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

Abstract

Tubular tree structures such as blood vessels and lung airways are central to many clinical tasks, including diagnosis, treatment planning, and surgical navigation. Accurate centerline extraction with correct topology is essential, as missing small branches can lead to incomplete assessments or overlooked abnormalities. We propose RefTr, a 3D image-to-graph framework that generates tubular tree centerlines via recurrent refinement of confluent trajectories. RefTr adopts a Transformer-based Producer–Refiner architecture in which the Producer predicts candidate trajectories and a shared Refiner iteratively aligns with target branches. The confluent trajectory representation enables whole-branch refinement while explicitly enforcing valid topology. This recurrent scheme improves precision and reduces decoder parameters by 2.4× compared to the state-of-the-art. We further introduce an efficient non-maximum suppression algorithm for spatial tree graphs to merge duplicate branches and extend evaluation metrics to be radius-aware for robust comparison. Experiments on multiple public datasets demonstrate higher recall and precision, faster inference, and substantially fewer parameters, highlighting the effectiveness of RefTr for 3D tubular tree analysis. Code is available at https://github.com/RomStriker/RefTr.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1612_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/RomStriker/RefTr

Link to the Dataset(s)

Trexplorer Super Centerline Datasets: https://zenodo.org/records/15888958 ATM’22: https://atm22.grand-challenge.org/ PARSE 2022: https://parse2022.grand-challenge.org/

BibTex

@InProceedings{NaeRom_RefTr_MICCAI2026,
        author = { Naeem, Roman AND Hagerman, David AND Alvén, Jennifer AND Kahl, Fredrik},
        title = { { RefTr: Recurrent Refinement of Confluent Trajectories for 3D Tubular Tree Centerlines } },
        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 work proposes a transformer based encoder-decoder model for extracting vascular centerlines from 3d volumetric scans using confluent trajectories where several partial branches are over-segmented, and merged using a specialised algorithm. The encoder consists of a SwinUNETR which outputs image features, that are used by the producer block to output several partial branches which are successively refined using a refiner network over many steps. Finally, a Tree non-maxiumum suppression (TNMS) algorithm is used to remove duplicate nodes/branches and appropriate bifurcations are introduced to obtain valid centerlines. Experiments are reported on three datasets and compared with relevant, recent baseline methods. The proposed method shows strong performance across datasets and most reported metrics while also being relatively efficient compared to the reported 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 proposed method of obtaining many confluent trajectories, refining and merging them using a self-attention is a novel contribution. This harps back to methods that used multiple hypothesis tracking for airway segmentation [0] (but is not discussed in the paper). The use of successive refinement and merging can improve overall centerline extraction capabilities as demonstrated in the evaluations.

    • Fig 1 is a useful illustration that clearly shows the different components and how they interact in the method.

    • Experimental evaluation is thorough, and the performance of the proposed method is competitive across datasets and metrics compared to the baseline methods.

    • Reporting ablation and efficiency metrics also shows further competitiveness of the method.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    • How are the token sets learned? The image encoder provides the image features to the producer, but where do the “learned input token sets” come from? How are they learned? At s=1 already, the tokens are learned which are only further refined by Producer and Refiner blocks. So, the input learned tokens are fixed?

    • The loss objective in Eq. 1 has several weights (alphas). No further discussion on how these were tuned and what the final values are is presented. I am also not sure I understand the different components as they are not described in sufficient detail.

    • TNMS is not described in detail. The authors say duplicate nodes are identified “using a radius-adaprive spatial threshold”. As far as I could check, there are no additional details provided about this method or the related paramters. And TNMS appears to be an important element going by Table 4 ablation.

    • The title says “vascular tree” only whereas the method is validated also for airways in ATM22 dataset [1]. Are the authors sure the centerlines in this dataset are of the vascular tree and not airway tree?

    • Why mention the concurrent submission and provide it as supplementary material? What is EVAR? It is unclear what the specific point the authors are trying to make with this claim. Paper should be self-contained.

    [0] Selvan et al. “Extracting tree structures in CT data by tracking multiple statistically ranked hypotheses. “ Medical Physics 46.10 (2019): 4431-4440. [1] https://atm22.grand-challenge. org/

  • 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.

    (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 is a good contribution but there are some points that need clarification which I think can be done during rebuttal.

  • 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.

    Rebuttal addresses many concerns raised in my review, and also other reviewers. I am inclined to accept the paper.



Review #2

  • Please describe the contribution of the paper

    This paper proposes RefTr, a Transformer-based image-to-graph framework for 3D vascular centerline extraction. The method introduces a Producer–Refiner architecture that iteratively refines “confluent trajectories,” aiming to improve centerline construction and enforce valid topology. The method is evaluated on multiple public datasets and compared primarily against the authors’ prior work (Trexplorer, Trexplorer Super), showing improvements in several metrics along with reduced parameter count and runtime. The problem is important and clinically relevant, and the paper is generally well written with solid experimental effort. However, the methodological novelty appears limited, and there are concerns regarding clarity, evaluation consistency, and comparison fairness that weaken the overall contribution.

  • 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.
    • Relevant and important problem: Accurate and topologically correct centerline extraction is critical for many clinical applications.
    • Technically sound framework: The proposed architecture is coherent and builds on a reasonable design (Transformer-based image-to-graph modeling).
    • Comprehensive experiments: o Multiple datasets (synthetic + real) o Multi-level evaluation (node / branch / graph) o Runtime and parameter comparisons o Ablation study covering key components
    • Efficiency improvements: The model achieves reduced decoder parameters and faster inference, which is practically meaningful.
  • 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) Limited novelty and incremental contribution (Major) The proposed method is heavily built upon the authors’ prior work (Trexplorer / Trexplorer Super), and the incremental contribution is relatively small:

    • The “recurrent refinement” mechanism appears to be a natural extension of iterative decoding already present in prior work, rather than a fundamentally new idea.
    • Many-to-one matching is a standard technique (e.g., DETR variants).
    • TNMS is a simple engineering extension (duplicate merging + pruning), with limited algorithmic novelty.
    • The main conceptual addition, “confluent trajectories,” is not sufficiently formalized or justified. Overall, the work reads more as an incremental refinement of an existing framework, with marginal gains relative to Trexplorer Super.

    (2) Unclear definition and overclaim of “confluent trajectories” (Major) The notion of “confluent trajectories” is central to the paper but:

    • The motivation of this choice is not convincingly explained and its concept is not rigorously defined. The current manuscript lacks: o formal definition o theoretical justification o comparison to alternative representations Moreover, the claim that this representation “explicitly enforces valid topology” is not fully supported:
    • Topology still relies on post-hoc reconstruction (divergence prediction + TNMS).
    • No formal guarantee of topology correctness is provided. This weakens the conceptual contribution.

    (3) Comparison bias toward authors’ own prior work (Major) The experimental comparison is heavily dominated by the authors’ previous methods (Trexplorer variants), which raises concerns about fairness and completeness. There is insufficient comparison to other categories of methods, including:

    • Topology-aware segmentation approaches (e.g., clDice-based methods: Shit et al., DOI: 10.1109/CVPR46437.2021.01629)
    • CNN-based centerline / tracing methods (e.g., DeepVesselNet: Tetteh et al., 2020, DOI: 10.3389/fnins.2020.592352) → more recent variants or follow-ups should also be considered
    • Skeletonization / topology-preserving learning approaches (Qiu, et.al, DOI:10.1016/j.media.2025.103578 or other recent works on differentiable skeletonization or topology constraints)
    • Recent graph-based or foundation models (e.g., VesselFM, Wittmann, et. al, DOI: arXiv:2411.17386) Given that the paper claims advantages in topology preservation, comparison against topology-preserving losses (e.g., clDice) is particularly important and currently missing.

    (4) Evaluation inconsistency with prior work (Major) The paper introduces new radius-aware metrics (rAP, rF1, etc.), which were not used in prior Trexplorer publications. This raises concerns:

    • Are results directly comparable to previously reported baselines?
    • Why are legacy metrics not reported alongside the new ones?
    • Could the new metrics favor the proposed method design? Additionally:
    • It is unclear whether the same dataset splits and evaluation protocols are used.
    • The performance metrics reported in this paper for comparison methods are different from previous publications. This limits the ability to fairly assess improvement over prior work.

    (5) Qualitative comparison inconsistency (Moderate–Major) The qualitative results (Fig. 2) show significantly weaker performance for prior methods compared to their original publications. This raises concerns about:

    • sample selection bias
    • comparability with previously reported visual results It would be more convincing to:
    • show the same cases as prior papers, or
    • include a broader set of representative examples

    (6) Weak justification for Parse dataset performance (Major) The authors attribute weaker performance on Parse 2022 to missing small branches in ground truth, supported by radiologist review. This explanation is not fully convincing:

    • The method appears worse overall, not only in recall.
    • The claim is not quantitatively validated. If ground truth is indeed flawed:
    • a stronger contribution would be to correct and release improved annotations, rather than relying on anecdotal justification.

    (7) Clarity issues in method description (Moderate)

    • The training and role of “learned input token sets” are not clearly explained.
    • The divergence prediction mechanism and topology reconstruction pipeline could be clarified.

    (8) Overlap with concurrent submission (Minor to Moderate) The authors provide a supplementary document that appears to be a separate concurrent MICCAI submission focusing on a clinical application (EVAR), which directly builds upon RefTr. While such parallel submissions are not uncommon, this raises potential concerns:

    • Double counting of novelty
    • Inflation of contribution through splitting method and application The methodological novelty is largely concentrated in the current paper, while the supplementary submission primarily demonstrates application-level extension. This is acceptable but should be considered during meta-review.
  • 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

    Suggestions for Improvement 1.Clearly define and formalize “confluent trajectories”, and justify how they differ from existing representations. 2.Provide a precise comparison vs Previous works (Trexplorer/ Trexplorer Super), ideally in tabular form, highlighting the true methodological delta. 3.Expand experimental comparisons to include: o topology-preserving segmentation o CNN-based tracing methods o skeletonization-based approaches 4.Report legacy evaluation metrics used in prior work for fair comparison. 5.Clarify dataset splits and ensure evaluation consistency. 6.Provide more representative qualitative comparisons, ideally using the same examples as prior work. 7.Address the Parse dataset issue more rigorously (or avoid speculative claims). 8.Improve clarity on: o learned token initialization and training o topology reconstruction process

  • 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?

    This is a technically solid but incremental work with meaningful engineering improvements. However, limited novelty, comparison bias, and evaluation inconsistencies prevent it from being a strong MICCAI contribution in its current form.

  • Reviewer confidence

    Very confident (4)

  • [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.

    The authors provided useful technical clarifications on some key issues, such as TNMS thresholds and same data splits as prior work. These responses improve reproducibility and readability. However, my main concerns remain largely unchanged.

    The authors articulate three sources of novelty. However, this does not substantially alter the novelty assessment. Trajectory reformulation is the main conceptual contribution, but still appears to be a reparameterization of the prior framework. Many-to-one inference ensembling is a practical extension, but conceptually straightforward. TNMS remains an engineering component. The claim of 2.4× parameter reduction and improved efficiency is valuable, but efficiency improvements do not by themselves establish strong methodological novelty.

    Baseline method is still one of my strongest concerns. The rebuttal argues that several omitted methods require voxel-level supervision and are therefore not directly comparable. However, this rationale appears overly restrictive. In particular, methods such as DeepVesselNet and related centerline-focused approaches do not necessarily require dense voxel labels and remain relevant references. More broadly, comparison is still concentrated on the authors’ own prior work, limiting the ability to assess performance against alternative methodological paradigms.

    While radius-aware metrics are conceptually reasonable, the proposed method does not consistently improve radius estimation accuracy, with worse Radius MAE reported across all datasets compared with their last model (Trexplorer Super) and even appear the worst in the Synthetic Dataset. This raises questions about whether the new metrics (considering radius as a factor) provide a more reliable assessment. Without a full verification on their newly proposed metrics, reporting full legacy metrics remains important to ensure conclusions are robust to the choice of evaluation metric. In the previous publication of Trexplorer Super, the Trexplorer Super has 0.5627 ± 0.0141 in radius MAE on Parse 2022 Dataset, but here in this paper, the value becomes 0.58 ± 0.025



Review #3

  • Please describe the contribution of the paper

    This paper introduces RefTr, a trajectory-level reformulation for 3D vascular centerline extraction that departs from conventional point-wise sequential tracking. The method represents centerline trees as confluent trajectories and refines them through a Producer–Refiner architecture. In addition, the paper proposes Tree Non-Maximum Suppression (TNMS) to suppress duplicate branches while preserving tree topology, and extends prior evaluation protocols with a radius-aware metric framework for more robust comparison across vessel scales. Extensive experiments on public datasets further suggest that RefTr can improve branch recovery and topological fidelity while reducing decoder complexity and inference cost.

  • 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 main strength of the paper lies in its reasonably clear methodological motivation. Rather than making only incremental architectural modifications, the work reformulates the reconstruction problem at the level of trajectory generation and recurrent refinement, which gives the method a degree of conceptual novelty. The experimental section is also relatively comprehensive, including multiple public datasets, point-/branch-/graph-level evaluation, efficiency comparisons, and ablation studies. In particular, the reported reduction in decoder parameters and runtime strengthens the practical relevance of the method. Finally, the ablation results broadly support the proposed design choices, especially the roles of recurrent refinement, many-to-one matching, past-memory attention, and TNMS.

  • 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.While the high-level concepts are understandable, the construction of the decoder, the organization of tags, and the specific roles of self-memory, cross-memory, and past memory attention are not clearly described, leaving the reader quite confused.

    2.Furthermore, some improvements seem to rely on a design loop: intentional over-prediction followed by suppression of duplicates via TNMS. This raises the question: how much of the gain stems from the core generation mechanism itself, and how much from post-processing?

    3.There is also insufficient justification for fixing the matching assignment from the first refinement step, as suboptimal early proposals may persist into subsequent refinement stages.

  • Please rate the clarity and organization of this paper

    Poor

  • 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.

    (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 offers a meaningful reformulation of vascular centerline extraction, and its main claims regarding effectiveness and efficiency are supported by reasonably solid experiments. But its presentation and evidential strength remain somewhat short of being fully convincing.

  • 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 rebuttal has addressed my main concerns regarding the method design and the role of TNMS. I maintain my positive recommendation.



Author Feedback

We thank the reviewers for their thoughtful feedback. Major concerns are addressed below.

1.Novelty (R2) RefTr’s novelty is threefold and acknowledged by R1 and R3: (i) trajectory-level reformulation: Trexplorer Super predicts one node at a time; RefTr jointly predicts and refines entire confluent trajectories; (ii) inference-time ensembling: many-to-one matching produces multiple candidates, merged at the patch level via pairwise divergence positions, boosting recall on small branches; (iii) TNMS: an O(N log N) algorithm removing residual duplicates globally in trees with thousands of nodes in under a second on CPU; no prior algorithm targets this task. Result: 2.4x decoder parameter reduction, faster inference, superior P/R across datasets.

2.Many-to-one vs DETR variants (R2) Group/H/Co-DETR use one-to-many only at training, discarding extras at inference. RefTr keeps and merges redundant predictions at inference, creating an inference-time ensemble, functionally distinct from DETR’s training-only mechanism.

3.Confluent trajectories and topology (R2) Formally defined in Sec. 2 and illustrated in Fig. 1(a). The representation (trajectories + end positions + pairwise divergence positions) defines a valid tree by construction; TNMS only merges duplicate branches. Empirically Betti-0 = Betti-1 = 0 on all datasets (Tab. 2).

4.Core mechanism vs TNMS (R2, R3) Table 4 ablates multiple core components (recurrent refinement, many-to-one matching, past-memory attention, and Producer/Refiner design choices), all with TNMS enabled. Removing any of these significantly degrades performance, establishing the core mechanism, not TNMS, as the source of the gains. Removing TNMS instead leaves recall essentially unchanged but reduces precision, since residual cross-patch duplicates are then counted as FPs. The core mechanism finds structures correctly; TNMS reclaims precision lost to residual duplicates.

5.Baselines (R2) RefTr and the compared baselines are image-to-graph models trained directly on centerline graphs (far cheaper to annotate). clDice, DeepVesselNet, VesselFM, and Qiu et al. are voxel-segmentation methods requiring costly voxel-level labels (as stated in Sec. 1) and address a different problem setting.

6.Metrics and Parse Performance (R2) Same splits and protocols as prior work; only metrics are upgraded. Radius-aware metrics are motivated as more robust across vessel scales. Legacy metrics were computed but omitted for space (e.g., RefTr Legacy Point-wise on ATM: Precision 0.6993, Recall 0.6413, F1 0.6549). On Parse, point-level metrics are better than Trexplorer Super; branch-level is comparable. The Parse GT issue was verified by a radiologist; public re-annotation is beyond scope. Fig. 2 shows challenging cases where prior methods struggle.

7.Architecture clarifications (R1, R3) Learned token sets are fixed trainable embeddings optimized end-to-end, analogous to DETR object queries [2]; image-agnostic until cross-attention. The decoder uses standard Transformer attention (self for inter-trajectory interactions, cross for image features, past-memory for ancestor context from up to 5 prior patches). We tested re-matching every step; reassignment disrupts refinement and destabilizes training. TNMS radius-adaptive threshold τ = max(τ_min, τ_pos * r), with τ_pos=0.3, τ_min=2.0; branches with duplicate fraction >= ρ=0.2 are merged. Loss weights (L1): alpha_pos=4.2, alpha_rad=1.2, alpha_end=0.9, alpha_div=0.3. 8.Other concerns (R1, R2) Surface similarities to MHT exist (multiple candidates, deferred decisions), but MHT prunes hypotheses while RefTr maintains them for ensembling. Title will be updated to “Tubular Tree” to cover airways. The concurrent EVAR paper is a clinical application of RefTr, disclosed per MICCAI transparency rules.

We will incorporate the above clarifications and additions in the camera-ready. Full implementation details, legacy metrics and more qualitative examples will follow with the code.



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.

    There are some discrepancies among the three reviewers, consisting of one weak reject and two weak accepts. The paper tackles an important task of 3D vascular centerline extraction, for which a Transformer-based image-to-graph framework is proposed, and evaluated on multiple public datasets with favorable results. Meanwhile, there are a number of aspects of the work that raise concerns. It would be beneficial if the authors provide further input and clarifications in the areas specifically brought out by the reviewers, including but not limited to, elucidating the main contributions, and the detailed description of the ingredients realizing the main tech contributions; improvement of the current presentation, and clarify key terms such as “confluent trajectories”, its motivation, definition and justification; critical evaluations with the SoTA baselines that go beyond own prior efforts; overlapping with the authors’ concurrent submission.

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Reject

  • Please justify your recommendation.

    Started with diverse comments and concerns and diverging ratings, the rebuttal is not effective in addressing the major concerns, including but not limited to technical contributions and empirical baselines. Overall it requires significant revision efforts that goes beyond minor nature to get into MICCAI in this round.



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 reviewers agreed that the problem is important and that the submitted work is technically coherent. Overall, this paper appears to make a useful and practically relevant contribution, supported by multi-dataset experiments, ablations, and efficiency improvements.



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 paper reconstructs tubular structures by generating many confluent trajectories and refining and merging them via recurrent self-attention.This paper reconstructs tubular structures by generating many confluent trajectories and refining and merging them via recurrent self-attention. The majority recommend acceptance. R1 and R3 both credit the trajectory-level reformulation as a genuine conceptual contribution and find the evaluation thorough—multiple public datasets, point-/branch-/graph-level metrics, efficiency analysis, and supportive ablations—maintaining their positive recommendations after rebuttal.

    R2 recommends rejection, but its two major points (limited novelty relative to the authors’ prior work, and comparison bias toward it) rest on inferring authorship from unpublished prior work, which violates the double-blind policy and cannot be weighed as a standard critique. R2’s remaining concerns are reasonable but largely concern presentation and should be addressable in the final version.

    On balance, the methodological contribution is substantive and the evaluation comprehensive enough to merit acceptance. The authors are encouraged to add a formal definition and justification of confluent trajectories and to broaden the discussion of related topology-aware and tracing methods.



back to top