Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

GR-Gaussian: Graph-Based Radiative Gaussian Splatting for Sparse-View CT Reconstruction

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Graph-based density-aware gradients suppress needle artifacts and improve sparse-view CT reconstruction.

desk verdict A plausible, incremental extension of R2-Gaussian with a useful denoised initialization and a graph-aware splitting heuristic, but the gradient mechanism in Eq. 16 is mathematically unsupported and the paper needs major revision before the numbers can be trusted. read the letter →

arxiv 2508.02408 v2 pith:INGGJS77 submitted 2025-08-04 eess.IV cs.CV

classification eess.IVcs.CV
keywords sparse-viewCTreconstruction3DGaussianSplattingradiativegraph-basedgradientneedle-likeartifactsuppressionpointcloudinitializationtomographicdensityfieldestimation
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 sets out to fix a specific failure of 3D Gaussian Splatting when applied to sparse-view CT reconstruction: the appearance of needle-like artifacts, which it attributes to splitting decisions that rely only on per-point average gradient magnitude. Its central claim is that representing the scanned object as a graph of radiative Gaussians and augmenting the splitting gradient with a graph-based density-difference term turns large, low-gradient kernels into proper split candidates, so the density field is refined instead of leaving long thin artifacts. It also claims that denoising the FDK-initialized point cloud before optimization reduces initialization error and accelerates convergence. If correct, the method would make high-quality CT reconstruction feasible from as few as 25 projection views, which matters because fewer views mean lower radiation dose and shorter scans. The reported gains over the strongest prior 3DGS method are 0.67 dB and 0.92 dB PSNR and 0.011 and 0.021 SSIM on simulated and real-world datasets.

What carries the argument

The load-bearing object is the Pixel-Graph-Aware Gradient (PGA), defined by Eq. (16) as $(g^c_i)_v = \sum_{pix=1}^{m^v_i} \frac{\partial L_v}{\partial \alpha^{v,pix}_i}\frac{\partial \alpha^{v,pix}_i}{\partial \mu^{i,v}_{ndc}} + \lambda_g \frac{\sum_{j\in\mathcal{N}(i)} \Delta\rho_{ij}}{k}$, where $\Delta\rho_{ij}=|\rho_i-\rho_j|$ is the absolute density difference between Gaussian $i$ and its graph neighbor $j$, $\mathcal{N}(i)$ is the KNN neighbor set, $k$ is a scaling factor, and $\lambda_g$ is a weight. This adds a scalar graph-density correction to the vector pixel gradient in NDC space, so kernels near density boundaries receive larger gradient magnitudes even when their per-pixel gradients are small. That pushes them over the splitting threshold $\tau_{pos}$ used by adaptive density control, which is the mechanism the paper invokes to explain why needle-like artifacts are suppressed and why the density field is better represented. The Denoised Point Cloud Initialization (Gaussian filtering of the FDK volume, Eq. (14)) and the graph-Laplacian regularization Eq. (18) support the same goal but are secondary.

What would settle it

Replace the density-difference term $\lambda_g \sum_{j\in\mathcal{N}(i)} \Delta\rho_{ij}/k$ in Eq. (16) with a random scalar perturbation of matched magnitude while keeping the same splitting schedule; if the PSNR gain over the baseline persists, the graph-density signal is not the cause of the artifact suppression, whereas if the gain disappears the specific density-difference content is doing the work, as the paper claims.

Watch

Extended reading notes

Core claim

GR-Gaussian's central discovery is that the needle artifacts in sparse-view radiative Gaussian splatting are not cured by better rendering but by changing what triggers densification. The paper represents the object as a graph whose vertices are Gaussian kernels, edges from bidirectional KNN, and each kernel's field includes weighted contributions from neighbors. It then replaces the usual average NDC-space gradient with an augmented gradient that adds $\lambda_g \frac{\sum_{j\in\mathcal{N}(i)} \Delta\rho_{ij}}{k}$ — a scaled sum of absolute density differences between a kernel and its graph neighbors — to the pixel-driven gradient. This makes kernels sitting at density boundaries, which may have small pixel gradients, generate large enough magnitudes to pass the splitting threshold $\tau_{pos}$, so they are split rather than left as elongated needles. A denoised initialization from a Gaussian-filtered FDK volume and a graph-Laplacian smoothness term complete the framework. The paper reports that this combination outperforms prior NeRF- and 3DGS-based methods on the X-3D and real-world 25-view datasets in both PSNR and SSIM, with visual results showing fewer streak artifacts.

Load-bearing premise

The method assumes that the augmented quantity in Eq. (16) — a pixel gradient plus a scalar density-difference penalty — is a meaningful gradient-like signal for splitting decisions, even though no derivation in the paper shows it is a true gradient or a valid descent direction.

Editorial extensions

If this is right

  • If the method is right, sparse-view CT reconstruction from around 25 projections can approach the quality of denser-view reconstructions without changing scanner hardware.
  • The graph-based augmented gradient should reduce needle-like artifacts in any radiative-Gaussian inverse problem where density is piecewise constant, not just CT.
  • The dynamic stopping criterion means the reported gains come with a built-in guard against overfitting on noisy real-world projections, making the results reproducible under the stated protocol.
  • Ablation results indicate the two innovations are complementary: the denoised initialization contributes most on real-world noise, while the graph-aware gradient contributes most in smooth regions.

Reading between the lines

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

  • Editorial extension: the density-difference term in Eq. (16) is functionally similar to a graph-Laplacian penalty on density, so the PGA mechanism may be re-expressing in gradient form the smoothness that the loss already enforces through $L_{lap}$; isolating the two effects would clarify which component actually carries the artifact suppression.
  • Editorial extension: all headline comparisons are at 25 views, so a natural stress test is 10-view and 50-view settings to see whether the graph-gradient advantage grows, shrinks, or plateaus as sparsity changes.
  • Editorial extension: because the gradient augmentation is color-free and density-based, the same idea could transfer to other mono-modal inverse problems such as PET or SPECT, where attenuation is also roughly piecewise constant.
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 / 5 minor

Summary. The paper proposes GR-Gaussian, a 3D Gaussian Splatting framework for sparse-view CT reconstruction. It augments the R2-Gaussian baseline with two components: a denoised FDK-based point cloud initialization (De-Init) and a pixel-graph-aware gradient strategy (PGA) that adds a KNN density-difference term to the per-Gaussian gradient, together with graph Laplacian regularization. Experiments on the simulated X-3D dataset and a real-world CT dataset under 25-view conditions report PSNR/SSIM improvements of 0.67 dB/0.011 and 0.92 dB/0.021 over R2-Gaussian, with ablations attributing gains to both components.

Significance. If the empirical claims hold, the contribution is modest but useful: consistent gains across two datasets and several object categories, supported by component ablations and a sensitivity analysis for k and sigma_d. The real-world validation and the visual reduction of needle-like artifacts are concrete strengths. The main caveat is that the central PGA mechanism, as written in Eq. (16), is not a mathematically well-defined gradient, so the mechanistic explanation for artifact suppression is unsupported unless the term is rederived or explicitly reframed as a heuristic. The small absolute gains and the absence of variance reporting temper, but do not eliminate, the practical interest of the method.

major comments (3)
  1. [§3.3.2, Eqs. (16)–(17)] Equation (16) is dimensionally ill-defined: the first term is a 2D NDC-space gradient vector, while the second term lambda_g * sum_{j in N(i)} Delta_rho_ij / k is a nonnegative scalar, and no operation is specified for adding them. Equation (17) then states that (g^c_i)_v is proportional to Delta_rho_ij, which can hold only if the photometric gradient term vanishes, but this is not argued. More fundamentally, Delta_rho_ij depends on the densities rho_i and rho_j, not on the NDC positions mu^{i,v}_{ndc}, so the added term is not the derivative of any term in L_total (Eq. (20)) with respect to mu^{i,v}_{ndc}; the graph Laplacian in Eq. (18) also depends on rho, not on mu. The PGA component is therefore not a gradient or a guaranteed descent direction, and the mechanistic explanation that it improves splitting accuracy is unsupported. The Table 2 ablation suggests a real heuristic benefit, but the paper should either derive the term as a proper gradient of a modified loss or present it explicitly as a heuristic augmentation and drop the gradient/proportionality claims.
  2. [§4.3.2, Table 3] The text states that k = 6 and sigma_d = 3 achieve the optimal reconstruction quality, but Table 3 shows that k = 8 yields higher PSNR and SSIM on both datasets (35.89/0.934 on X-3D and 36.01/0.860 on real-world, versus 35.86/0.933 and 35.95/0.858 for k = 6). If k = 6 is chosen as an efficiency-quality tradeoff, the paper should say so and report the tradeoff explicitly. In addition, both k and sigma_d are selected on the same X-3D and real-world datasets used for the final comparisons (Tables 1 and 3), which introduces a mild circularity; a validation split or a clear statement of the selection protocol is needed to support the state-of-the-art claim.
  3. [§4.1.2 and §4.3.3] No variance across runs is reported anywhere, and the dynamic stopping criterion (Iter_stop) evaluates PSNR every 500 iterations and stops when PSNR declines by more than 0.5%, with PSNR being the same metric used for final evaluation. Early stopping on the evaluation metric can bias the reported numbers upward. The authors should report mean +/- std over at least three seeds and clarify whether the stopping PSNR is computed on the same slices/volumes as the final reported PSNR, or use a validation-based stopping rule.
minor comments (5)
  1. [§3.2, Eqs. (7) and (12)] The symbol k is used both for the number of KNN neighbors in Eq. (7) and as the scaling factor in the edge-weight denominator of Eq. (12); these are different quantities and should be denoted separately.
  2. [§4.1.1–4.1.2] There are typos: 'ponton scatter' should be 'photon scatter' (Sec. 4.1.1), and 'television volume level' should be 'total variation volume' (Sec. 4.1.2).
  3. [§4.3.2, Table 4] Table 4 introduces the 'SSGU extension' and reports SDS/DDS/CoSD timings, but SSGU is never defined in the method or experiments, and its connection to the proposed framework is unclear.
  4. [§4.2, Table 1] It is not stated whether the baseline numbers (FDK, SART, ASD-POCS, NAF, SAX-NeRF, R2-GS) are rerun under identical conditions or taken from prior papers; this should be clarified for a fair comparison.
  5. [Fig. 4] The caption's equation g_i^c = g_i^{c'} + lambda_g * sum Delta_rho_ij (mu_ndc,x, mu_ndc,y) is ambiguous: if the scalar is broadcast to both components, this should be stated explicitly, since it directly relates to the issue in Eq. (16).

Circularity Check

2 steps flagged · score 2.0 of 10

PGA's density-variation claim is tautological (Eq. 17 restates the term inserted in Eq. 16), and final PSNR gains use hyperparameters and early stopping on the same test datasets; the central empirical comparison is still externally anchored.

  1. self definitional [Section 3.3.2, Eqs. (16)-(17)]
    "the augmented gradient is defined as: (g_i^c)_v = Σ_{pix=1}^{m_i^v} ∂L_v/∂α_i^{v,pix} · ∂α_i^{v,pix}/∂μ_{ndc}^{i,v} + λ_g · (Σ_{j∈N(i)} Δρ_ij)/k ... By leveraging graph-based relationships, the augmented gradient effectively captures density variations: (g_i^c)_v ∝ Δρ_ij, ∀j ∈ N(i)."

    Equation (17) is not a derived consequence of the rendering model; it is a restatement of Equation (16) because the density-difference sum was inserted directly into the definition of (g_i^c)_v. The photometric gradient term depends on the NDC positions μ_ndc, not on the densities ρ_i and ρ_j, so the only part of the augmented quantity that has any built-in dependence on Δρ_ij is the added λ_g·ΣΔρ/k term. Thus the paper's mechanistic explanation that the augmented gradient 'captures density variations' is true by construction, and the claimed improvement in splitting decisions is attributed to a property that was put in by hand, not to a gradient of any term in L_total (Eq. 20).

  2. fitted input called prediction [Section 4.3.2 (Table 3), Section 4.1.2, Section 4.2 (Table 1)]
    "Sensitivity analysis on the X-3D and real-world datasets (Table 3) shows optimal reconstruction quality at k = 6 and σd = 3, achieving the highest PSNR and SSIM. ... A dynamic stopping criterion (Iter stop) evaluates PSNR every 500 iterations, terminating if PSNR decreases by more than 0.5%. ... it achieves a PSNR increase of 0.67 dB and an SSIM improvement of 0.011."

    The reported improvements in Table 1 are measured on the same X-3D and real-world datasets that were used in Table 3 to analyze and select the hyperparameters k and σd, and the dynamic stopping criterion directly uses the PSNR computed on those same datasets. The final numbers are therefore not independent evaluations: the model's hyperparameters and stopping point are chosen using the evaluation metric on the evaluation set, making the headline PSNR/SSIM gains test-set-optimized rather than out-of-sample predictions. This is an evaluation-loop circularity, distinct from the logical derivation of the method.

full rationale

The paper's central empirical claim — that GR-Gaussian outperforms R2-Gaussian by 0.67/0.92 dB PSNR on the X-3D and real-world datasets — is an externally anchored measurement against ground-truth volumes and independent baselines, not an algebraically forced consequence of the model's definition. There is no load-bearing self-citation: the cited [45] is an external baseline, and the authors' earlier works [27]-[31] are unrelated background. The method is therefore not a renamed known result and does not import a uniqueness theorem from the authors. Two localized issues contribute a small circular component. First, Eq. (17) is presented as evidence that the augmented gradient captures density variations, but this is tautological because the density-difference term is inserted into Eq. (16); the photometric gradient term does not depend on ρ, so Eq. (17) restates the definition. Second, the evaluation loop is circular: k and σd are analyzed/selected on the same X-3D and real-world datasets used for the final comparisons (Table 3 vs Table 1), and the dynamic stopping criterion uses test PSNR to decide when to stop. The scalar-to-vector addition in Eq. (16) and the unsupported proportionality in Eq. (17) are also correctness risks — the augmented quantity is not a true gradient of L_total — but that is a mechanistic weakness, not itself a circularity. Overall, because the central contribution is an empirical method whose headline result does not reduce to its fit or to a self-citation chain, the circularity score is low.

Assumptions & free parameters 10 free parameters · 7 assumptions · 0 invented entities

The central claim rests on a set of hand-chosen loss weights, a KNN graph constructed on Gaussian positions, and the domain assumption that density differences between neighbors are a useful splitting signal. No new physical entities are introduced; the 'graph-structured radiative Gaussian' is a modeling construct built from existing Gaussian kernels plus edges. The grad student check is therefore on the heuristic gradient augmentation and the hyperparameter choices.

free parameters (10)
  • lambda_g (graph gradient weight) = 1e-4
    Weight of the density-difference augmentation in Eq. 16; set by hand and not included in the sensitivity analysis.
  • k (KNN neighbors) = 6
    Number of graph neighbors; chosen from Table 3 sensitivity analysis on the same X-3D and real-world datasets used for the final comparison.
  • sigma_d (denoising width) = 3
    Gaussian filtering standard deviation in the Denoised Point Cloud Initialization; chosen from Table 3 sensitivity analysis on the test datasets.
  • lambda_lap = 8e-4
    Weight of graph Laplacian regularization in Eq. 19.
  • lambda_tv = 0.05
    Weight of 3D total variation regularization.
  • lambda_ssim = 0.25
    Weight of D-SSIM loss.
  • tau (density threshold) = 0.001
    Threshold for excluding empty regions during point cloud sampling.
  • M (number of Gaussians) = 50000
    Kernel count set in implementation details.
  • Iter_stop PSNR decline threshold = 0.5%
    Dynamic stopping criterion: terminate if PSNR drops more than 0.5% over 500 iterations; may select based on test PSNR.
  • edge weight denominator k in Eq. 12 = unstated
    The scaling factor in Eq. 12 controls sensitivity of edge weights to distance; the paper does not specify whether it equals the KNN count k.
assumptions (7)
  • domain assumption X-ray attenuation can be represented by an isotropic density field without view-dependent color.
    Used in Sec. 3.1.1 to justify omitting color channels from radiative Gaussians.
  • domain assumption Similar tissues and materials have approximately constant attenuation coefficients, so density differences between neighboring kernels are a meaningful splitting signal.
    Used in Sec. 3.3.2 to motivate Eq. 16.
  • domain assumption Gaussian filtering of the FDK volume removes sparse-view noise while preserving structural detail.
    Assumed in De-Init (Sec. 3.3.1); no quantitative validation beyond final PSNR.
  • domain assumption The graph Laplacian penalty on densities preserves boundaries while smoothing interiors.
    Used in Sec. 3.3.4, Eq. 18.
  • ad hoc to paper A KNN graph with small k captures relevant spatial relationships for splitting decisions.
    Chosen via sensitivity analysis; no theoretical justification.
  • domain assumption Differentiable voxelization and 3D tile culling approximate the continuous density field accurately.
    Inherited from the R2-Gaussian voxelizer; Sec. 3.3.3.
  • standard math The rendering and voxelization operations are differentiable, enabling end-to-end gradient descent.
    Needed for Adam optimization; standard in 3DGS.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GR-Gaussian: Graph-Based Radiative Gaussian Splatting for Sparse-View CT Reconstruction." pith.science (2026). https://pith.science/paper/INGGJS77

@misc{pith2026250802408,
  author       = {Pith},
  title        = {Pith review of: GR-Gaussian: Graph-Based Radiative Gaussian Splatting for Sparse-View CT Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/INGGJS77}},
  note         = {Machine review of arXiv:2508.02408}
}
read the original abstract

3D Gaussian Splatting (3DGS) has emerged as a promising approach for CT reconstruction. However, existing methods rely on the average gradient magnitude of points within the view, often leading to severe needle-like artifacts under sparse-view conditions. To address this challenge, we propose GR-Gaussian, a graph-based 3D Gaussian Splatting framework that suppresses needle-like artifacts and improves reconstruction accuracy under sparse-view conditions. Our framework introduces two key innovations: (1) a Denoised Point Cloud Initialization Strategy that reduces initialization errors and accelerates convergence; and (2) a Pixel-Graph-Aware Gradient Strategy that refines gradient computation using graph-based density differences, improving splitting accuracy and density representation. Experiments on X-3D and real-world datasets validate the effectiveness of GR-Gaussian, achieving PSNR improvements of 0.67 dB and 0.92 dB, and SSIM gains of 0.011 and 0.021. These results highlight the applicability of GR-Gaussian for accurate CT reconstruction under challenging sparse-view conditions.

Figures

Figures reproduced from arXiv: 2508.02408 by the authors.

Figure 1
Figure 1. Visual results of GR-GAUSSIAN. We compare GR-GAUSSIAN to two NeRF-based methods NAF[44], SAX-NeRF[4]) and R 2 - GS[45] in terms of visual quality and PSNR (dB). Our method mitigates needle-like artifacts and achieves superior CT reconstruction quality under sparse-view conditions. Abstract 3D Gaussian Splatting (3DGS) has emerged as a promis￾ing approach for CT reconstruction. However, existing methods rely on the a… view at source ↗
Figure 2
Figure 2. Training pipeline of GR-Gaussian. (a) Overall training pipeline. (b) Denoised Point Cloud Initialization Strategy. (c) Pixel-Graph￾Aware Gradient Strategy. Deep learning-based methods[5, 23, 24] demonstrate strong performance but require large labeled datasets and long training times[29, 31]. They also struggle with out￾of-distribution objects. NeRF-based[27, 30, 33] approaches show promise in per-case reconstructio… view at source ↗
Figure 3
Figure 3. The scanned object is represented as graph-based radiative Gaussians, optimized using real X-ray projections to retrieve the density volume via voxelization. contribution is computed as: (g r i ) v = ∂Lv ∂µi,v ndc = mi Xv pix=1 X 3 j=1 ∂Lv ∂cpix j · ∂cpix j ∂αi v,pix · ∂αi v,pix ∂µi,v ndc , (4) Here, α i v,pix represents the contribution factor of a pixel to G3 i , c pix j is the intensity of the j-th color channel,… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: The Pixel-Graph-Aware Gradient Strategy leverages density differences between Gaussian kernels by constructing a graph to encode point-to-point relationships, enhancing gradient computation and enabling effective splitting of large kernels with low gradients. 5 [PITH_…
Figure 5
Figure 5. Figure 5: The left shows CT reconstructions from the X-3D dataset, covering three categories: chest, bonsai, and teapot. The right displays Real-world dataset reconstructions with colorized slices to highlight details, all under 25-view conditions. Baseline De-Init PGA X-3D Real…
Figure 6
Figure 6. Figure 6: Ablation study results highlight the impact of PGA and De-Init in enhancing reconstruction quality. the real-world dataset. For the ablation study, the baseline model omits the De-Init and PGA components, with Gaus￾sians initialized via an FDK-based approach. To valida…
Figure 7
Figure 7. Figure 7: Iteration Analysis Reconstruction results of GR-Gaussian across different iterations, illustrating the impact of iteration count on PSNR. ensures that the model avoids overfitting and maintains op￾timal reconstruction quality. 5. Conclusion This paper introduces GR-Gau…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. URS-Stereo: Uncertainty-Guided Residual Search for Real-Time Stereo Matching

    cs.CV 2026-07 conditional novelty 4.5 of 10

    Uncertainty-modulated residual offsets relocate local cost-volume centers in coarse-to-fine stereo matching, improving zero-shot disparity accuracy while keeping real-time speed.

Reference graph

Works this paper leans on

46 extracted references · 36 canonical work pages · cited by 1 Pith paper

  1. [1]

    Simultaneous alge- braic reconstruction technique (sart): a superior implemen- tation of the art algorithm

    Anders H Andersen and Avinash C Kak. Simultaneous alge- braic reconstruction technique (sart): a superior implemen- tation of the art algorithm. Ultrasonic imaging, 6(1):81–94,

  2. [2]

    Lose the views: Limited angle ct reconstruction via implicit sino- gram completion

    Rushil Anirudh, Hyojin Kim, Jayaraman J Thiagarajan, K Aditya Mohan, Kyle Champley, and Timo Bremer. Lose the views: Limited angle ct reconstruction via implicit sino- gram completion. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages 6343– 6352, 2018. 3

  3. [3]

    Radiative gaussian splatting for efficient x-ray novel view synthesis

    Yuanhao Cai, Yixun Liang, Jiahao Wang, Angtian Wang, Yu- lun Zhang, Xiaokang Yang, Zongwei Zhou, and Alan Yuille. Radiative gaussian splatting for efficient x-ray novel view synthesis. In European Conference on Computer Vision , pages 283–299. Springer, 2024. 3

  4. [4]

    Structure-aware sparse-view x-ray 3d re- construction

    Yuanhao Cai, Jiahao Wang, Alan Yuille, Zongwei Zhou, and Angtian Wang. Structure-aware sparse-view x-ray 3d re- construction. In Proceedings of the IEEE/CVF Conference 8 on Computer Vision and Pattern Recognition, pages 11174– 11183, 2024. 1, 2, 3, 6

  5. [5]

    Solving 3d inverse problems us- ing pre-trained 2d diffusion models

    Hyungjin Chung, Dohoon Ryu, Michael T Mccann, Marc L Klasky, and Jong Chul Ye. Solving 3d inverse problems us- ing pre-trained 2d diffusion models. ieee. InCVF Conference on Computer Vision and Pattern Recognition, pages 10–12,

  6. [6]

    Industrial applications of computed tomography

    Leonardo De Chiffre, Simone Carmignato, J-P Kruth, Robert Schmitt, and Albert Weckenmann. Industrial applications of computed tomography. CIRP annals, 63(2):655–677, 2014. 3

  7. [7]

    Adam: A method for stochastic optimiza- tion

    Kingma Diederik. Adam: A method for stochastic optimiza- tion. (No Title), 2014. 6

  8. [8]

    Prac- tical cone-beam algorithm

    Lee A Feldkamp, Lloyd C Davis, and James W Kress. Prac- tical cone-beam algorithm. Journal of the Optical Society of America A, 1(6):612–619, 1984. 1, 3, 6

Show all 46 references
  1. [9]

    Fast graph repre- sentation learning with pytorch geometric

    Matthias Fey and Jan Eric Lenssen. Fast graph repre- sentation learning with pytorch geometric. arXiv preprint arXiv:1903.02428, 2019. 6

  2. [10]

    Deep learning- based sinogram completion for low-dose ct

    Muhammad Usman Ghani and W Clem Karl. Deep learning- based sinogram completion for low-dose ct. In 2018 IEEE 13th Image, Video, and Multidimensional Signal Processing Workshop (IVMSP), pages 1–5. IEEE, 2018. 3

  3. [11]

    Sugar: Surface- aligned gaussian splatting for efficient 3d mesh reconstruc- tion and high-quality mesh rendering

    Antoine Gu ´edon and Vincent Lepetit. Sugar: Surface- aligned gaussian splatting for efficient 3d mesh reconstruc- tion and high-quality mesh rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5354–5363, 2024. 2

  4. [12]

    Cuda c programming guide

    Design Guide. Cuda c programming guide. NVIDIA, July, 29:31, 2013. 6

  5. [13]

    Bloomscene: Lightweight structured 3d gaussian splatting for crossmodal scene generation.Proceed- ings of the AAAI Conference on Artificial Intelligence, 39(4): 3536–3544, 2025

    Xiaolu Hou, Mingcheng Li, Dingkang Yang, Jiawei Chen, Ziyun Qian, Xiao Zhao, Yue Jiang, Jinjie Wei, Qingyao Xu, and Lihua Zhang. Bloomscene: Lightweight structured 3d gaussian splatting for crossmodal scene generation.Proceed- ings of the AAAI Conference on Artificial Intellig...

  6. [14]

    Computed medical imaging

    Godfrey N Hounsfield. Computed medical imaging. Science, 210(4465):22–28, 1980. 3

  7. [15]

    Fatesgs: Fast and accurate sparse-view surface reconstruction using gaussian splatting with depth-feature consistency

    Han Huang, Yulun Wu, Chao Deng, Ge Gao, Ming Gu, and Yu-Shen Liu. Fatesgs: Fast and accurate sparse-view surface reconstruction using gaussian splatting with depth-feature consistency. Proceedings of the AAAI Conference on Arti- ficial Intelligence, 39(4):3644–3652, 2025. 3

  8. [16]

    Computer-aided di- agnosis in chest radiography

    Shigehiko Katsuragawa and Kunio Doi. Computer-aided di- agnosis in chest radiography. Computerized Medical Imag- ing and Graphics, 31(4-5):212–223, 2007. 3

  9. [17]

    3d gaussian splatting for real-time radiance field rendering

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139–1,

  10. [18]

    Improving 3d imag- ing with pre-trained perpendicular 2d diffusion models

    Suhyeon Lee, Hyungjin Chung, Minyoung Park, Jonghyuk Park, Wi-Sun Ryu, and Jong Chul Ye. Improving 3d imag- ing with pre-trained perpendicular 2d diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10710–10720, 2023. 3

  11. [19]

    Micro- macro wavelet-based gaussian splatting for 3d reconstruction from unconstrained images

    Yihui Li, Chengxin Lv, Hongyu Yang, and Di Huang. Micro- macro wavelet-based gaussian splatting for 3d reconstruction from unconstrained images. Proceedings of the AAAI Con- ference on Artificial Intelligence, 39:5057–5065, 2025. 3

  12. [20]

    Ani- matable gaussians: Learning pose-dependent gaussian maps for high-fidelity human avatar modeling

    Zhe Li, Zerong Zheng, Lizhen Wang, and Yebin Liu. Ani- matable gaussians: Learning pose-dependent gaussian maps for high-fidelity human avatar modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19711–19722, 2024. 3

  13. [21]

    Gs-ir: 3d gaussian splatting for inverse rendering

    Zhihao Liang, Qi Zhang, Ying Feng, Ying Shan, and Kui Jia. Gs-ir: 3d gaussian splatting for inverse rendering. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21644–21653, 2024. 2

  14. [22]

    Decoupling appearance variations with 3d consistent features in gaussian splatting

    Jiaqi Lin, Zhihao Li, Binxiao Huang, Xiao Tang, Jianzhuang Liu, Shiyong Liu, Xiaofei Wu, Fenglong Song, and Wen- ming Yang. Decoupling appearance variations with 3d consistent features in gaussian splatting. Proceedings of the AAAI Conference on Artificial Intelligence, 39(5):...

  15. [23]

    Learning deep intensity field for extremely sparse-view cbct reconstruction

    Yiqun Lin, Zhongjin Luo, Wei Zhao, and Xiaomeng Li. Learning deep intensity field for extremely sparse-view cbct reconstruction. In International Conference on Medical Im- age Computing and Computer-Assisted Intervention , pages 13–23. Springer, 2023. 2, 3

  16. [24]

    Dolce: A model-based probabilistic diffusion framework for limited-angle ct reconstruction

    Jiaming Liu, Rushil Anirudh, Jayaraman J Thiagarajan, Stewart He, K Aditya Mohan, Ulugbek S Kamilov, and Hy- ojin Kim. Dolce: A model-based probabilistic diffusion framework for limited-angle ct reconstruction. In Proceed- ings of the IEEE/CVF International conference on compu...

  17. [25]

    Tomogan: low- dose synchrotron x-ray tomography with generative adver- sarial networks: discussion

    Zhengchun Liu, Tekin Bicer, Rajkumar Kettimuthu, Doga Gursoy, Francesco De Carlo, and Ian Foster. Tomogan: low- dose synchrotron x-ray tomography with generative adver- sarial networks: discussion. Journal of the Optical Society of America A, 37:422–434, 2020. 3

  18. [26]

    Scaffold-gs: Structured 3d gaussians for view-adaptive rendering

    Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654–20664, 2024. 2

  19. [27]

    Follow your pose: Pose- guided text-to-video generation using pose-free videos

    Yue Ma, Yingqing He, Xiaodong Cun, Xintao Wang, Siran Chen, Xiu Li, and Qifeng Chen. Follow your pose: Pose- guided text-to-video generation using pose-free videos. In Proceedings of the AAAI Conference on Artificial Intelli- gence, pages 4117–4125, 2024. 2

  20. [28]

    Follow-your-emoji: Fine-controllable and expressive freestyle portrait animation

    Yue Ma, Hongyu Liu, Hongfa Wang, Heng Pan, Yingqing He, Junkun Yuan, Ailing Zeng, Chengfei Cai, Heung-Yeung Shum, Wei Liu, et al. Follow-your-emoji: Fine-controllable and expressive freestyle portrait animation. In SIGGRAPH Asia 2024 Conference Papers, pages 1–12, 2024. 1

  21. [29]

    Follow-your-creation: Empowering 4d creation through video inpainting

    Yue Ma, Kunyu Feng, Xinhua Zhang, Hongyu Liu, David Junhao Zhang, Jinbo Xing, Yinhan Zhang, Ayden Yang, Zeyu Wang, and Qifeng Chen. Follow-your-creation: Empowering 4d creation through video inpainting. arXiv preprint arXiv:2506.04590, 2025. 2

  22. [30]

    Follow-your-click: Open-domain regional image animation via motion prompts

    Yue Ma, Yingqing He, Hongfa Wang, Andong Wang, Leqi Shen, Chenyang Qi, Jixuan Ying, Chengfei Cai, Zhifeng Li, Heung-Yeung Shum, et al. Follow-your-click: Open-domain regional image animation via motion prompts. In Proceed- 9 ings of the AAAI Conference on Artificial Intelligen...

  23. [31]

    Follow-your-motion: Video motion transfer via efficient spatial-temporal decoupled finetuning

    Yue Ma, Yulong Liu, Qiyuan Zhu, Ayden Yang, Kunyu Feng, Xinhua Zhang, Zhifeng Li, Sirui Han, Chenyang Qi, and Qifeng Chen. Follow-your-motion: Video motion transfer via efficient spatial-temporal decoupled finetuning. arXiv preprint arXiv:2506.05207, 2025. 2

  24. [32]

    Trans- mission maximum-likelihood reconstruction with ordered subsets for cone beam ct

    Stephen H Manglos, George M Gagne, Andrzej Krol, F Deaver Thomas, and Rammohan Narayanaswamy. Trans- mission maximum-likelihood reconstruction with ordered subsets for cone beam ct. Physics in Medicine & Biology , 40:1225, 1995. 3

  25. [33]

    Nerf: Representing scenes as neural radiance fields for view syn- thesis

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis. Communications of the ACM, 65(1):99–106, 2021. 2, 3

  26. [34]

    Pytorch: An imperative style, high-performance deep learning library

    A Paszke. Pytorch: An imperative style, high-performance deep learning library. arXiv preprint arXiv:1912.01703 ,

  27. [35]

    Hogsa: Bimanual hand-object interaction under- standing with 3d gaussian splatting based data augmenta- tion

    Wentian Qu, Jiahe Li, Jian Cheng, Jian Shi, Chenyu Meng, Cuixia Ma, Hongan Wang, Xiaoming Deng, and Yinda Zhang. Hogsa: Bimanual hand-object interaction under- standing with 3d gaussian splatting based data augmenta- tion. Proceedings of the AAAI Conference on Artificial In- t...

  28. [36]

    A local update strategy for iterative reconstruction from projections

    Ken Sauer and Charles Bouman. A local update strategy for iterative reconstruction from projections. IEEE Transactions on Signal Processing, 41(2):534–548, 2002. 3

  29. [37]

    Image reconstruction in circular cone-beam computed tomography by constrained, total-variation minimization

    Emil Y Sidky and Xiaochuan Pan. Image reconstruction in circular cone-beam computed tomography by constrained, total-variation minimization. Physics in Medicine & Biol- ogy, 53:4777, 2008. 3, 6

  30. [38]

    X-ray tomographic datasets, 2024

    The Finnish Inverse Problems Society. X-ray tomographic datasets, 2024. 2, 6

  31. [39]

    Image quality assessment: from error visibility to structural similarity

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Si- moncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600–612, 2004. 6

  32. [40]

    4d gaussian splatting for real-time dynamic scene rendering

    Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 20310–2032...

  33. [41]

    Mip-splatting: Alias-free 3d gaussian splat- ting

    Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splat- ting. In Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition , pages 19447–19456,

  34. [42]

    Warp-and-project to- mography for rapidly deforming objects

    Guangming Zang, Ramzi Idoughi, Ran Tao, Gilles Lubineau, Peter Wonka, and Wolfgang Heidrich. Warp-and-project to- mography for rapidly deforming objects. ACM Transactions on Graphics (TOG), 38:1–13, 2019. 3

  35. [43]

    Intratomo: self-supervised learning- based tomography via sinogram synthesis and prediction

    Guangming Zang, Ramzi Idoughi, Rui Li, Peter Wonka, and Wolfgang Heidrich. Intratomo: self-supervised learning- based tomography via sinogram synthesis and prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1960–1970, 2021. 2, 3

  36. [44]

    Naf: neural attenuation fields for sparse-view cbct reconstruction

    Ruyi Zha, Yanhao Zhang, and Hongdong Li. Naf: neural attenuation fields for sparse-view cbct reconstruction. In In- ternational Conference on Medical Image Computing and Computer-Assisted Intervention , pages 442–452. Springer,

  37. [45]

    R 2-gaussian: Rectifying radiative gaussian splatting for tomographic reconstruction

    Ruyi Zha, Tao Jun Lin, Yuanhao Cai, Jiwen Cao, Yanhao Zhang, and Hongdong Li. R 2-gaussian: Rectifying radiative gaussian splatting for tomographic reconstruction. In Ad- vances in Neural Information Processing Systems (NeurIPS),

  38. [46]

    Large images are gaussians: High-quality large image representation with lev- els of 2d gaussian splatting

    Lingting Zhu, Guying Lin, Jinnan Chen, Xinjie Zhang, Zhenchao Jin, Zhao Wang, and Lequan Yu. Large images are gaussians: High-quality large image representation with lev- els of 2d gaussian splatting. Proceedings of the AAAI Con- ference on Artificial Intelligence, 39:10977–10...

Pith tools

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