Abstract

Robotic cone-beam computed tomography (CBCT) enables flexible, non-circular source trajectories, offering potential reductions in metal and cone-beam artifacts compared with standard circular orbits. Accurate reconstruction for such trajectories remains challenging: classical filtered backprojection (FBP) fails to account for view-dependent redundancy, while iterative methods are computationally expensive. To address sinusoidal non-circular trajectories, we present TCCT, a trajectory-conditioned CBCT reconstruction framework for a sinusoidal non-circular orbit family. TCCT employs a geometry-aware spectral operator to infer redundancy weights from the acquisition trajectory. Concretely, a Fourier neural operator learns the continuous mapping from per-view geometric descriptors to weights in the Radon domain, followed by a lightweight CNN that refines them spatially. During training, the predicted weights are integrated into a differentiable shift-variant FBP pipeline and end-to-end optimized using reconstruction-driven supervision on simulated volumes, without requiring ground-truth redundancy weights. Experimental results demonstrate high-quality reconstructions while achieving an order-of-magnitude speedup compared with iterative methods. TCCT lays the foundation for general non-circular trajectory CBCT reconstruction and enables future end-to-end dual-domain optimization for noise suppression, artifact reduction, and trajectory optimization. Code is available at: https://github.com/ChengzeYe/TCCT.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3821_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/ChengzeYe/TCCT

Link to the Dataset(s)

N/A

BibTex

@InProceedings{YeChe_TCCT_MICCAI2026,
        author = { Ye, Chengze AND Schneider, Linda-Sophie AND Mei, Siyuan AND Sun, Yipeng AND Liu, Daiqi AND Scheuplein, Joshua AND Wu, Fei AND Bayer, Siming AND Pérez-Toro, Paula Andrea AND Maier, Andreas},
        title = { { TCCT: Trajectory-Conditioned CBCT Reconstruction for Sinusoidal Non-circular Orbits with a Fourier Neural Operator } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16888},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    TCCT is a trajectory-conditioned reconstruction framework for non-circular CBCT. A Fourier neural operator and a lightweight CNN predict trajectory-dependent redundancy weights, which are inserted into a differentiable shift-variant FBP pipeline. The aim is to avoid trajectory-specific retraining while maintaining reconstruction quality within a family of non-circular trajectories.

  • 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 central design choice is the right one: redundancy weighting is exactly the trajectory-dependent piece of SV-FBP, and the rest of the pipeline can stay classical. Factoring the problem this way, by keeping the known operator and learning only the component that is genuinely hard to specify analytically, is what makes this paper interesting.
    • Conditioning on trajectory parameters rather than retraining per orbit is a practically useful direction for robotic CBCT.
    • Evaluation on clinically derived test anatomy is a meaningful step up from phantom-only studies.
    • The runtime advantage over ART and the competitive numbers against trajectory-specific differentiable SV-FBP give the method a clear practical argument.
  • 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 weakness is the gap between the title and what is actually shown. Non-circular orbits suggests broad trajectory generalization, but the experiments cover a single sinusoidal family, and the unseen-frequency robustness study samples within the training range. This is interpolation rather than out-of-distribution generalization, and the paper should say so plainly.
    • Training uses only 24 synthetic phantoms with 6 for validation, which is small enough to raise some concern about robustness.
    • Test projections are simulated from clinical CT volumes, so measured robotic CBCT performance is not demonstrated.
    • The Section 3.3 operating threshold is selected on the test set, which makes the sensitivity/specificity point much weaker than the AUC result.
    • The main comparisons are to trajectory-specific differentiable SV-FBP and ART(50), and prior geometry-conditioned reconstruction work is not engaged with as clearly as it could be.
  • 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

    1.Please reposition the generalization claim. What is demonstrated is interpolation within a sinusoidal orbit family and within the training frequency range, which is useful but narrower than general non-circular trajectory reconstruction. Stating this directly would actually strengthen the paper. 2.The Section 3.3 threshold selected on the test set is the issue I would most want addressed in rebuttal. Either fix the threshold on validation data, or present the clinical point as an AUC-only result. 3.Please clarify the clinical evaluation. Clinical CT volumes are a strength, but simulated projections are not the same as measured robotic CBCT data, and the current text slightly blurs this. 4.Positioning relative to prior geometry-conditioned or multi-geometry reconstruction work would sharpen the novelty claim. 5.The ART versus TCCT runtime comparison needs more detail. Were hardware and software conditions comparable, and was the ART implementation reasonably optimized? 6.Please state explicitly that the unseen-frequency experiment is interpolation within the training range. 7.From a reproducibility perspective, the simulation setup should be described more explicitly or released. In particular, it would help to provide the phantom-generation procedure, trajectory-sampling details, and the projection/simulation pipeline used for training. 8.The practical takeaway, namely comparable reconstruction quality to trajectory-specific SV-FBP without per-trajectory retraining, is the real selling point and could be stated more directly in the discussion, especially since the quantitative differences are modest. 9.As a minor presentation point, the math near Eq. (1) on page 3 appears to extend into the margin.

  • 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 think this is a genuinely interesting paper. What makes it stand out is the way it factors the problem: instead of replacing reconstruction wholesale, it keeps the SV-FBP pipeline intact and learns only the trajectory-dependent redundancy weights, which is a technically clean and well-motivated design. The trajectory-conditioned formulation is also practically useful, since it aims to avoid per-trajectory retraining while maintaining competitive reconstruction quality. What keeps my score at weak accept rather than higher is that the experiments demonstrate interpolation within a single sinusoidal trajectory family on a small synthetic training set with simulated test projections, and the clinical operating point is selected on the test set. Even with those limitations, I think the core idea is strong and interesting enough to support acceptance.

  • Reviewer confidence

    Somewhat confident (2)

  • [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.

    Sticking with accept. The core contribution is clean: TCCT keeps the differentiable SV-FBP pipeline fixed and learns only trajectory-dependent redundancy weights from trajectory descriptors, removing the per-trajectory retraining that limits prior DRACO-style methods.

    The rebuttal addressed my scope concern. The authors acknowledge that generalization is demonstrated within a family of sinusoidal orbits and that the unseen-frequency experiment is interpolation within the training range, not OOD. They also clarified that clinical evaluation uses simulated projections from clinical CT volumes (not robotic CBCT), that the efficiency win is amortized over training, and that 200 projections are used.

    Limitations remain — one trajectory family, no real robotic CBCT, small synthetic training set, and TCCT doesn’t uniformly beat trajectory-specific SV-FBP (notably SSIM at higher frequencies) — but I read these as an acceptable trade-off for a methodological contribution, provided the camera-ready narrows claims accordingly.

    My original comment about a test-set-selected threshold/AUC analysis was a mistake on my part, and I’m not weighing it.



Review #2

  • Please describe the contribution of the paper

    This paper proposes a TC-SNO framework for 3D image reconstruction in non-circular orbits, which is highly relevant for robotic CBCT applications. The authors adopt the known-operator principle by fixing the mathematically proven, differentiable SV-FBP pipeline while restricting the learning process solely to the redundancy-weight operator. By predicting these weights from trajectory parameters, the framework enables end-to-end training on simulated data using a reconstruction-driven objective, eliminating the need for ground-truth redundancy weights and trajectory-specific retraining.

  • 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.

    Practical Problem Definition: Addressing the trajectory variability in robotic CBCT is a practical problem with clinical and technical value.

    Efficient Hybrid Architecture: The approach of combining an established mathematical foundation (SV-FBP) with an AI module (operator-learning) specifically for dynamic weight calculation is structurally sound. It ensures mathematical stability while leveraging deep learning for complex, variant components.

    GT-Free Training: Successfully training the model using only a reconstruction-driven objective on simulated data, without requiring explicit ground-truth for redundancy weights, is a commendable achievement.

  • 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.

    Lack of Practical Effectiveness (Core reason for rejection): The paper highlights robustness across various sinusoidal trajectories (frequencies 1 through 9) as its main contribution. However, according to Table 1, under highly variable trajectory conditions (frequencies 7, 8, and 9), the proposed method shows a degradation in SSIM compared to existing baseline methods, even if the PSNR is slightly improved over SV-FBP. Since SSIM is a critical metric for evaluating structural image quality, this drop severely undermines the claim that explicitly analyzing and learning trajectory features provides a robust advantage over traditional SV-FBP in complex environments.

    Unsubstantiated Claims on Efficiency and Generalization: The authors claim that “Unlike differentiable SV-FBP, which requires trajectory-specific training, TCCT generalizes across trajectories without retraining, providing greater flexibility in practical applications.” However, the manuscript lacks concrete quantitative evidence to substantiate these advantages.

    • Lack of Computational Metrics: There is no comparative analysis of computational efficiency, such as training time, inference time, or resource utilization, to demonstrate exactly how much more efficient TCCT is compared to the trajectory-specific SV-FBP baseline.
    • Insufficient Proof of Domain Generalization: The paper does not adequately evaluate the completely novel trajectory domains. Without showing the actual performance drop or robustness when confronted with a different type of trajectory without retraining, the claimed superiority and practical flexibility of the proposed methodology remain unproven and difficult to understand.

    Questionable Necessity Compared to Modern 3D Reconstruction Techniques: The framework explicitly designs a module to learn trajectory features. However, in scenarios where camera intrinsic and extrinsic parameters are already known, recent methodologies like 3D Gaussian Splatting (GS) can directly utilize these geometric descriptors for highly accurate 3D reconstruction. Without a comparative analysis against these parameter-based rendering approaches, the fundamental necessity of the proposed trajectory-learning module remains questionable.

    Logical Disconnect in the Abstract: The flow of the abstract is somewhat disjointed. Right after proposing the trajectory-conditioned framework, it abruptly introduces what the Fourier Neural Operator learns. Without reading the main text, the relationship between the overarching TC-SNO framework and the specific Fourier Neural Operator is unclear, which hampers readability.

    Omission of Critical Experimental Details: Although the task may not strictly be a sparse-view CBCT reconstruction, reconstructing a CBCT volume from projection images inherently depends on the number of views. The paper fails to clearly state how many projection images were used in the experiments, which limits the reproducibility and objective assessment of the proposed method’s performance.

    Insufficient Baselines: To comprehensively prove the superiority of the proposed method, the experiments must include comparisons not only with traditional FBP-based methods but also with state-of-the-art view-synthesis or GS-based methods that utilize camera parameters (e.g., R2-GS [1], DIF-GS [2]).

    [1] R. Zha, et al, “R^2-Gaussian: Rectifying Radiative Gaussian Splatting for Tomographic Reconstruction”, NeurIPS, 2024.[2] Y. Lin, et al, “Learning 3D Gaussians for Extremely Sparse-View Cone-Beam CT Reconstruction”, MICCAI, 2025.

  • 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 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.

    (3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

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

    The authors tackle an interesting and challenging task in robotic CBCT reconstruction. However, the current experimental results do not sufficiently support the core claims of the paper. I recommend the following improvements:

    1.Analyze the performance degradation: Please provide an in-depth analysis of why the performance decreased below baseline. Discussing these limitations transparently and proposing structural improvements to mitigate this structural degradation is necessary. Furthermore, please describe the advantages of TCCT compared to previous methods in detail, particularly focusing on its efficiency.

    2.Include modern baselines: Add quantitative and qualitative comparisons with modern parameter-based rendering/reconstruction methods (e.g., R2-GS, DIF-GS). This will help clarify the unique advantages (e.g., computational efficiency, noise suppression) of your SV-FBP hybrid approach over novel synthesis methods.

    3.Clarify the abstract and settings.

  • 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.

    After reviewing the authors’ rebuttal, I update my recommendation to accept. The authors have adequately addressed my major concerns, and I accept the clarifications provided. However, I would like to note that R2-GS and DIF-GS have recently been applied to CBCT reconstruction settings, making them at least conceptually relevant to this work. While a direct experimental comparison may not be necessary given the differences in problem formulation, the camera-ready version should include a brief discussion of these methods in the related work section, clarifying the distinction in objectives and applicable scenarios.



Review #3

  • Please describe the contribution of the paper

    This paper proposes TCCT, a trajectory-conditioned CBCT reconstruction framework that addresses a key limitation of prior differentiable SV-FBP approaches such as DRACO, which require retraining a separate model for each new trajectory. The core contribution is a trajectory-conditioned spectral neural operator (TC-SNO), which learns a continuous mapping from per-view geometric descriptors to redundancy weights in the radon domain, enabling a single trained model to generalize across multiple sinusoidal trajectories without trajectory-specific retraining. TC-SNO is integrated into a DRACO pipeline and trained end-to-end using reconstruction-driven supervision, without requiring ground-truth redundancy weights.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    I find this work to be a meaningful methodological extension of prior work. The per-trajectory retraining requirement of DRACO is a genuine practical bottleneck in robotic CBCT, where acquisition trajectories may vary across procedures. Framing redundancy-weight estimation as a trajectory-conditioned operator-learning problem is therefore a natural and well-motivated extension. A particularly non-trivial aspect is that the reconstruction-driven training paradigm established in DRACO remains effective in this setting, even when the redundancy weights are dynamically predicted by TC-SNO rather than optimized as fixed trajectory-specific parameters.

    I also find the architectural choice to be well motivated. The use of a Fourier Neural Operator is well motivated by the paper’s formulation of the trajectory-to-weight mapping as a function-to-function problem. For the sinusoidal trajectory family considered here, where the acquisition geometry evolves smoothly and periodically, the spectral bias of FNO is a plausible and well-justified inductive bias.

  • 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 weakness is that the scope of the title, abstract, and several claims in the main text exceeds the actual experimental evidence. The paper is framed as addressing general non-circular trajectory CBCT reconstruction, yet all experiments are conducted within a sinusoidal trajectory family parameterized by frequency. The discussion itself acknowledges that the reported results represent interpolation within the training range rather than out-of-distribution generalization, and explicitly leaves more complex trajectories and real-world validation to future work. In this sense, the current framing appears broader than what is directly supported by the experiments. This point is especially important because the direct predecessor of this work, DRACO, evaluated a wider range of orbit types, including circular, sinusoidal, circle-plus-arc, random nearest-neighbor, and real projection data from a stochastic multi-tilt circular orbit. Therefore, I believe the paper would be substantially strengthened by narrowing its framing toward trajectory-conditioned generalization within a sinusoidal orbit family.

    A second weakness is that the ablation study is not described clearly enough to fully support the paper’s architectural conclusions. In Table 2, the “SV-FBP baseline” is described as being trained across trajectories and converging to a trajectory-agnostic compromise solution, but the manuscript does not specify its exact configuration in enough detail. As a result, it is difficult to determine precisely how trajectories are handled in this baseline, whether any explicit trajectory conditioning is used, and to what extent the observed gain should be attributed to trajectory conditioning itself versus the specific Fourier-based implementation adopted in TC-SNO. Clarifying this point would make the ablation more interpretable and strengthen the paper’s central methodological claim.

    Beyond the two main weakness discussed above, there is also a minor point that would benefit from clarification. The abstract states that the proposed framework enables future end-to-end “dual-domain optimization” for noise suppression, artifact reduction, and trajectory optimization, yet this term is neither defined nor revisited in the main text. It would therefore be helpful if the authors clarified what the two domains refer to in this context and how this claim connects to the current TCCT framework.

  • 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 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 proposed TCCT framework addresses a genuine and practically important limitation of the prior DRACO approach, and the core idea of modeling redundancy weights as a trajectory-conditioned function via a Fourier neural operator is well-motivated and technically interesting. The method demonstrates stable performance across sinusoidal trajectories without trajectory-specific retraining.

    My primary concern is that the paper’s framing overstates the scope of its contributions relative to what the experiments actually demonstrate. The title and abstract suggest broad applicability to general non-circular trajectories, yet all evaluations are confined to a sinusoidal orbit family, and the discussion itself acknowledges that the results reflect interpolation within the training range. This mismatch, combined with the observation that the Fourier-based inductive bias in TC-SNO is specifically suited to smooth, periodic trajectories and may not extend naturally to discontinuous orbits, makes the current framing difficult to accept as is. Additionally, the ablation study does not provide sufficient detail about the SV-FBP baseline configuration to fully support the paper’s central architectural claim.

    I therefore recommend a weak accept, because these concerns appear addressable through revision rather than requiring substantial additional experimentation.

  • Reviewer confidence

    Confident but not absolutely certain (3)

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

    Accept

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

    The rebuttal adequately addresses my main concern regarding the mismatch between framing and experimental scope, and the authors’ commitment to narrowing the claims to sinusoidal orbit families is appropriate.



Author Feedback

We thank all reviewers for their constructive feedback. TCCT is a trajectory-conditioned CBCT reconstruction framework for sinusoidal non-circular trajectories. Our main contributions are: (1) one model for a continuously varying sinusoidal orbit family (R1-3); (2) a known-operator design that keeps SV-FBP fixed and learns only redundancy weights, preserving FBP-like efficiency (R1-3); and (3) reconstruction-driven training without ground-truth weights (R2), evaluated on clinical CT-derived simulated CBCT data (R1). Scope beyond the evidence (R1,R3): We agree that the wording should better match the evidence. We’ll narrow the claims to sinusoidal non-circular CBCT orbits. The unseen-frequency study shows interpolation within the training frequency range, not OOD generalization to arbitrary trajectories, as stated in Sec.5.However, to our knowledge, no prior FBP-like pipeline handles variable trajectories without per-orbit retraining, a key bottleneck for applying fast iteration-free inversion formulas in robotic CT. Training phantoms and clinical simulation (R1): The 24 phantoms are not fixed cases: each uses randomly sampled sinusoidal trajectories. Since TCCT learns geometry-to-weight mapping rather than an anatomy prior, the phantoms supervise trajectory-dependent redundancy learning. Real robotic CBCT data were not used: Sec.3 uses clinical CT volumes to simulate CBCT projections under Artis zeego geometry. This evaluates phantom-to-anatomy transfer under controlled simulation. To our knowledge, no public dataset provides real robotic CBCT projections with accurate non-circular trajectory metadata matching our setting. Real validation remains future work. SSIM degradation at high-frequency trajectories (R2): We acknowledge the lower SSIM for some high-frequency orbits. For frequencies 7-9, however, the drop is modest, averaging 0.0088 SSIM, i.e., about 1.0% relative to trajectory-specific SV-FBP. Tab.1 compares one TCCT model against separately trained per-orbit SV-FBP models, a strong per-orbit reference. Our claim is competitive quality without per-orbit retraining, not per-orbit optimality. We’ll state this trade-off more explicitly. Efficiency and computational details (R1,R2): Sec.3 states training used an A40 GPU. We’ll further specify that trajectory-specific SV-FBP requires about 21h per orbit, whereas TCCT trains once for the full sinusoidal family in about 24h, thus Tab.1 needs 9 SV-FBP trainings versus one TCCT training. For inference, Sec.3.2 reports about 2s/volume for TCCT versus 30s/volume for ART(50), both on an RTX 4080S GPU. TCCT is not faster than trajectory-specific SV-FBP per inference. Its efficiency comes from avoiding per-trajectory retraining, which is the intended practical advantage for variable orbits. Tab.2 SV-FBP configuration and GS comparison (R2,R3): We’ll clarify that Tab.2 uses the non-conditioned SV-FBP from Sec.2.2: it receives no trajectory parameters G and learns shared weights across sinusoidal trajectories. We’ll remove “baseline” to avoid confusion with the per-orbit SV-FBP models in Tab.1.TCCT conditions only the SV-FBP redundancy-weight operator while keeping the reconstruction analytic, making trajectory-specific SV-FBP the closest baseline. NeRF/GS methods mainly target sparse-view reconstruction with object-specific/test-time optimization. At our 200-view scale, their objective and computational setting are not comparable. Other experimental and terminology clarifications (R1–R3): Sec.3 already specifies 200 projections. We’ll make this more visible. As stated, code will be released upon acceptance, including phantom generation, trajectory sampling and the simulation pipeline. We’ll describe them more explicitly. We’ll clarify the abstract and “dual-domain optimization” in relation to TCCT. R1’s Sec.3.3 threshold comment may refer to another manuscript: our paper reports only MSE/PSNR/SSIM/Params ablations, with no AUC/sensitivity/specificity, or test-set-selected threshold.




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

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

    The paper presents an interesting and technically sound trajectory-conditioned reconstruction framework that addresses a practical limitation of prior SV-FBP-based methods. Reviewers generally appreciate the hybrid known-operator design and the potential to avoid trajectory-specific retraining. However, several important concerns remain, particularly regarding the scope of generalization claims (currently limited to interpolation within a sinusoidal trajectory family), limited experimental validation (synthetic training and simulated projections), insufficient comparison with modern baselines, and some unclear or potentially overstated claims. Additional clarifications on experimental setup, efficiency, and evaluation methodology are also requested. Given these points, the authors are invited to provide a rebuttal addressing the reviewers’ concerns and clarifying the contributions and limitations of the work.

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

    Accept

  • Please justify your recommendation.

    The authors have addressed the reviewer’s comments. Please update those correspondingly in your camera-ready version.



Meta-review #2

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

    Accept

  • Please justify your recommendation.

    The authors have largely clarified the previously unclear points in their rebuttal, and the reviewers appear satisfied with the responses, with all agreeing on acceptance. The AC therefore recommends 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.

    The authors clarified the concerns in the rebuttal.



back to top