List of Papers Browse by Subject Areas Author List
Abstract
Cone-Beam CT (CBCT) systems utilizing continuous gantry rotation and flat-panel detector readout introduce complex acquisition dynamics, such as angular motion blur and rolling shutter distortion. Conventional reconstruction algorithms, however, typically rely on a static geometry assumption, inevitably leading to the blurred structures and geometric inconsistencies. In this paper, we propose PhyRadGS, a physics-aware 3DGS framework that integrates a novel spatial temporal discretization strategy. Specifically, We explicitly model rolling shutter by spatially partitioning projections and motion blur by temporally segmenting the exposure duration. Furthermore, we implement a parallelized multi-pose batch rendering pipeline to mitigate computational overhead, maintaining the real-time efficiency inherent to 3DGS. Experiments on synthetic and real-world datasets demonstrate PhyRadGS significantly outperforms state-of-the-art static baselines, improving PSNR by +2.3 dB and SSIM by +0.043. Qualitative comparisons further confirm that PhyRadGS effectively recovers high-frequency clinical details even under rapid scanning protocols. The source code and dataset is available at https://github.com/liaw05/PhyRadGS.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0017_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/liaw05/PhyRadGS
Link to the Dataset(s)
Static synthetic dataset: https://github.com/Ruyi-Zha/r2_gaussian/tree/main
Synthetic Flying Dataset & Real Phantom Data: https://github.com/liaw05/PhyRadGS
BibTex
@InProceedings{LiAnw_PhyRadGS_MICCAI2026,
author = { Li, Anwei AND Zheng, Zhongyang AND Sun, Tianyuan AND Zhang, Yubing AND Wang, Rongqiu},
title = { { PhyRadGS: Physics-Aware Radiative Gaussian Splatting for Fly-Scanning CBCT Reconstruction } },
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
The study presents a physics-driven 3D Gaussian splatting framework to model motion blur and rolling shutter effects in fly-scanning CBCT reconstruction. The approach is intuitive and well-motivated, and demonstrates consistent improvements over existing methods.
- 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 physics-driven modeling and reconstruction offer an elegant solution to a challenging reconstruction problem.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The study assumes known scanner kinematics, but there can exist calibration errors that render the prior knowledge less reliable in the optimization pipeline. The method’s robustness to such errors have not been evaluated. 2.In many modern CBCT systems with high frame rates and short exposure times, the fly-scanning-related motion effects can be relatively minor, limiting the development’s clinical utility. 3.The comparison study is somewhat limited. All compared methods are static-based, while the known motion information can be similarly incorporated in other methods for a more fair comparison. 4.Clinical evaluation on real data is lacking.
- 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?
The study presents a nice physics-driven framework for fly-scanning CBCT reconstruction and shows improvements over conventional methods. It could be further improved with additional robustness tests under motion parameter uncertainties, and through comparison with stronger baselines on real non-phantom scans.
- 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 authors have responded well to my comments.
Review #2
- Please describe the contribution of the paper
1.they propose a framework integrates the known mechanical kinematics of the scanner and the readout characteristics of the detector, allowing it to recover sharp anatomical micro-structures from severely blurred detector projections. 2.the authors propose a mathematical discretization strategy that involves spatial partitioning to model the rolling shutter effect and temporal segmentation to model the exposure duration (motion blur)
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The paper explicitly models the complex acquisition dynamics of continuous gantry rotation (fly-scanning) and flat-panel detector readout mechanisms. 2.The evaluation is exceptionally thorough, utilizing both synthetic datasets with simulated fly-scanning dynamics and real-world double-layer sphere phantom data acquired from a commercial CBCT system.
- 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.Inadequate baseline comparisons (lack of dynamic/motion-correction baselines) and missing literature review on concurrent 3DGS works. The paper compares PhyRadGS exclusively against reconstruction methods that rely on a strict static pose assumption, namely FDK, TV-regularized ASD-POCS, NAF, SAX-NeRF, and R2-Gaussian. While it successfully proves that PhyRadGS beats static methods on dynamic data, this is somewhat of an unfair fight. Because fly-scanning inherently introduces motion, static baselines are mathematically expected to fail. 2.Lack of in-vivo clinical data validation The real-world evaluation in the paper is entirely restricted to a double-layer sphere phantom scanned on a commercial CBCT system. A sphere phantom represents highly simplified, high-contrast geometry. It completely lacks the intricate, overlapping, and low-contrast high-frequency anatomical structures found in actual human biology (e.g., bone trabeculae, soft tissue boundaries). Without real-world in-vivo patient scans, the robustness of the algorithm in a true clinical environment remains unproven. The authors missed concurrent works such as: 3DGR-CT, X^2 Gaussian. 3.As the authors explicitly acknowledge in their conclusion, the current physical model strictly assumes rigid motion. The framework successfully models the known mechanical kinematics of the scanner (continuous gantry rotation), but it ignores the patient entirely.
- 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?
1.Inadequate baselines. 2.The real-world evaluation relies solely on a double-layer sphere phantom. 3.The model assumes perfectly rigid mechanical motion based on scanner kinematics. It entirely ignores patient-induced non-rigid deformations (e.g., breathing, tremors, cardiac motion), which are inevitable in clinical fly-scanning and represent a major hurdle for clinical adoption.
- 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.
All my concerns have been resolved.
Review #3
- Please describe the contribution of the paper
The paper proposes PhyRadGS, a physics-aware radiative Gaussian splatting framework for fly-scanning CBCT reconstruction that models motion blur from continuous gantry rotation and rolling-shutter distortion from row-wise detector readout. The core technical idea is a spatio-temporal discretization that partitions each projection into spatial strips and temporal sub-intervals, converts continuous acquisition dynamics into calibrated micro-states, and enforces scanner kinematics in the forward model. Experiments on synthetic fly-scanning data and real phantom data show improved reconstruction quality versus static baselines while maintaining practical runtime.
- 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.This work aims to address a practical issue and targets a realistic CBCT setting (continuous fly-scanning).
2.The forward model is clear and the pipeline links theory with the implementation well.
3.Baselines cover analytic/iterative methods, implicit neural representations, and 3DGS-based methods, making a strong comparative scope.
4.The ablation studies are useful and gain insights for the hyperparameter choices.
- 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.Efficiency gain is one of the main superiority of PhyRadGS. However, it remains unclear if all baselines are implemented under the same condition. For example, R2-Gaussian runs 30K iterations by default while PhyRadGS is 20K. Does it means the advantage comes from reducing number of iterations rather than other designs?
2.Real-data evaluation is based on a sphere phantom setup rather than in-vivo patient studies. This is a reasonable first step but limits evidence for anatomical complexity, metal artifacts, and clinical workflow variability.
3.The ablation scope can be expanded. While L1 fidelity + SSIM + TV is a standard loss setup, an extra L1 regularization term is new. The authors claim it is for eliminating floating artifacts, but more ablation studies should be conducted to verify the claim.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I recommend acceptance because the paper addresses a well-motivated, practically important failure mode of CBCT under fly-scanning. PhyRadGS makes a clear methodological contribution by embedding a physics-aware acquisition model. The evaluation is convincing for the claimed scope, combining controlled synthetic fly-scanning simulations and real phantom data, and comparing against diverse strong baselines (analytic, iterative, implicit fields, and splatting). Remaining issues, such as limited in-vivo evidence, and the need for broader robustness analyses, are reasonable extensions rather than fundamental flaws.
- 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 response addressed my concerns.
Author Feedback
We thank all reviewers for their constructive feedback.
- Lack of In-Vivo Clinical Data (Reviewer1 Comment 4, Reviewer2 Comment 2, Reviewer3 Comment 2)
Real dynamic fly-scanning patient data is extremely difficult to obtain due to ethical constraints. To bridge this gap, we are the first to synthesize fly-scanning data using TIGRE on publicly available anatomical volumes (LIDC-IDRI and Open SciVis) that are widely adopted by NAF (Ref. 27), R2-Gaussian (Ref. 26), SAX-NeRF (Ref. 4), and X-Gaussian (Ref. 3), providing a reproducible benchmark for future research. Our commercial CBCT experiment (2048×2048) is among the very few real scans in this field. The controlled phantom cleanly isolates fly-scanning dynamics: direct validation for our forward model. In-vivo data is planned for future work.
- Static-Only Baselines Fairness (Reviewer1 Comment 3, Reviewer2 Comment 1)
Static methods represent current clinical practice: their failure under fly-scanning directly motivates our work. To directly address the fairness concern, we integrated our Spatio-Temporal (ST) module into R2-Gaussian (Ref. 26) (“R2-Gaussian w/ ST,” Table 1): this achieves +2.29 dB PSNR and +0.038 SSIM over static R2-Gaussian, confirming the gain comes from physics modeling. PhyRadGS achieves better quality than “R2-Gaussian w/ ST” while being ~4x faster (~5 min vs. ~19 min), widening further on high-resolution real data (~9 min vs. ~51 min, Table 1).
- Concurrent Works (Reviewer2 Comment 2)
Regarding concurrent works: 3DGR-CT improves rasterization accuracy for cone-beam geometry but remains fully static. X^2-Gaussian targets 4D CT for self-supervised respiratory motion: motion caused by patient not the system. Both these works try to solve fundamentally different problem, which is why they are not included in comparison.
- Iteration Count Fairness (Reviewer3 Comment 1)
As stated in Sec. 3.1, all 3DGS-based methods (R2-Gaussian, PhyRadGS) were trained under identical settings: 20K iterations on a single RTX 4090, with wall-clock time reported. In our reproduction, R2-Gaussian converges by 20K with performance comparable to its original 30K setting (Ref. 26). Speedup between PhyRadGS (~5 min) vs. “R2-Gaussian w/ ST” (~19 min, Table 1) reflects batch rendering efficiency, not fewer iterations.
- Calibration Error (Reviewer1 Comment 1)
Commercial CBCT systems achieve calibration precision below 0.1° angular uncertainty: two orders of magnitude smaller than the 14.4° integration arc per exposure (fly-scanning protocol). Our real phantom experiment on a commercial system empirically confirms robustness: PhyRadGS achieves 37.43 dB PSNR vs. 35.13 dB for PhyRadGS w/o ST (Table 1).
- Clinical Utility at High Frame Rates (Reviewer1 Comment 2)
Rolling shutter distortion is hardware-intrinsic and frame-rate-independent. Motion blur is caused by integration arc per view, which is determined by angular spacing (360°/N_views) but not frame rate or exposure time: for 25 (50) views, each projection spans 14.4° (7.2°). While increasing view count mitigates motion blur, it raises radiation dose: a key clinical trade-off. PhyRadGS achieves 31.23 dB at ST 25 views (static 50 views 31.96 dB), making smaller dose possible.
- L1 Regularization Ablation (Reviewer3 Comment 3)
As stated in Sec. 2.4, L1 density regularization is a standard technique for suppressing floating artifacts in sparse-view reconstruction, analogous to the TV regularization adopted by R2-Gaussian (Ref. 26). Our ablation (Table 2) focuses on our main contribution: novel parameters M and N. Extended loss ablations are acknowledged for future work.
- Rigid Motion Assumption (Reviewer2 Comment 3)
This limitation is explicitly acknowledged in our Conclusion. PhyRadGS models scanner mechanical kinematics, which are inherently rigid. Patient non-rigid motion is an orthogonal problem addressable by combining our forward model with deformation fields (Refs. 23, 24). This extension is planned as future work.
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 a physics-driven framework for fly-scanning CBCT reconstruction with meaningful improvements and clear methodological contributions. However, several reviewers’ concerns remain, including reliance on phantom-based evaluation, insufficient comparison with strong baselines, and limited robustness to non-rigid motion. The authors are encouraged to address and clarify these points in the rebuttal. The AC recommends proceeding to rebuttal before making a final decision.
- 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 clarified the previously unclear points in their rebuttal, and all reviewers appear satisfied with the responses. The AC recommends proceeding with acceptance.
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 recognize the well-motivated problem, the clear physics-aware forward model for fly-scanning CBCT reconstruction, and the solid evaluation combining synthetic and real phantom data. Concerns about baseline fairness, lack of in-vivo validation, and the rigid motion assumption were raised but adequately addressed in the rebuttal, where the authors clarified the integration of their module into R2-Gaussian as a fairer comparison, the practical difficulty of obtaining clinical fly-scanning data, and the inherent scope of the rigid motion assumption.
All three reviewers updated their recommendation to accept after the rebuttal. The paper is therefore recommended 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.
The authors have addressed most of the reviewers’ comments. Limitations are well justified and explained. I recommend acceptance.