall system is coherent, well motivated, and supported by encouraging results across multiple datasets. My main concerns are the limited conceptual grounding of S.E.E.D., the strong reliance on first-frame consistency in the reliability gate, and the fact that the comparisons do not fully isolate the benefit of the proposed design from the benefit of richer prompting. Overall, however, I think the practical value is sufficient to justify acceptance.
- Reviewer confidence
Very confident (4)
- [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
This paper presents EchoPilot, a training-free framework for ultrasound video segmentation that uses a single point and a category name as input. The method combines frozen foundation models to address initialization ambiguity and temporal drift by introducing a criterion for selecting an optimal context scale, along with a reliability-gated memory mechanism.
- 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 training-free setting with minimal interaction is practical and clinically relevant, especially for ultrasound where annotations are costly and limited.
-
The integration of VLM and VFM for semantic guidance and spatial refinement is well-motivated and effectively leverages pretrained priors.
-
The method is evaluated across multiple datasets and backbones, showing consistent improvements over SAM-based 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 reliability-gated memory relies on a fixed first-frame anchor and a cosine similarity threshold. While the ablation shows overall effectiveness and stability to the threshold, it does not evaluate robustness under significant appearance changes, occlusions, or initialization errors, where this design may still lead to incorrect acceptance or rejection decisions.
-
The comparison with baselines may not be entirely fair, as most baselines rely solely on point prompts, while the proposed method additionally leverages textual input, which introduces additional semantic information.
-
The prompt refinement step selects up to three auxiliary points via cosine similarity and NMS, but the choice of the number of prompts is not well justified.
-
The method relies on a single category name as the text prompt, but it is unclear how sensitive performance is to prompt phrasing or whether richer textual descriptions could improve or destabilize the results.
-
- 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 minor concerns, the paper presents a well-motivated and practical approach with strong empirical results.
- 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
N/A
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.
This paper presents EchoPilot, a training-free framework for ultrasound video object segmentation that requires only a single point click and an anatomical category name on the first frame. The method addresses initialization ambiguity and temporal drift via Scale-Space Semantic Prompting (S.E.E.D.) and a Reliability-Gated memory mechanism. Additionally, it introduces a new placenta US video dataset.
Reviewers agree that the problem formulation is clinically relevant and that the training-free, low-interaction setting is well motivated. The framework is coherent, well presented and shows consistent improvements over training-free baselines.
The raised concerns relate to limited methodological novelty (R#2), justification of design choices (R#2, R#4), and incomplete evaluation (missing ablations) (R#1, R#2, R#4).
Overall, considering the practical relevance of the work, the empirical performance and its presentation, I support early acceptance.