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

Abstract

Trackerless freehand 3D ultrasound reconstructs volumes from 2D B-mode sequences by estimating probe motion without external tracking. In B-mode ultrasound, accurate motion estimation hinges on recognizing subtle inter-frame appearance changes, yet repetitive speckle and weak visual anchors make correspondences ambiguous and drift accumulates when composing relative transforms. We propose a correspondence-amplifying conditioning pipeline with a flow-matching transformer for pose-delta sequence generation. For each adjacent frame pair, we build a bidirectionally supported token correspondence distribution from ViT features, use it to align current-frame evidence onto the previous token grid, and derive a correspondence-induced lattice displacement cue. We further apply entropy-based confidence weighting to suppress ambiguous regions and aggregate high-dimensional correspondence-aware tokens into compact context tokens via a Q-Former. Conditioned on this context, a flow-matching DiT generates coherent pose-delta sequences through ODE sampling. Experiments on tracked freehand ultrasound sequences show improved step accuracy and substantially reduced long-horizon drift.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/6350_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

TUS-REC2024 dataset: https://github-pages.ucl.ac.uk/tus-rec-challenge/TUS-REC2024/data.html

BibTex

@InProceedings{KimSan_Trackerless_MICCAI2026,
        author = { Kim, Sang-yun AND Oh, Seok-Hwan AND Kim, Myeong-Gee AND Kim, Young-Min AND Jung, Guil AND Lee, Hyeonjik AND Son, Jungjae AND Kwon, Hyuk-Sool AND Bae, Hyeon-Min},
        title = { { Trackerless Ultrasound Pose Estimation via Correspondence-Aware Tokenization and Flow-Matching Transformers } },
        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 proposes a correspondence-aware conditioning framework for trackerless freehand ultrasound pose estimation, where explicit token correspondences, entropy-based confidence weighting, Q-Former compression, and a flow-matching DiT are combined to predict pose-delta sequences.

  • 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 problem is important and well motivated. The proposed correspondence-aware tokenization is technically sensible and well aligned with the challenges of ultrasound B-mode imagery. Quantitative results show clear gains over the baseline and improvements over reported trackerless methods on TUS-REC.

  • 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.

    Limited clinical generalizability:The algorithm is validated only on a single forearm dataset (TUS-REC). It lacks evaluation on more clinically challenging anatomies that are substantially affected by respiratory motion and tissue deformation. Lack of fine-grained ablation analysis:The ablation study provides results for Baseline, Baseline + CAT, and Baseline + CAT + QF, which partially demonstrates the benefit of progressively adding the proposed components. However, it still does not fully isolate the contribution of QF, since there is no comparison against alternative compression strategies or a more detailed analysis of whether the gain comes specifically from the query-based aggregation mechanism. Therefore, while the added value of CAT is relatively clear, the necessity and distinct contribution of QF remain less well justified.

  • 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?

    This paper addresses an important problem in trackerless freehand ultrasound pose estimation and proposes a technically coherent framework that combines explicit correspondence modeling, confidence weighting, Q-Former compression, and flow-matching DiT in a well-motivated way. The reported results show clear improvements over both the baseline and prior trackerless methods, especially in reducing long-horizon drift, which supports the practical value of the approach. At the same time, the current evidence does not yet fully support all major claims, since the validation is limited to a single dataset and the specific contribution of Q-Former is not yet fully isolated through finer-grained ablations. Overall, I find the work promising and technically interesting, with sufficient merit for acceptance, but I believe some of its claims would benefit from stronger validation and more detailed component analysis.

  • 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.

    The authors have clarified my main concerns regarding the role of different modules, and the evaluation protocol. The correspondence-aware tokenization module is the core contribution of this work, and its motivation and effectiveness are now clearer. In particular, explicitly modeling inter-frame correspondences is well aligned with the challenges of B-mode ultrasound pose estimation and provides meaningful improvements in drift reduction. My remaining concern is that the clinical generalizability is still not fully demonstrated. I hope the authors can further discuss this limitation and investigate the method on more diverse clinical ultrasound settings in future work. Overall, I think this paper can be accepted. 23.After reading the rebuttal, I think the authors have adequately clarified my main concerns. The core contribution, correspondence-aware tokenization, is technically meaningful. Although the current validation is still limited to one dataset and one anatomical scenario, I view this more as a limitation for future work rather than a fatal weakness. Overall, I support acceptance of this paper.



Review #2

  • Please describe the contribution of the paper

    Explicitly construct and amplify inter-frame correspondences​ by building a mutual, row-stochastic token correspondence distribution from ViT features, aligning current-frame evidence onto the previous token grid, and deriving a correspondence-induced lattice displacement cue. This makes subtle motion-related appearance changes explicit, overcoming the ambiguity caused by repetitive speckle and weak visual anchors. Integrate entropy-based confidence weighting​ to suppress unreliable regions (e.g., homogeneous speckle) by downweighting low-confidence correspondences, ensuring only robust motion evidence dominates the conditioning signal. Adopt a Q-Former​ to distill high-dimensional correspondence-aware tokens into compact context tokens, efficiently aggregating task-relevant spatial information for pose generation. Use a flow-matching DiT​ conditioned on these context tokens to generate coherent pose-delta sequences via ODE sampling, improving stability and reducing long-horizon drift.

  • 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 most significant methodological novelty lies in explicitly constructing and amplifying inter-frame correspondences before pose estimation—a departure from prior works that relied on networks to implicitly discover patch-level alignments.

    2.To handle ambiguous regions (e.g., homogeneous speckle or low-texture areas), the authors introduce entropy-based confidence weighting for correspondence-aware tokens.

    3.The pose-delta sequence is generated using a flow-matching Diffusion Transformer (DiT), a novel application of generative modeling to ultrasound motion estimation.

  • 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.Although the integration​ of components into a correspondence-amplifying conditioning framework is novel, some individual elements have precedents in prior work: ViT Feature Extraction & MAE Pretraining: The use of a frozen ViT-B pretrained via masked autoencoding (MAE) for ultrasound feature extraction builds directly on established practices in medical imaging (e.g., He et al., 2022). While effective, this choice is not unique to this work. Flow Matching for Generative Modeling: The application of flow matching (Lipman et al., 2022) to pose-delta generation is innovative in the ultrasound domain, but flow matching itself is a general-purpose generative technique previously applied to other domains (e.g., image synthesis, molecular dynamics). The novelty here lies in adapting it to ultrasound motion estimation, not in inventing the method. Q-Former for Context Compression: The Q-Former (Li et al., 2023) was originally proposed for language-image pre-training. Its use here to compress correspondence-aware tokens is a reasonable adaptation but not a fundamentally new contribution. Implication: The paper’s strength is in the orchestration of these components for a specific ultrasound challenge, but individual parts may lack novelty when viewed in isolation.

    2.The method relies on a frozen ViT-B backbone​ pretrained on B-mode images via MAE. While this leverages powerful general-purpose features, it introduces two limitations: Domain Adaptation Gap: The pretrained ViT is not fine-tuned on the specific TUS-REC dataset or ultrasound-specific tasks beyond MAE. Prior work (e.g., Guo et al., 2022) has explored end-to-end training of feature extractors for ultrasound motion estimation, which may better adapt to dataset-specific nuances (e.g., speckle patterns, probe orientations). Computational Trade-off: Freezing the ViT avoids retraining costs but may limit performance gains achievable through joint optimization of feature extraction and pose estimation. The paper does not ablate the impact of freezing vs. fine-tuning, leaving open whether end-to-end training could further improve results.

    3.The TUS-REC dataset includes characteristic sweep shapes (L-, C-, S-shaped), and the method is optimized for these. However: Generalization to Irregular Trajectories: Uncommon or irregular scan paths (e.g., spiral, random) are not explicitly tested. Prior work (e.g., Li et al., 2023) has noted that sequence modeling can struggle with out-of-distribution trajectories, which may affect the method’s applicability in freehand scanning scenarios with unpredictable operator behavior.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

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

    In summary, the paper’s strengths lie in its novel, problem-specific methodology​ (explicit correspondence amplification, confidence weighting, Q-Former compression, flow-matching DiT), robust design choices​ tailored to ultrasound’s unique challenges, and comprehensive evaluation​ demonstrating real-world impact. These contributions collectively advance the state of trackerless freehand 3D ultrasound reconstruction.

  • 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

    This paper proposes a novel method for predicting the relative pose between ultrasound frames by incorporating correspondence-aware tokenization, entropy-based confidence weighting, a Q-Former, and flow matching.

  • 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 paper proposes a novel method to address challenges such as repetitive speckle and weak visual anchors by explicitly constructing and enhancing inter-frame correspondences. The approach introduces correspondence-aware tokenization, incorporates entropy-based confidence weighting, and leverages a Q-Former together with flow matching.

    The Results section includes an ablation study of each proposed component and provides comparisons with several methods from prior work.

  • 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 main weaknesses relate to the data split and the choice of comparison methods. The test set is too small to convincingly demonstrate the effectiveness of the proposed method, and the comparisons are primarily against lower-ranked teams of the challenge from which the dataset is derived rather than top-performing ones. These issues do not provide sufficient evidence to support the claimed performance. In addition, the Methods section is difficult to follow, with several terms (e.g., “ODE”) not clearly defined.

  • 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

    1.What does “pose deltas” refer to? Is this a standard term in the field? If not, it would be helpful to include a brief explanation. 2.Section 2.1: T_k denotes the pose for each frame. Could the authors clarify what this pose represents, for example, the coordinate system in which it is defined and the transformation it encodes? 3.The definition of H_f and W_f is missing and should be clearly specified. 4.Equation (4) is not clearly presented. Additional details or clarification would help improve understanding. 5.What does “ODE” refer to in Section 2.4? Please define the term when it is first introduced. 6.Section 3.1: In the experiments, the test set includes only three subjects, which is too small to reliably demonstrate the effectiveness of the proposed method. 7.The authors re-implemented several methods from prior work; however, the manuscript lacks a clear textual comparison between these methods, and no statistical significance analysis (e.g., p-values) is provided. Also, std is missing in Table 1.8.The performance of the proposed method is promising; however, the comparisons are made only against the fourth- and fifth-ranked teams. Why were the top three teams not included in the comparison?

  • 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?

    Although the method is promising and interesting, the presentation in the Methods section is difficult to follow. Moreover, the experimental design is not sufficiently rigorous. The dataset split is not well justified, with only three subjects in the test set, and the comparisons do not include the top-ranking methods for this challenge dataset.

  • Reviewer confidence

    Confident but not absolutely certain (3)

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

    Reject

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

    Although the 72 test scans cover different scan settings, they are collected from only 3 subjects. This makes the evaluation highly sensitive to subject-specific anatomy, acquisition variability, and noise, and may not provide sufficient evidence that the reported gains would generalise beyond these subjects.

    This concern is particularly important for the ablation study. Since the claimed improvements of individual modules are evaluated on the same small three-subject test set, the observed differences may reflect variability in this particular split rather than robust contributions of the proposed components. A subject-disjoint protocol is necessary, but not sufficient, to prove that the observed improvements between methods or ablation variants are reliable when the test set contains very few subjects.

    The authors could consider re-splitting the available data rather than strictly following the original split. Re-conducting the experiments with a more meaningful split would make the results more convincing.



Author Feedback

We thank the reviewers and meta-reviewer for thoughtful feedback. We appreciate that they identified our focus on ultrasound-specific challenges and explicit inter-frame correspondence modeling as key strengths, aligning with our design intent. We address the concerns below.

[Meta,R3] Dataset split and test-set size We used a subject-disjoint split with publicly labeled TUS-REC2024 sequences: 1,080 (45 subjects) for training, 120 (5 subjects) for validation, and 72 (3 subjects) for testing in Table 1.The 72 test sequences are the official public validation set and were never used for training. They cover both arms, scan plane orientations, scan directions, and L/C/S trajectory shapes. All ablations and comparisons were evaluated under the same subject-disjoint protocol, so the relative comparison is appropriate for assessing CAT. If accepted, we will clarify this dataset description.

[Meta,R3] Comparative baseline selection criteria To ensure a fair and reproducible comparison, we selected baselines relevant to trackerless freehand ultrasound pose estimation that had public assets and could be evaluated under the same subject-disjoint split. We therefore included recent trackerless approaches and strong TUS-REC2024 dataset-based methods. All comparative methods were trained and evaluated under our split for a fair comparison. Under this protocol, the proposed method improves over the strongest reproduced baseline in Table 1. [R1,R2] Role of CAT, Q-Former, and flow matching CAT is the primary methodological contribution. It constructs inter-frame token correspondences, aligns current-frame evidence to the previous token grid, derives lattice-displacement cues, and downweights uncertain matches. Q-Former and flow-matching DiT are complementary components fitting this CAT representation. QF compresses dense CAT tokens via learnable query aggregation while retaining task-relevant evidence. Flow-matching DiT decodes the context into pose-delta sequences. In the ablation, Baseline+CAT uses linear layer compression, while Baseline+CAT+QF uses learnable query-based compression. The larger gain from Baseline to Baseline+CAT shows CAT is the dominant source and CAT+QF further improves GPE and FDR.

[R2] Frozen B-mode ultrasound-pretrained ViT encoder As stated in the manuscript, the feature encoder is a large-scale B-mode ultrasound pretrained ViT. The MAE objective was used for ultrasound-domain representation learning, so tokens are aligned with B-mode representations. We froze this encoder to keep the correspondence space stable and to mitigate overfitting on the limited pose-labeled TUS-REC split.

[R1,R2] Scan-protocol diversity and clinical potential TUS-REC provides realistic in-vivo freehand scans under standardized yet diverse protocols reflecting common ultrasound scanning conditions. Each subject includes 24 scans spanning both forearms, scan plane orientations, bidirectional sweeps, and L/C/S trajectories. Thus, the 72-sequence test set evaluates CAT across multiple scan-protocol variations. By converting subtle inter-frame B-mode appearance changes into explicit correspondence evidence, CAT provides an image-only motion representation that is not restricted to a particular anatomy or scan trajectory. Together with the multi-anatomy B-mode pretrained encoder, CAT is a flexible conditioning mechanism for attention-based inter-frame ultrasound pose models and opens the possibility of broader clinical freehand reconstruction scenarios.

[R3] Clarifications “Pose delta” is the relative rigid transform between consecutive calibrated ultrasound image planes, Delta T_k=T_{k-1}^{-1}T_k. T_k is the calibrated frame pose in tracker coordinates. H_f,W_f are the ViT token-grid height/width. Eq. (4) applies row-/column-wise softmax, followed by mutual support and row normalization in Eqs. (5-6). ODE means ordinary differential equation; inference integrates the learned flow with Euler steps.



Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    This paper proposes a markerless pose tracking algorithm for handheld ultrasound manipulation. The main technical contribution lies in establishing local correspondences via ViT features and then processing them through attention mechanisms that selectively value more highly the correspndences from meaningful motions. The method is tested on the TUS-REC challenge dataset, comparing it against different recent baselines.

    All reviewers agree that the technical contribution is valuable. However, R3 raises concerns about the experiment protocol. I would recommend a rebuttal so that authors can argue for their current choice of baselines and data split in relation to the TUS-REC challenge results (noting that running new experiments is out of scope for rebuttal and shouldn’t be considered)

  • 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.

    Reviewer disagreement remains after rebuttal , with 2 accepts and 1 reject.

    The main unresolved issue is related to the test data containing sequences from only 3 subjects. While the limitations are understandable, the authors are following the evaluation protocol from TUS-REC2024, making their work more reproducible and comparable to prior literature. In my assessment this is a reasonable choice in the context of a page-limited MICCAI submission - a small test set that is directly comparable to prior literature and easily reproducible is a reasonable choice vs. a larger test set with a non-standard protocol not comparable to prior literature.

    A larger publication would enable to test both cases.



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.

    This paper remains borderline after rebuttal. Despite the limitations in the experimental setting, the novelty has been recognized by all the reviewers, which motivates my support for acceptance.



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.

    After the rebuttal, one reviewer maintained an Accept recommendation, and another maintained a Reject recommendation. The third reviewer, who initially provided a Weak Accept recommendation, did not provide an updated review. While the paper has limitations in the evaluation of clinical generalizability, it also demonstrates several strengths, including a well-motivated technical contribution and promising experimental results. Therefore, the paper is recommended for acceptance to MICCAI 2026.The authors are encouraged to incorporate the necessary changes discussed during the rebuttal process into the final camera-ready version.



back to top