pith. sign in

arxiv: 2312.15676 · v2 · submitted 2023-12-25 · 📡 eess.IV · cs.CV

3DGR-CT: Sparse-View CT Reconstruction with a 3D Gaussian Representation

Yingtai Li , Xueming Fu , Han Li , Shang Zhao , Ruiyang Jin , S. Kevin Zhou This is my paper

Pith reviewed 2026-05-24 05:51 UTC · model grok-4.3

classification 📡 eess.IV cs.CV
keywords sparse-view CT3D Gaussian representationdifferentiable CT projectorFBP initializationlow-dose imagingvolume reconstructionGaussian splatting
0
0 comments X

The pith

A 3D Gaussian representation reconstructs sparse-view CT volumes more accurately and with faster convergence than implicit neural methods.

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

Sparse-view CT lowers radiation dose by using fewer X-ray projections but creates noise and artifacts that are hard to remove. The paper replaces implicit neural representations with an explicit 3D Gaussian representation, initialized from filtered back-projection images and optimized through a differentiable CT projector. This change yields higher reconstruction accuracy and quicker optimization on multiple datasets. The same representation also supports real-time physical simulation, a task that remains difficult for neural alternatives. The core result is that the Gaussian approach can serve as a practical, efficient substitute for neural volume representations in low-dose CT imaging.

Core claim

We introduce 3DGR-CT, which represents a CT volume as a set of 3D Gaussians. These Gaussians are initialized using FBP images and then refined by direct differentiation through a CT projector, delivering higher reconstruction accuracy, faster convergence, and the ability to run real-time physical simulations that implicit neural representations struggle to support.

What carries the argument

3D Gaussian representation equipped with FBP-image-guided initialization and integration into a differentiable CT projector.

If this is right

  • Higher reconstruction accuracy than state-of-the-art methods across diverse datasets.
  • Faster convergence during the optimization process.
  • Support for real-time physical simulation that is clinically relevant.
  • A practical alternative to implicit neural representations for low-dose CT.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same initialization and projector strategy could be tested on other limited-data tomography problems such as limited-angle CT or tomosynthesis.
  • Real-time simulation capability might allow on-the-fly dose estimation during interventional procedures.
  • Because the representation is explicit, it may be easier to combine with hardware-accelerated rendering pipelines than neural fields.

Load-bearing premise

That 3D Gaussian splatting, once initialized from FBP images and coupled to a CT projector, will reliably outperform implicit neural representations on sparse-view CT tasks.

What would settle it

A head-to-head test on a new clinical sparse-view CT dataset in which reconstruction error or optimization time is equal to or worse than the leading neural baseline.

Figures

Figures reproduced from arXiv: 2312.15676 by Han Li, Ruiyang Jin, Shang Zhao, S. Kevin Zhou, Xueming Fu, Yingtai Li.

Figure 1
Figure 1. Figure 1: Comparison of 3D Gaussian Representation (3DGR) and Implicit Neural Representation (INR). (a) Prior-Informed Initialization: Our FBP-image guided initialization technique effectively avoids placing Gaussians in void regions and allocates their density (corresponds to modeling capacity) according to region complexity. Computations for each voxel only involve nearby Gaussians. In contrast, INR need to comput… view at source ↗
Figure 2
Figure 2. Figure 2: Our method achieves similar quantitative performance with state-of-the-art INR based method with only half of its time, and continue to arrive at a much better visual and quantitative result given more time. 2.2. Related Work Basis Functions in CT Reconstruction Basis functions play a fundamental role in computerized tomography reconstruction (Herman, 2009). The choice and properties of basis functions sig… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of pipeline: Initially, projection data is acquired from various viewpoints and subsequently processed using Filtered Back Projection (FBP) to yield the FBP-reconstructed image. This reconstructed image is then utilized as a prior for initializing the 3D Gaussians. During the training phase, these 3D Gaussians are discretized into a volumetric image, which then undergoes a forward projection proce… view at source ↗
Figure 4
Figure 4. Figure 4: Visualization of reconstructed images. The window level and width are optimized for visualizing brain tissue, while other anatomical structures are displayed using standard CT value ranges. Observe that NeRP, without prior information, generates blurry reconstructions lacking intricate details. NAF, on the other hand, is prone to introducing additional noise and artifacts. Compared to INRs, our proposed 3D… view at source ↗
Figure 5
Figure 5. Figure 5: Zoomed-in comparison of critical anatomical structures. optimizer, with (𝛽1 , 𝛽2 ) = (0.9, 0.999), the learning rate for 𝜇 starts from 2e-3, with an exponential decay to 2e-6 by the end of the training process. We employ constant learning rates of 0.05, 0.005, and 0.001 for intensity 𝑡, scaling 𝑠, and rotation 𝑞 respectively. All models are trained for 15,000 iterations on an Nvidia 3090 GPU. Reconstructed… view at source ↗
Figure 7
Figure 7. Figure 7: Comparison of performance under different number of projections. The radius represents standard deviation. 4.2. Main Results We compare the proposed 3D Gaussian representation with INR based methods, such as NeRP without prior (Shen et al., 2022) and NAF (Zha et al., 2022). NeRP represents im￾ages using a multi-layer perceptron with a Fourier encoder, and NAF (Zha et al., 2022) replace Fourier encoder with… view at source ↗
Figure 6
Figure 6. Figure 6: Comparison on energy distribution difference. 20 40 80 120 Number of Projections 24 32 40 48 PSNR (dB) Comparison under different number of Projections FBP SART ASD-POCS NeRP NAF Ours [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: Demonstration of physical simulation with 3D Guassian representation. We use a blue arrow to represent the applied point and direction of the transient implusive force. The second right columns display the maximum deformation of the coronary artery. The last columns show the state at the time point when the coronary artery first begins to return to its initial position. 4.3. Demonstration of Physical Simul… view at source ↗
Figure 9
Figure 9. Figure 9: compares our proposed FBP-image-guided initial￾ization strategy with a uniform initialization approach. For uniform initialization, we use the same threshold 𝜏 to filter the air regions and initialize Gaussians uniformly across the foreground. All Gaussisans are initialized with the same covariance and intensity. The results clearly demonstrate that initializing Gaussians from FBP reconstructions consisten… view at source ↗
Figure 10
Figure 10. Figure 10: Ablation on number of Gaussians and adaptive density control. 4.4.3. Number of Gaussian Functions. To understand the number of Gaussian functions and the reconstruction quality, we experiment with a range of Gaussian counts from 5,000 to 500,000. These experiments are conducted without employing adaptive density control to avoid the perturbation on Gaussian numbers. As shown in [PITH_FULL_IMAGE:figures/f… view at source ↗
Figure 12
Figure 12. Figure 12: Ablation on activation function for intensity and scaling. performance compared to using the sigmoid function. For scaling, we experiment with both sigmoid and exponential activation functions. We observe that both choices result in comparable performance, while the exponential activation function generally performs slightly better. 4.4.7. Adaptive Density Control Thresholds We conduct an ablation study t… view at source ↗
Figure 11
Figure 11. Figure 11: Ablation on learning rate for scaling. 4.4.5. Learning Rate for Gaussian Scale [PITH_FULL_IMAGE:figures/full_fig_p012_11.png] view at source ↗
Figure 13
Figure 13. Figure 13: Reconstruction of real-world data. (a) Example of raw projection data. (b), (c), (d) are the reconstruction results from different views. 5.2. Comparison of interpoltation ability of Gaussian representation with INRs We compare the interpolation ability of Gaussian rep￾resentations with INRs through following experiments: we directly fitting a low resolution image (which is obtained through downsampling t… view at source ↗
read the original abstract

Sparse-view computed tomography (CT) reduces radiation exposure by acquiring fewer projections, making it a valuable tool in clinical scenarios where low-dose radiation is essential. However, this often results in increased noise and artifacts due to limited data. In this paper we propose a novel 3D Gaussian representation (3DGR) based method for sparse-view CT reconstruction. Inspired by recent success in novel view synthesis driven by 3D Gaussian splatting, we leverage the efficiency and expressiveness of 3D Gaussian representation as an alternative to implicit neural representation. To unleash the potential of 3DGR for CT imaging scenario, we propose two key innovations: (i) FBP-image-guided Guassian initialization and (ii) efficient integration with a differentiable CT projector. Extensive experiments and ablations on diverse datasets demonstrate the proposed 3DGR-CT consistently outperforms state-of-the-art counterpart methods, achieving higher reconstruction accuracy with faster convergence. Furthermore, we showcase the potential of 3DGR-CT for real-time physical simulation, which holds important clinical applications while challenging for implicit neural representations.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

0 major / 1 minor

Summary. The paper proposes 3DGR-CT, a sparse-view CT reconstruction method that replaces implicit neural representations with an explicit 3D Gaussian representation. It introduces two adaptations—FBP-image-guided Gaussian initialization and integration with a differentiable CT projector—and claims that extensive experiments and ablations on diverse datasets show consistent outperformance over state-of-the-art methods in reconstruction accuracy and convergence speed, plus potential for real-time physical simulation with clinical value.

Significance. If the empirical performance claims hold, the work could be significant by demonstrating that explicit 3D Gaussian representations can serve as a faster-converging, more efficient alternative to implicit methods for low-dose CT, with added benefits for real-time simulation applications that are difficult for NeRF-style approaches.

minor comments (1)
  1. [Abstract] Abstract: the claim of outperformance, higher accuracy, and faster convergence is stated without any quantitative metrics, dataset names, baseline methods, or ablation summaries, which weakens the abstract's ability to convey the central empirical result.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the constructive summary and positive recommendation of minor revision. We are pleased that the potential significance of explicit 3D Gaussian representations for sparse-view CT is recognized. Since no specific major comments were raised, we will focus on addressing any minor issues identified during the revision process.

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper presents an empirical method for sparse-view CT reconstruction using 3D Gaussian splatting with two proposed adaptations (FBP-guided initialization and differentiable projector integration). The central claim is that these adaptations yield superior accuracy and convergence on tested datasets, validated via experiments and ablations. No derivation chain, fitted parameter renamed as prediction, or self-citation load-bearing the result is present; the argument rests on direct experimental comparison rather than reducing to its own inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review provides no explicit free parameters, axioms, or invented entities; no details on optimization hyperparameters or modeling assumptions are stated.

pith-pipeline@v0.9.0 · 5734 in / 998 out tokens · 27621 ms · 2026-05-24T05:51:14.464008+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. A Survey on 3D Gaussian Splatting

    cs.CV 2024-01 unverdicted novelty 2.0

    A survey compiling principles, applications, benchmarks, and challenges of 3D Gaussian Splatting for explicit 3D scene representation.

Reference graph

Works this paper leans on

27 extracted references · 27 canonical work pages · cited by 1 Pith paper

  1. [1]

    URL: https://doi.org/10.5281/zenodo.249479

    Operator discretization library (odl). URL: https://doi.org/10.5281/zenodo.249479. Andersen, A.H., Kak, A.C.,

    work page doi:10.5281/zenodo.249479

  2. [2]

    IEEE Conf

    Lose the views: Limited angle CT reconstruction via implicit sinogram completion, in: Proc. IEEE Conf. Comput. Vis. and Pattern Recog., pp. 6343–6352. Barron,J.T.,Mildenhall,B.,Verbin,D.,Srinivasan,P.P.,Hedman,P.,2022. Mip-nerf 360: Unbounded anti-aliased neural radiance fields, in: Proc. IEEE Int. Conf. Comput. Vis., pp. 5470–5479. Bian,J.,Siewerdsen,J.H...

    work page 2022

  3. [3]

    X2vision:3dCT reconstruction from biplanar x-rays with deep structure prior, in: Lect

    Cafaro, A., Spinat, Q., Leroy, A., Maury, P., Munoz, A., Beldjoudi, G., Robert,C.,Deutsch,E.,Grégoire,V.,Lepetit,V.,2023. X2vision:3dCT reconstruction from biplanar x-rays with deep structure prior, in: Lect. Notes Comput. Sci., Springer. Springer. pp. 699–709. Cai,Y.,Liang,Y.,Wang,J.,Wang,A.,Zhang,Y.,Yang,X.,Zhou,Z.,Yuille, A.,

    work page 2023

  4. [4]

    arXiv preprint arXiv:2403.04116

    Radiative gaussian splatting for efficient x-ray novel view synthesis. arXiv preprint arXiv:2403.04116 . Carr, J.C., Beatson, R.K., Cherrie, J.B., Mitchell, T.J., Fright, W.R., Mc- Callum, B.C., Evans, T.R.,

    work page arXiv

  5. [5]

    Chen, Y., Chen, Z., Zhang, C., Wang, F., Yang, X., Wang, Y., Cai, Z., Yang,L.,Liu,H.,Lin,G.,2024

    Does 3d gaussian splatting need accurate volumetric rendering? arXiv preprint arXiv:2502.19318 . Chen, Y., Chen, Z., Zhang, C., Wang, F., Yang, X., Wang, Y., Cai, Z., Yang,L.,Liu,H.,Lin,G.,2024. Gaussianeditor:Swiftandcontrollable 3d editing with gaussian splatting, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp....

    work page arXiv 2024

  6. [6]

    IEEE Conf

    Solving 3d inverse problems using pre-trained 2d diffusion models, in: Proc. IEEE Conf. Comput. Vis. and Pattern Recog., pp. 22542–22551. Ding,C.,Zhang,Q.,Wang,G.,Ye,X.,Chen,Y.,2023. Learnedalternating minimizationalgorithmfordual-domainsparse-viewCTreconstruction, in: Lect. Notes Comput. Sci., Springer. pp. 173–183. Ding, Q., Ji, H., Gao, H., Zhang, X.,

    work page 2023

  7. [7]

    arXiv preprint arXiv:2211.17048

    Snaf: Sparse-view cbct reconstruction with neural attenuation fields. arXiv preprint arXiv:2211.17048 . Feldkamp, L.A., Davis, L.C., Kress, J.W.,

    work page arXiv

  8. [8]

    Josa a 1, 612–619

    Practical cone-beam algorithm. Josa a 1, 612–619. Gao,Z.,Planche,B.,Zheng,M.,Chen,X.,Chen,T.,Wu,Z.,2024. DDGS- CT: direction-disentangled gaussian splatting for realistic volume ren- dering. arXiv preprint arXiv:2406.02518 . Ge, R., He, Y., Xia, C., Sun, H., Zhang, Y., Hu, D., Chen, S., Chen, Y., Li, S., Zhang, D.,

    work page arXiv 2024

  9. [9]

    Ddpnet: a novel dual-domain parallel network for low-doseCTreconstruction,in:Lect.NotesComput.Sci.,pp.748–757. Gharleghi, R., Adikari, D., Ellenberger, K., Ooi, S.Y., Ellis, C., Chen, C.M., Gao, R., He, Y., Hussain, R., Lee, C.Y., Li, J., Ma, J., Nie, Z., Oliveira, B., Qi, Y., Skandarani, Y., Vilaça, J.L., Wang, X., Yang, S., Sowmya, A., Beier, S., 2022a....

    work page doi:10.1016/j.compmedimag.2022.102049 2022

  10. [10]

    Eagles: Efficient accelerated 3d gaussians with lightweight encodings, in: European Conference on Computer Vision, Springer. pp. 54–71. Guo, P., Zhao, C., Yang, D., Xu, Z., Nath, V., Tang, Y., Simon, B., Belue, M.,Harmon,S.,Turkbey,B.,etal.,2024. Maisi:Medicalaiforsynthetic imaging. arXiv preprint arXiv:2409.11169 . Herman, G.T.,

    work page arXiv 2024

  11. [11]

    Springer Science & Business Media

    Fundamentals of computerized tomography: image reconstruction from projections. Springer Science & Business Media. hetan, A., Yang, G., Wang, Z., Marschner, S., Hariharan, B., . Accurate differential operators for hybrid neural fields. CoRR abs/2312.05984. Ho,J.,Jain,A.,Abbeel,P.,2020. Denoisingdiffusionprobabilisticmodels. Advances in neural information ...

    work page arXiv 2020

  12. [12]

    arXiv preprint arXiv:2407.01029

    Endosparse:Real-timesparseviewsynthesisofendoscopicscenesusing gaussian splatting. arXiv preprint arXiv:2407.01029 . Liebl, H., Schinz, D., Sekuboyina, A., Malagutti, L., Löffler, M.T., Bayat, A., El Husseini, M., Tetteh, G., Grau, K., Niederreiter, E., Baum, T., Wiestler,B.,Menze,B.,Braren,R.,Zimmer,C.,Kirschke,J.S.,2021. A computed tomography vertebral ...

    work page arXiv 2021

  13. [13]

    arXiv preprint arXiv:2407.01090

    Learning 3d gaussians for extremely sparse-view cone-beam CT reconstruction. arXiv preprint arXiv:2407.01090 . Liu, H., Liu, Y., Li, C., Li, W., Yuan, Y.,

    work page arXiv

  14. [14]

    arXiv preprint arXiv:2406.16073

    LGS: A light-weight 4d gaussian splatting for efficient surgical scene reconstruction. arXiv preprint arXiv:2406.16073 . Löffler, M.T., Sekuboyina, A., Jacob, A., Grau, A.L., Scharr, A., El Hus- seini, M., Kallweit, M., Zimmer, C., Baum, T., Kirschke, J.S.,

    work page arXiv

  15. [15]

    Low-dose ct image and projection dataset. Med. phys. 48, 902–911. Molaei, A., Aminimehr, A., Tavakoli, A., Kazerouni, A., Azad, B., Azad, R.,Merhof,D.,2023. Implicitneuralrepresentationinmedicalimaging: A comparative survey, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 2381–2391. Müller, T., Evans, A., Schied, C., Keller, A.,

    work page 2023

  16. [16]

    Nilchian, M., Ward, J.P., Vonesch, C., Unser, M.,

    Gaspct:GaussiansplattingfornovelCTprojectionviewsynthesis.arXiv preprint arXiv:2404.03126 . Nilchian, M., Ward, J.P., Vonesch, C., Unser, M.,

    work page arXiv

  17. [17]

    IEEETrans- actions on Image Processing 24, 3826–3833

    Optimized kaiser–besselwindowfunctionsforcomputedtomography. IEEETrans- actions on Image Processing 24, 3826–3833. doi:10.1109/TIP.2015. 2451955. Park, S., Kim, S., Song, I.S., Baek, S.J.,

    work page doi:10.1109/tip.2015 2015

  18. [18]

    Notes Comput

    3d teeth reconstruction from panoramic radiographs using neural implicit functions, in: Lect. Notes Comput. Sci., Springer. pp. 376–386. Rahaman, N., Baratin, A., Arpit, D., Draxler, F., Lin, M., Hamprecht, F., Bengio,Y.,Courville,A.,2019. Onthespectralbiasofneuralnetworks, in: ICML, pp. 5301–5310. Reed, A.W., Kim, H., Anirudh, R., Mohan, K.A., Champley, ...

    work page 2019

  19. [19]

    Medical Image Analysis 73, 102166

    Verse: A vertebrae labelling and segmentation benchmark for multi-detector ct images. Medical Image Analysis 73, 102166. URL: https://www.sciencedirect.com/science/article/pii/ S1361841521002127, doi:https://doi.org/10.1016/j.media.2021.102166. Shen, L., Pauly, J., Xing, L.,

    work page doi:10.1016/j.media.2021.102166 2021

  20. [20]

    Solving inverse problems in med- ical imaging with score-based generative models

    Song, Y., Shen, L., Xing, L., Ermon, S., 2021a. Solving inverse problems in medical imaging with score-based generative models. arXiv preprint arXiv:2111.08005 . Song,Y.,Sohl-Dickstein,J.,Kingma,D.P.,Kumar,A.,Ermon,S.,Poole,B., 2021b. Score-based generative modeling through stochastic differential equations,in:Int.Conf.onLearn.Represent. URL: https://open...

    work page arXiv

  21. [21]

    arXiv preprint arXiv:2102.05181

    Coil: Coordinate-basedinternallearningforimaginginverseproblems. arXiv preprint arXiv:2102.05181 . Takikawa,T.,Müller,T.,Nimier-David,M.,Evans,A.,Fidler,S.,Jacobson, A., Keller, A.,

    work page arXiv

  22. [22]

    - SIGGRAPH Asia Conf

    Compact neural graphics primitives with learned hash probing, in: Proc. - SIGGRAPH Asia Conf. Pap., SA, pp. 120:1– 120:10. Tancik,M.,Srinivasan,P.,Mildenhall,B.,Fridovich-Keil,S.,Raghavan,N., Singhal,U.,Ramamoorthi,R.,Barron,J.,Ng,R.,2020. Fourierfeatures letnetworkslearnhighfrequencyfunctionsinlowdimensionaldomains. Advances in neural information process...

    work page 2020

  23. [23]

    Notes Comput

    Mepnet: A model- drivenequivariantproximalnetworkforjointsparse-viewreconstruction and metal artifact reduction in ct images, in: Lect. Notes Comput. Sci., Springer. pp. 109–120. Wang, K., Yang, C., Wang, Y., Li, S., Wang, Y., Dou, Q., Yang, X., Shen, W., 2024a. Endogslam: Real-time dense reconstruction and trackinginendoscopicsurgeriesusinggaussiansplatt...

    work page arXiv 2024

  24. [24]

    arXiv preprint arXiv:2405.20693

    R2-gaussian: Rectifying radiative gaussian splatting for tomographic reconstruction. arXiv preprint arXiv:2405.20693 . Zha, R., Zhang, Y., Li, H.,

    work page arXiv

  25. [25]

    Notes Comput

    Naf: neural attenuation fields for sparse- view cbct reconstruction, in: Lect. Notes Comput. Sci., pp. 442–452. Zhang,J.,Chao,H.,Xu,X.,Niu,C.,Wang,G.,Yan,P.,2021. Task-oriented low-dose CT image denoising, in: Lect. Notes Comput. Sci., pp. 441–

    work page 2021

  26. [26]

    TOGS: gaussian splatting with temporal opacity offset for real-time 4d DSA rendering

    Zhang,S.,Zhao,H.,Zhou,Z.,Wu,G.,Zheng,C.,Wang,X.,Liu,W.,2024. TOGS: gaussian splatting with temporal opacity offset for real-time 4d DSA rendering. arXiv preprint arXiv:2403.19586 . Zhang, Z., Liang, X., Dong, X., Xie, Y., Cao, G.,

    work page arXiv 2024

  27. [27]

    arXiv preprint arXiv:2401.11535

    Deformable endo- scopic tissues reconstruction with gaussian splatting. arXiv preprint arXiv:2401.11535 . Zwicker, M., Pfister, H., Van Baar, J., Gross, M.,

    work page arXiv