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

Abstract

Echocardiography is a critical tool for detecting heart diseases, yet its steep operational difficulty causes a shortage of skilled personnel. Probe guidance systems, which assist in acquiring high-quality images, offer a promising solution to lower this operational barrier. However, robust probe guidance remains challenging due to significant individual variability. This variability manifests as differences in low-level features within two-dimensional (2D) images, which complicates image feature understanding, and differences in individual three-dimensional (3D) structures, which poses challenges for precise navigation. To address these challenges, we first propose leveraging the robust image representations learned by ultrasound foundation models from vast datasets. Yet, applying these models to probe navigation is non-trivial due to their lack of understanding of individual 3D structures. To this end, we meticulously design a Vision-Action Adapter (VA-Adapter) to online inject the capability of understanding individual 3D structures. Specifically, by embedding the VA-Adapter into the foundation model’s image encoder, the model can infer cardiac anatomy from historical vision-action sequences, mimicking the cognitive process of a sonographer. Extensive experiments on a dataset with over 1.31M samples demonstrate that the VA-Adapter outperforms strong probe guidance models while requiring approximately 33 times fewer trained parameters. Code is available at https://github.com/LeapLabTHU/VA-Adapter.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0575_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/VA-Adapter

Link to the Dataset(s)

N/A

BibTex

@InProceedings{WanTen_VAAdapter_MICCAI2026,
        author = { Wang, Teng AND Jiang, Haojun AND Wang, Yuxuan AND Sun, Zhenguo AND Deng, Yujiao AND Song, Shiji AND Huang, Gao},
        title = { { VA-Adapter: Adapting Ultrasound Foundation Model to Echocardiography Probe Guidance } },
        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 introduces VA-Adapter, an efficient module that equips pre‑trained ultrasound foundation models with probe‑guidance capability for echocardiography. By inserting the adapter into deep layers of a frozen image encoder and modelling historical vision‑action sequences with a transformer‑based interaction module, the method learns subject‑specific 3D cardiac anatomy. Experiments on a large dataset show that VA-Adapter reduces navigation error.

  • 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 idea of repurposing diagnostic foundation models for probe guidance via a lightweight adapter that explicitly models vision‑action sequences is novel. Most prior work trains guidance models from scratch or uses clinician’s statistical experience with visual guidance.

    2.With only 2.6M trainable parameters, the method achieves strong performance, making it attractive for resource‑constrained clinical settings.

    3.The paper incorporates a large dataset with image‑action pairs and trajectories obtained from robot teleoperation, which is substantial and also covers ten standard echocardiographic planes.

  • 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 evaluation is mainlu metric‑based. There is no user study to measure success rate, time to target plane, or image quality. Without such validation, the real‑world impact remains unclear.

    2.The model uses only L=4 frames. The effect of longer histories or adaptive sampling is not explored, and the paper does not discuss how robust the method is to different temporal granularities.

  • 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 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.

    (5) Accept — should be accepted, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The paper introduces a novel and parameter‑efficient VA‑Adapter that successfully adapts ultrasound foundation models for probe guidance, achieving strong accuracy with only 2.6M trainable parameters. The evaluation is thorough, including comparisons with strong baselines, extensive ablations, and real‑time latency analysis on a large dataset. While dataset diversity and clinical validation are limited, the core technical contribution is significant, well‑validated, and clearly presented.

  • 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 #2

  • Please describe the contribution of the paper

    This paper proposes VA-Adapter, a task-specific parameter-efficient module that adapts ultrasound foundation models such as EchoCLIP, USFM, and BiomedCLIP for echocardiography probe guidance. The method keeps the pretrained image encoder frozen, inserts VA-Adapters into its deeper layers, and processes short history sequences of sampled images and relative probe motions by interleaving visual and action tokens in a bottleneck space. These are updated through a vision-action interaction module, then aggregated by a GRU and mapped to 10 target-specific heads that predict 6-DoF motion from the current frame to each one of the 10 standard cardiac views. On a dataset of 356 expert scans from 178 adults (1.31M image-pose pairs), the method reduces translation and rotation MAE compared with strong single-frame and sequential baselines while using about 33× fewer trainable parameters and adding only minimal inference latency.

  • 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 vision–action adapter design that injects interleaved action tokens into deeper layers of a frozen foundation model is innovative and well-motivated for capturing patient-specific 3D structure from 2D sequences.
    • The residual mapping of adapted visual tokens back to the backbone feature space and progressive propagation of action tokens across adapter layers is a thoughtful architectural choice that balances adaptation and preservation of pre-trained knowledge.
    • Segmental sampling of histories encourages larger inter-frame motion and richer anatomical variation compared to consecutive frames, which is well aligned with the goal of structure inference.
    • The paper evaluates across three different ultrasound foundation backbones and compares PEFT strategies (LoRA, Prefix Tuning, vanilla adapters) with consistent gains for VA-Adapter.
    • Ablations on the interaction module and adapter dimension provide insight into where improvements come from and how capacity impacts performance/efficiency.
    • Inference-time profiling shows the proposed adapters add marginal latency and meet real-time needs on commodity GPUs.
    • The overall method is clearly described, with a coherent forward pass across interleaved vision–action tokens and a clean training strategy (frozen backbone, tuned adapters).
    • Dataset acquisition and label formation (relative action to standard views via recorded poses) are explained in a reproducible way.
    • Tackling operator dependence in echocardiography via learning 3D structure from 2D sequences is a clinically relevant problem.
    • Demonstrating that ultrasound foundation models can be efficiently adapted to probe guidance meaningfully connects two active research directions (diagnosis and acquisition/navigation).
  • 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 argues that VA-Adapter enables the model to infer patient-specific 3D cardiac structure from vision-action history, but the evidence is entirely downstream: lower action-prediction error. There is no intermediate analysis showing that the learned representation actually encodes anatomy, spatial state, or geometric consistency better than the alternatives. So the structural-reasoning interpretation is plausible, but not really proven by the presented experiments.
    • The metric is purely MAE on 6-DoF action vectors; there is no closed-loop evaluation (e.g., success rate to reach target plane, steps-to-target) or user study to establish clinical relevance of the observed error reductions.
    • The rotation parameterization and associated MAE in degrees are not specified (Euler vs. axis–angle), raising concerns about discontinuities and comparability across poses.
    • The equal weighting of translation (mm) and rotation (deg) by normalization may not reflect their differential impact on view quality and acquisition success.
    • No external validation: single-center, single scanner/probe, two experts. Generalization to different vendors, probes, operators, and patient populations (e.g., high BMI) is untested.
    • Absence of per-plane results and dispersion measures (e.g., SD/CI) makes it difficult to assess robustness and whether gains are uniform across views.
    • The sequential baselines’ configurations are not fully specified (where/how sequence fusion occurs, whether encoders are frozen, and whether parameter counts are matched), complicating fairness assessments.
    • Limited ablations: no study of sequence length L, history sampling strategies (segmental vs. consecutive), or adapter placement depth.
    • While general baselines are covered, related ultrasound guidance works using quality assessment signals or robotic autonomy (e.g., QA-NET-style guidance, recent compact autonomous liver US systems) are not discussed, and clinical navigation success metrics commonly used in robotic US are absent.
    • Fig. 1(c) is too small and should be enlarged to ensure the details are clearly visible. Additionally, Fig. 1(a) and (b) do not need to be presented as figures; the statistics of the public FM dataset and the private dataset could be more effectively conveyed in the text or summarized in a table.
    • Fig. 2 should present the architecture more clearly. It is currently unclear whether each pair of consecutive red layers is connected via one green linear layer and one VA-Adapter; if that is the intended design, it should be shown explicitly in the figure. Additionally, the indexing of the red layers is inconsistent (k in the left panel vs. i in the right) and should be unified. If the right panel is intended as a zoomed-in view of the left, this relationship should be clearly observable.
    • The figures would benefit from more detailed and informative captions that better explain their components and relationships.
    • Fig. 2 appears incomplete, as the final output of the model is not shown. For completeness, consider including the predictor head, the model’s final output, and the associated loss functions in the diagram.
  • 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 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

    This is a strong submission with a real technical idea and a clinically meaningful target task. The strongest evidence in the paper is the ablation against the vanilla adapter and the matched-backbone comparison against full fine-tuning. Those experiments make a convincing case that the gain is coming from the proposed vision-action interaction design rather than from a generic PEFT effect.

    My main recommendation is to tighten the framing of what the paper has shown. The experiments support that VA-Adapter improves offline sequential action prediction for target-plane guidance. They do not yet fully establish robust closed-loop guidance, nor do they directly verify the stronger claim that the model has learned an individualized 3D structure. I would therefore encourage the authors to present that interpretation a bit more cautiously.

    The paper would also benefit from a more explicit acknowledgment that MAE is only a proxy for guidance success. Even a short discussion of what level of translation/rotation error is likely clinically acceptable would help the reader interpret the reported gains. Finally, since the dataset is one of the important parts of this work, more details on subject diversity and hard cases would strengthen the paper.

    Questions for Authors

    • What exact blocks/layers receive adapters for each backbone (EchoCLIP CNN, USFM/BiomedCLIP Transformers)? Please provide a schematic/table so others can reproduce your insertion strategy.
    • How were the sequential baselines configured (fusion point, frozen vs. fine-tuned encoders, GRU settings)? Can you clarify why “sequential fully tuning” sometimes underperforms freezing?
    • Did you evaluate different sequence lengths L or compare segmental sampling against consecutive frames quantitatively? How sensitive is performance to L and the sampling scheme?
    • How many standard-plane annotations per subject are available, and how do you compute targets when some planes are missing or sparsely visited?
    • Can you provide any closed-loop proxy evaluation (e.g., iterative application of predicted actions with pose retrieval, steps-to-target, success rate) or phantom tests to link MAE to navigational success?
    • How does the model generalize across operators, scanners, and probes? Any evidence from leave-one-operator-out or cross-vendor splits?
    • How are translation and rotation scales normalized for the loss? Did you explore alternative weighting or a multi-task uncertainty weighting scheme?
  • 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?

    This paper presents a timely and well-executed approach to adapt ultrasound foundation models for echocardiography probe guidance via a novel vision-action adapter. The method is conceptually elegant, computationally efficient, and empirically strong on a sizable in-house dataset, with consistent gains over solid baselines and PEFT alternatives. However, the evaluation focuses on pose MAE without closed-loop navigation metrics or external validation, and baseline configurations need clearer specification to establish fairness fully. The work’s potential impact is high, but additional evidence, per-plane robustness, sequence-length/sampling sensitivity, rotation handling, and some form of closed-loop or prospective validation, would substantially strengthen the case for clinical relevance. I view this as a strong, innovative contribution, provided the authors address the key concerns and clarify experimental details.

    Also, please remove the last part of your introduction regarding your other anonymized related manuscript (Ultrastar), and also remove it from the references, as it is not cited in this paper. “We also submit an anonymized related manuscript [2] (in the supplementary material) that investigates sequence modeling for probe guidance. The two submissions are complementary and non-trivially different in methodological focus and experimental emphasis.”

  • 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



Review #3

  • Please describe the contribution of the paper

    This paper introduces a VA-adapter module that can be combined with previously developed foundation models for echocardiography for the probe guidance task. Probe guidance data collection is challenging because it must be acquired prospectively, but foundation models for ultrasound were trained on millions of retrospective images, so this paper explores if we can utilize them as a transfer learning technique. VA-adapter is added in the deep layers of foundation models, and is trained to encode images together with timestep embeddings and the action pairs. Image encoder takes in a current frames and samples 4 past frames and produces latent vectors for each 5 frames, they are subsequently fed into a GRU head which predicts 10 different heads denoting 10 different views and how to move the probe to get to them. As a result, they show that their sequential approach combined with transfer learning from previous foundation models outperform single-frame-no-transfer-learning approaches, and that adapter fine-tuning for this task is superior to PEFT fine-tuning strategies.

  • 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.
    • Innovative adaptation of ultrasound foundation models for the task of probe guidance
    • Strong ablations and comparison between different foundation models and different PEFT strategies
  • 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.
    • Visualizations of ‘Retrieved Plane’ in Figure 6 is a bit misleading because your search space is limited to only the sampled frames (you take KNN neighbor to retrieve the plane). You should provide how big of a difference was between a predicted Trans and Rot and a nearest neighbor.
    • Outdated sources, no source is from 2025.Please add a comparison metrics or discussion with EchoWorld https://arxiv.org/abs/2504.13065 and Cardiac Copilot https://arxiv.org/html/2406.13165v1
    • The paper argues that the main benefit is using pre-trained weights from previous foundation models. To support this argument, the authors should present an ablation where the same architectures are used but weights are initialized either at random or from ImageNet.
  • 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 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?

    It actually makes the most sense to do only USFM because it’s trained as masked autoencoder, for CLIP (EchoCLIP, BioMedCLIP) they are trained to learn concepts not necessarily structures visible so potentially they don’t see the structure.

  • 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?

    This paper would be of interests to the community because it describes an effective adaptation of previously developed foundation models for a new task - probe guidance. However, the authors need to provide more evidence (as explained above) in their rebuttal to strengthen main claims.

  • 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 the reviewers’ constructive comments and suggestions. We have carefully revised the manuscript and address the main concerns below. <R1,R2: Offline MAE Metric and Clinical Relevance> In the field of ultrasound probe navigation, evaluating model performance via offline proxy metrics is a prevalent practice [1-3]. As a practical metric, offline MAE strongly correlates with navigation quality and serves as a prerequisite for closed-loop success. For a study focused on methodological innovation, this metric directly and rigorously quantifies the model’s decision-making precision at the algorithmic level. We agree that closed-loop validation is an excellent suggestion to further enhance clinical impact. We are actively overcoming hardware and ethical barriers to evaluate real-world closed-loop performance for a future journal extension. Regarding the clinical relevance of the observed error reductions, key cardiac anatomical landmarks are on the millimeter scale, even minor mm/° improvements are clinically meaningful. As a clinical reference, the average expert repositioning error is reported to be 5.7° [2]. By reducing the rotation error from approximately 7.8° to 6.7°, our method significantly narrows the gap to expert-level proficiency. <R1,R2: Sequence Length and Sampling Strategy> To ensure real-time clinical navigation, we set L=4 as a practical trade-off between capturing sufficient context and maintaining low inference latency. We adopted segmental sampling over consecutive sampling because the latter causes minimal inter-frame motion and high redundancy, hindering 3D structural understanding. By dividing the history into equal segments and randomly sampling one frame per segment, we effectively mitigate redundancy and capture richer cardiac information. Regarding robustness to temporal granularities, our sampling mechanism exposes the model to highly diverse temporal granularities and large spatial spans during training. Furthermore, our model relies on the actual relative probe motion between sampled frames rather than temporal indices, which intrinsically guarantees its robustness to temporal granularities. <R2: Generalization> To assess patient-level generalization, we evaluated our method on a large validation cohort composed entirely of unseen subjects, including male and female subjects, ages ranging from 18 to 69, BMIs ranging from 19.0 to 34.6, and diverse cardiac pathologies. This setting provides a direct evaluation of patient-level generalization. For device-level generalization, the main impact of different ultrasound systems is typically reflected in low-level image appearance. During our data acquisition, sonographers adjusted imaging settings adaptively for individual patients, which introduced substantial appearance diversity into the training data. In addition, recent work [4] has shown that a navigation model trained on 247,000 images from a single device can generalize to other ultrasound devices; in comparison, our model is trained on over one million images. We will add multi-center external validation to future work. <R2: Robustness across views> In our original experiments, we performed a comprehensive evaluation across each individual view to ensure the system’s reliability. While the manuscript reported averaged metrics to provide a more concise overview, our internal data from the initial assessment confirms that the performance gains are indeed uniform: the VA-Adapter consistently outperforms the baselines across all 10 standard views in both translation and rotation dimensions. We will add a detailed explanation in the revision. Due to space constraints, we cannot address every minor comment here. However, we sincerely value all remaining suggestions and will meticulously incorporate them into the camera-ready version. [1] Pose-GuideNet, MICCAI’24 [2] Real-world Visual Navigation for Cardiac Ultrasound View Planning, MICCAI’24 [3] EchoWorld, CVPR’25 [4] UltraBot, Nat. Commun.’25



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.

    The reviews are consistently positive, with scores of 5, 4, and 4.The paper presents VA-Adapter, a parameter-efficient approach for adapting ultrasound foundation models to echocardiography probe guidance. The strengths are the clear technical contribution, strong empirical performance on a sizable robotic echocardiography dataset, comparisons across multiple foundation backbones and PEFT methods, useful ablations, and low trainable-parameter/runtime cost.

    The main weaknesses are the lack of closed-loop or user-study validation, the absence of external validation, reliance on offline translation/rotation MAE as a proxy for guidance success, and limited analysis of robustness across views, sequence lengths, and acquisition settings. These are important limitations, but they do not outweigh the technical contribution and positive reviewer consensus. I recommend acceptance.



back to top