Pith. sign in

REVIEW 3 major objections 6 minor 30 references

DIGS: Dynamic CBCT Reconstruction using Deformation-Informed 4D Gaussian Splatting and a Low-Rank Free-Form Deformation Model

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The paper claims a deformation-informed 4D Gaussian splatting method reconstructs dynamic CBCT about six times faster than a HexPlane baseline while producing sharper images.

desk verdict A coherent DI-4DGS method for dynamic CBCT with a plausible speedup, but the simulated evaluation is partly self-fulfilling because four of six test cases generate motion from the same rank-2 B-spline FFD family the method assumes. read the letter →

arxiv 2506.22280 v1 pith:RKLSLIET submitted 2025-06-27 eess.IV cs.CV

Yuliang Huang , Imraj Singh , Thomas Joyce , Kris Thielemans , Jamie R. McClelland This is my paper · ORCID
classification eess.IVcs.CV
keywords DynamicCBCT4DGaussianSplattingFree-FormDeformationLow-rankmotionmodelMotion-compensatedreconstructionvectorfieldCone-beamCTRadiotherapy
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper takes aim at a clinical bottleneck: cone-beam CT scans for radiotherapy are corrupted by breathing motion, and the standard fix of sorting projections into respiratory phases breaks down when breaths vary. It proposes to reconstruct a full dynamic image sequence with 4D Gaussian Splatting, using an explicit low-rank free-form deformation model to move the Gaussian kernels instead of learning implicit neural motion. The central claim is that this deformation-informed design is both faster and sharper: about six times faster than the HexPlane-based 4DGS baseline, with higher PSNR and fewer motion artifacts across six simulated datasets. If the claim holds, dynamic CBCT becomes a practical, motion-resolved alternative to phase-binned reconstruction, and it produces a continuous deformation vector field that can be used for tumor tracking.

What carries the argument

The load-bearing object is the deformation-informed free-form deformation model: a continuous DVF built as a low-rank sum of spatial B-spline basis functions (a control-point lattice with cubic B-spline interpolation) and temporal B-spline weights. Its job is twofold. It acts as the motion representation, so reconstruction estimates a small number of control-point displacements instead of per-Gaussian attributes, and it enforces consistency because every Gaussian attribute at time $t$ is derived from the same $\boldsymbol{D}$ and its Jacobian $\nabla_{\boldsymbol{\mu}}\boldsymbol{D}$. The FFD also supplies spatial regularization: adjacent Gaussians are forced to move coherently, which the ablations show is the main source of the quality improvement.

What would settle it

Generate a test dataset whose ground-truth deformation is not low-rank B-spline, for example from a free-breathing trace with sliding organs or a five-component PCA motion model, and run the same comparison: if the PSNR gain over HexPlane falls to near the 0.6 dB seen in the unmatched XCAT case while the wall-clock advantage holds, the claim that the FFD model is the source of both speed and quality is weakened.

Watch

Extended reading notes

Core claim

On its own terms, the discovery is that explicit, spatially regularized motion beats implicit neural motion for dynamic CBCT. The method, called DIGS, represents the attenuation volume as a cloud of 3D Gaussian kernels at a reference time and lets one continuous free-form deformation field $\boldsymbol{D}(\boldsymbol{x},t)=\boldsymbol{x}+\sum_{r=1}^{N_r}\omega_r(t)\boldsymbol{u}_r(\boldsymbol{x})$ warp them: means move along the field and covariances transform by its Jacobian, so position, scale, and rotation stay mutually consistent. The spatial basis functions $\boldsymbol{u}_r$ are interpolated from a learnable B-spline control-point lattice and the temporal weights $\omega_r(t)$ from 1D B-splines, with $N_r=2$ in the experiments. This yields a $6\times$ speedup over HexPlane (about 15 minutes versus 1 hour 40 minutes on the same GPU), higher mean PSNR (28.14 dB vs 25.86 dB), and a directly usable deformation vector field; ablations attribute the quality gain mostly to the FFD's spatial regularization.

Load-bearing premise

The evaluations that show the largest improvement are generated with a motion model of exactly the same kind the method uses (two PCA components and B-spline free-form deformation), so the claimed advantage may rely on real motion being low-rank and smooth in the same way.

Editorial extensions

If this is right

  • Dynamic CBCT can be performed without phase sorting, so irregular breathing no longer has to be averaged into sorted phases.
  • A continuous deformation vector field is produced alongside the images, giving radiotherapy a voxel-wise motion estimate for tumor tracking and dose accumulation.
  • Reconstruction time drops to about 15 minutes on a single consumer GPU, roughly six times faster than the HexPlane 4DGS baseline, making the approach closer to clinical workflow.
  • Removing the FFD spatial regularization degrades mean PSNR by about 0.85 dB, showing that coherent Gaussian motion is what carries the quality gain.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The speed and quality advantages may shrink on real patient motion that is not well described by two PCA components; the XCAT cases in the paper already show a smaller gain (0.6 dB) when the ground truth does not come from the matched low-rank model.
  • The same deformation-informed coupling could be carried over to other tomographic reconstruction problems, such as PET or dynamic CT, where a per-projection motion field is needed.
  • Because the DI framework yields both images and DVFs, the estimated motion field could be checked against independent surrogates (e.g., breathing belts or MR-derived motion) to validate it on real scans.
  • A testable extension would be to increase $N_r$ adaptively during optimization, which could handle motion of higher spatial or temporal complexity without giving up the speed advantage.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes DIGS, a 4D Gaussian-splatting method for dynamic CBCT reconstruction. Motion is represented explicitly by a low-rank free-form deformation model D(x,t)=x+∑_{r=1}^{Nr} ω_r(t)u_r(x), with cubic-B-spline spatial basis functions u_r and 1D B-spline temporal weights ω_r. Gaussian means are moved by D and covariances are updated by the Jacobian of D (Eqs. 7–8), so all Gaussian attributes evolve consistently under one DVF. The method jointly optimizes reference Gaussians and motion basis against measured projections. It is evaluated on six simulated CBCT datasets (two XCAT-based, four based on 4DCT patient data with PCA-compressed deformations) and compared with SuPReMo and a HexPlane-based 4DGS baseline, reporting better PSNR/RMSE and about a 6x speedup over HexPlane. Two ablation studies isolate the effect of the deformation-informed update and of the FFD spatial regularization. Code is provided.

Significance. If the reported gains are robust, DIGS would be a useful step toward practical dynamic CBCT: it is faster than a HexPlane-style 4DGS baseline, produces a continuous DVF that can be used for tumor tracking, and ships with code. The covariance update in Eq. (8) is standard fluid/continuum mechanics and is cleanly derived, and the ablation isolating the FFD spatial regularization is informative. However, the central performance claim is weakened by the evaluation design: four of the six test cases use ground-truth motion generated from exactly the same low-rank PCA B-spline FFD family that the method assumes (Nr=2, B-spline spatial basis). The non-matched XCAT cases show much smaller gains over HexPlane (e.g., 0.63 dB in case 2), so the generality of the claimed improvement is not established. The speedup is plausible but the comparison protocol with HexPlane is incompletely specified.

major comments (3)
  1. [§4.1 and §4.2] Cases 3–6 are generated by registering 4DCT phases with NiftyReg (a cubic B-spline FFD) and then applying PCA that retains only the mean deformation plus two principal components (§4.1), while the proposed model sets Nr=2 and uses B-spline FFD spatial basis functions (§4.2, Eq. 9–10). The generating motion therefore lies in the same hypothesis class as the method itself. The reported PSNR gains over HexPlane on cases 3–6 (roughly 2.1–2.7 dB) partly reflect a matched inductive bias rather than general superiority. The XCAT cases 1–2, which are not generated from this model family, show much smaller gains (case 2: 21.78 vs 21.15 dB). To support the central claim, please add evaluations with mismatched motion, e.g., retaining more than two PCA components (Nr=3 or 4), using non-B-spline deformations, or using independently measured respiratory motion, and report per-case results. Also report sensitivity to Nr.
  2. [§5, Table 1] The claimed 6× speedup over HexPlane is based on wall-clock times of about 1h27m–1h50m for HexPlane versus 14–16m for DIGS, but the paper does not state whether the two methods were given the same number of training iterations, the same learning-rate schedule, or the same early-stopping criterion. Without this information or a runtime–quality curve, the speedup may partly reflect optimization settings rather than representation efficiency. Please specify the exact comparison protocol (iterations, convergence criteria, hardware, measured times) or provide performance-versus-time plots.
  3. [§5, Discussion] The conclusion that the method "produces sharper images with fewer motion artifacts" is supported primarily by mean PSNR/RMSE on the matched simulation cases and by a few example slices (Fig. 2). Since the ground truth is created by warping reference images with the simulated DVFs and the method explicitly models a DVF of the same family, the evaluation metric is aligned with the method's inductive bias. To strengthen the claim, please include quantitative artifact metrics or a reader study on the non-matched cases, and ideally on an independently acquired dataset or on clinical CBCT projections.
minor comments (6)
  1. [§3.1, Eq. (4)] The tilde notation is used for both the projected 2D Gaussian and the projected coordinates/covariance, which is confusing; consider using a distinct symbol, e.g., a subscript '2D'.
  2. [§4.2] Nr is set to 2 without any ablation or justification; since Nr is a key hyperparameter of the low-rank motion model, a sensitivity study (Nr=1,3,4) would help the reader understand its role.
  3. [§4.1] The description of the 4DCT-based simulation states that 310 time points are uniformly sampled from 20 breathing cycles, but it is not clear whether the projections are presented to the method in chronological order or shuffled; please clarify, since the temporal B-spline basis assumes a temporal ordering.
  4. [§5, Table 1] Table 1 reports mean PSNR/RMSE over cases but not standard deviations or per-case values for the ablations; reporting these would help assess variability across datasets.
  5. [§2.2] The claim that this is the first application of a physics-based deformation-informed framework to image reconstruction is strong given that PhysGaussian and related methods apply the same covariance update to Gaussian kernels; please clarify the distinction (image reconstruction versus graphics rendering) and soften the novelty claim if appropriate.
  6. [References] Reference [9] is cited as an arXiv preprint; if a peer-reviewed version has appeared, please cite that version instead.

Circularity Check

0 steps flagged · score 1.0 of 10

No circularity in the derivation; the matched low-rank FFD simulation for cases 3-6 is a generalizability concern, not a circular step.

full rationale

The paper's derivation chain is self-contained: Eqs. (7)-(8) follow from applying the deformation D (Eq. 9) and its Jacobian to Gaussian means and covariances, and the B-spline FFD parameterization (Eq. 10) is an explicit model choice rather than a result derived from the data. The image-quality claims are empirical, benchmarked against SuPReMo and HexPlane on simulated projections, and the reconstruction never uses the simulated DVFs as inputs—only the projection data are optimized against. No parameter is fitted to the test ground truth and then repackaged as a prediction. The self-citations ([6], [7], [10], [11]) are used for data post-processing, baselines, or background, and none carries the logical load of the derivation. The main reviewer concern is that cases 3-6 generate ground-truth motion by PCA-truncated NiftyReg B-spline FFDs, which is the same low-rank FFD family the method assumes (Nr=2). This is a real external-validity and benchmark-matching limitation, and the independent XCAT cases show a smaller advantage (case 2: 21.15 to 21.78 dB PSNR), but it is not circularity: the evaluation still requires the optimizer to recover unknown Gaussian parameters and motion from projections, and the ground-truth DVFs are not used during optimization. Under the stated definition of circularity—a prediction equivalent to its inputs by construction—this paper does not exhibit such a step.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The method relies on standard Gaussian splatting and continuum mechanics approximations, plus a low-rank B-spline motion model. No new physical entities are introduced. The key assumptions are the representational power of Gaussians, local linearity of deformation, and a rank-2 motion model, the latter being aligned with the simulation setup.

free parameters (4)
  • Nr (number of motion basis functions) = 2
    Hand-set without sensitivity analysis; matches the 2 PCA components used to generate ground-truth motion in cases 3-6.
  • B-spline control point spacing = 8 voxels spatial, 4 time points
    Chosen without ablation; affects the smoothness and flexibility of the deformation field.
  • Initial learning rates for space and time grids = 1e-4 and 1e-2
    Standard Adam hyperparameters, not tuned per dataset or reported with sensitivity analysis.
  • Number of Gaussians = 80K
    Set in line with prior work [9]; no analysis of how this affects reconstruction quality or speed.
assumptions (5)
  • domain assumption A 3D Gaussian projection can be approximated by a 2D Gaussian (Eq. 4)
    Borrowed from R2-Gaussian [28]; fundamental to the rendering model and not re-derived for this application.
  • domain assumption The attenuation field can be represented as a sum of Gaussians (Eq. 3)
    Core representation assumption of Gaussian splatting; not proven for CBCT attenuation fields.
  • domain assumption The deformation is locally linear over each Gaussian's support (Eq. 8)
    Standard continuum mechanics approximation; assumes small strain within each Gaussian, not validated for respiratory motion magnitudes.
  • ad hoc to paper Motion is low-rank with Nr=2 basis functions (Eq. 9)
    The rank is set to match the simulated motion's 2 PCA components for cases 3-6; not justified for general patient motion.
  • domain assumption Gaussian density remains constant over time
    Assumes no attenuation change during deformation, reasonable for incompressible tissue but unvalidated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DIGS: Dynamic CBCT Reconstruction using Deformation-Informed 4D Gaussian Splatting and a Low-Rank Free-Form Deformation Model." pith.science (2026). https://pith.science/paper/RKLSLIET

@misc{pith2026250622280,
  author       = {Pith},
  title        = {Pith review of: DIGS: Dynamic CBCT Reconstruction using Deformation-Informed 4D Gaussian Splatting and a Low-Rank Free-Form Deformation Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RKLSLIET}},
  note         = {Machine review of arXiv:2506.22280}
}
read the original abstract

3D Cone-Beam CT (CBCT) is widely used in radiotherapy but suffers from motion artifacts due to breathing. A common clinical approach mitigates this by sorting projections into respiratory phases and reconstructing images per phase, but this does not account for breathing variability. Dynamic CBCT instead reconstructs images at each projection, capturing continuous motion without phase sorting. Recent advancements in 4D Gaussian Splatting (4DGS) offer powerful tools for modeling dynamic scenes, yet their application to dynamic CBCT remains underexplored. Existing 4DGS methods, such as HexPlane, use implicit motion representations, which are computationally expensive. While explicit low-rank motion models have been proposed, they lack spatial regularization, leading to inconsistencies in Gaussian motion. To address these limitations, we introduce a free-form deformation (FFD)-based spatial basis function and a deformation-informed framework that enforces consistency by coupling the temporal evolution of Gaussian's mean position, scale, and rotation under a unified deformation field. We evaluate our approach on six CBCT datasets, demonstrating superior image quality with a 6x speedup over HexPlane. These results highlight the potential of deformation-informed 4DGS for efficient, motion-compensated CBCT reconstruction. The code is available at https://github.com/Yuliang-Huang/DIGS.

Figures

Figures reproduced from arXiv: 2506.22280 by the authors.

Figure 1
Figure 1. Schematic illustration of the proposed method. A continuous free-form defor￾mation (FFD) is constructed from a set of learnable spatial and temporal basis, which is used to warp the reference Gaussians though a deformation-informed (DI) framework. Projections rendered from the dynamic Gaussians are compared with the measured projections and the optimiser updates the reference Gaussians and the motion basis to minimi… view at source ↗
Figure 2
Figure 2. Ground-truth and reconstructed mean position images for case 2 (top), case 3 (middle) and case 5 (bottom). Dynamic range: 0.00 - 0.03 mm−1 ment, but remains comparable to results of using a unified DVF. Ablation Study 2 results in a noticeable performance drop [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 19 canonical work pages

  1. [1]

    International Journal of Radiation Oncology* Biology* Physics86(2), 372–379 (2013)

    Balik, S., Weiss, E., Jan, N., Roman, N., Sleeman, W.C., Fatyga, M., Christensen, G.E.,Zhang,C.,Murphy,M.J.,Lu,J.,etal.:Evaluationof4-dimensionalcomputed tomography to 4-dimensional cone-beam computed tomography deformable image registration for lung cancer adaptive radiation therapy. International Journal of Radiation Oncology* Biology* Physics86(2), 372...

    work page 2013

  2. [2]

    Cambridge university press (1997)

    Bonet, J., Wood, R.D.: Nonlinear continuum mechanics for finite element analysis. Cambridge university press (1997)

    work page 1997

  3. [3]

    Advances in Neural Information Processing Systems (2024)

    Cai, J., Yang, Y., Yuan, W., He, Y., Dong, Z., Bo, L., Cheng, H., Chen, Q.: Gaussian-informed continuum for physical property identification and simulation. Advances in Neural Information Processing Systems (2024)

    work page 2024

  4. [4]

    In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion

    Cao, A., Johnson, J.: Hexplane: A fast representation for dynamic scenes. In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion. pp. 130–141 (2023)

    2023

  5. [5]

    International journal of radiation oncology, biology, physics87(1), 33–45 (2013)

    De Los Santos, J., Popple, R., Agazaryan, N., Bayouth, J.E., Bissonnette, J.P., Bucci, M.K., Dieterich, S., Dong, L., Forster, K.M., Indelicato, D., et al.: Image guided radiation therapy (igrt) technologies for radiation therapy localization and delivery. International journal of radiation oncology, biology, physics87(1), 33–45 (2013)

    work page 2013

  6. [6]

    Physics in Medicine & Biology 65(16), 165005 (2020)

    Eiben, B., Bertholet, J., Menten, M.J., Nill, S., Oelfke, U., McClelland, J.R.: Con- sistent and invertible deformation vector fields for a breathing anthropomorphic phantom: a post-processing framework for the xcat phantom. Physics in Medicine & Biology 65(16), 165005 (2020)

    work page 2020

  7. [7]

    In: Biomedical Image Registra- tion: 8th International Workshop, WBIR 2018, Leiden, The Netherlands, June 28-29, 2018, Proceedings 8

    Eiben, B., Tran, E.H., Menten, M.J., Oelfke, U., Hawkes, D.J., McClelland, J.R.: Statistical motion mask and sliding registration. In: Biomedical Image Registra- tion: 8th International Workshop, WBIR 2018, Leiden, The Netherlands, June 28-29, 2018, Proceedings 8. pp. 13–23. Springer (2018)

    work page 2018

  8. [8]

    arXiv preprint arXiv:2401.15318 (2024) 10 Y

    Feng, Y., Feng, X., Shang, Y., Jiang, Y., Yu, C., Zong, Z., Shao, T., Wu, H., Zhou, K., Jiang, C., et al.: Gaussian splashing: Unified particles for versatile motion synthesis and rendering. arXiv preprint arXiv:2401.15318 (2024) 10 Y. Huang et al

    arXiv 2024

Show all 30 references
  1. [9]

    arXiv preprint arXiv:2501.04140 (2025)

    Fu, Y., Zhang, H., Cai, W., Xie, H., Kuo, L., Cervino, L., Moran, J., Li, X., Li, T.: Spatiotemporal gaussian optimization for 4d cone beam ct reconstruction from sparse projections. arXiv preprint arXiv:2501.04140 (2025)

    2025 arXiv

  2. [10]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Huang,Y.,Eiben,B.,Thielemans,K.,McClelland,J.R.:Resolvingvariablerespira- tory motion from unsorted 4d computed tomography. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 588–597. Springer (2024)

    2024

  3. [11]

    Physics in Medicine & Biology69(2), 025020 (2024)

    Huang, Y., Thielemans, K., Price, G., McClelland, J.R.: Surrogate-driven respi- ratory motion model for projection-resolved motion estimation and motion com- pensated cone-beam ct reconstruction from unsorted projection data. Physics in Medicine & Biology69(2), 025020 (2024)

    2024

  4. [12]

    Physics in Medicine & Biology66(21), 215018 (2021)

    Jailin, C., Roux, S., Sarrut, D., Rit, S.: Projection-based dynamic tomography. Physics in Medicine & Biology66(21), 215018 (2021)

    2021

  5. [13]

    ACM Trans

    Kerbl, B., Kopanas, G., Leimkühler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph.42(4), 139–1 (2023)

    2023

  6. [14]

    In: European Conference on Computer Vision

    Kratimenos, A., Lei, J., Daniilidis, K.: Dynmf: Neural motion factorization for real- time dynamic view synthesis with 3d gaussian splatting. In: European Conference on Computer Vision. pp. 252–269. Springer (2024)

    2024

  7. [15]

    Medical physics35(10), 4649–4659 (2008)

    Leng, S., Zambelli, J., Tolakanahalli, R., Nett, B., Munro, P., Star-Lack, J., Pali- wal, B., Chen, G.H.: Streaking artifacts reduction in four-dimensional cone-beam computed tomography. Medical physics35(10), 4649–4659 (2008)

    2008

  8. [16]

    arXiv preprint arXiv:2411.16800 (2024)

    Lin, J., Wang, Z., Jiang, S., Hou, Y., Jiang, M.: Phys4dgen: A physics-driven framework for controllable and efficient 4d content generation from a single image. arXiv preprint arXiv:2411.16800 (2024)

    2024

  9. [17]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Lin, Y., Dai, Z., Zhu, S., Yao, Y.: Gaussian-flow: 4d reconstruction with dynamic 3d gaussian particle. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21136–21145 (2024)

    2024

  10. [18]

    Computer methods and programs in biomedicine98(3), 278–284 (2010)

    Modat, M., Ridgway, G.R., Taylor, Z.A., Lehmann, M., Barnes, J., Hawkes, D.J., Fox, N.C., Ourselin, S.: Fast free-form deformation using graphics processing units. Computer methods and programs in biomedicine98(3), 278–284 (2010)

    2010

  11. [19]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin- Brualla, R.: Nerfies: Deformable neural radiance fields. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5865–5874 (2021)

    2021

  12. [20]

    Rit, S., Oliva, M.V., Brousmiche, S., Labarbe, R., Sarrut, D., Sharp, G.C.: The reconstruction toolkit (rtk), an open-source cone-beam ct reconstruction toolkit basedontheinsighttoolkit(itk).In:JournalofPhysics:ConferenceSeries.vol.489, p. 012079. IOP Publishing (2014)

    2014

  13. [21]

    Medical physics 36(6Part1), 2283–2296 (2009)

    Rit, S., Wolthaus, J.W., van Herk, M., Sonke, J.J.: On-the-fly motion-compensated cone-beam ct using an a priori model of the respiratory motion. Medical physics 36(6Part1), 2283–2296 (2009)

    2009

  14. [22]

    IEEE transactions on medical imaging18(8), 712–721 (1999)

    Rueckert, D., Sonoda, L.I., Hayes, C., Hill, D.L., Leach, M.O., Hawkes, D.J.: Non- rigid registration using free-form deformations: application to breast mr images. IEEE transactions on medical imaging18(8), 712–721 (1999)

    1999

  15. [23]

    Medical physics37(9), 4902–4915 (2010)

    Segars, W.P., Sturgeon, G., Mendonca, S., Grimes, J., Tsui, B.M.: 4d xcat phantom for multimodality imaging research. Medical physics37(9), 4902–4915 (2010)

    2010

  16. [24]

    Medical physics32(4), 1176–1186 (2005)

    Sonke, J.J., Zijp, L., Remeijer, P., Van Herk, M.: Respiratory correlated cone beam ct. Medical physics32(4), 1176–1186 (2005)

    2005

  17. [25]

    Radiation Oncology7(1), 1–8 (2012) DIGS for dynamic CBCT 11

    Sweeney, R.A., Seubert, B., Stark, S., Homann, V., Müller, G., Flentje, M., Guck- enberger, M.: Accuracy and inter-observer variability of 3d versus 4d cone-beam ct based image-guidance in sbrt for lung tumors. Radiation Oncology7(1), 1–8 (2012) DIGS for dynamic CBCT 11

    2012

  18. [26]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Xie, T., Zong, Z., Qiu, Y., Li, X., Feng, Y., Yang, Y., Jiang, C.: Physgaussian: Physics-integrated 3d gaussians for generative dynamics. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4389– 4398 (2024)

    2024

  19. [27]

    In: Interna- tional Conference on Medical Image Computing and Computer-Assisted Interven- tion

    Yang, S., Li, Q., Shen, D., Gong, B., Dou, Q., Jin, Y.: Deform3dgs: Flexible defor- mation for fast surgical scene reconstruction with gaussian splatting. In: Interna- tional Conference on Medical Image Computing and Computer-Assisted Interven- tion. pp. 132–142. Springer (2024)

    2024

  20. [28]

    In: Advances in Neural Information Processing Systems (NeurIPS) (2024)

    Zha, R., Lin, T.J., Cai, Y., Cao, J., Zhang, Y., Li, H.: R2-gaussian: Rectifying radiative gaussian splatting for tomographic reconstruction. In: Advances in Neural Information Processing Systems (NeurIPS) (2024)

    2024

  21. [29]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Zha, R., Zhang, Y., Li, H.: Naf: neural attenuation fields for sparse-view cbct reconstruction. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 442–452. Springer (2022)

    2022

  22. [30]

    IEEE Transactions on Visualization and Computer Graphics8(3), 223–238 (2002)

    Zwicker, M., Pfister, H., Van Baar, J.,Gross, M.: Ewasplatting. IEEE Transactions on Visualization and Computer Graphics8(3), 223–238 (2002)

    2002

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.