GR-Gaussian: Graph CT Reconstruction

• GR-Gaussian introduces a graph-based CT reconstruction framework that models volumes as radiative Gaussians optimized from sparse X-ray projections.
• The method employs a denoised point cloud initialization and a Pixel-Graph-Aware gradient strategy to mitigate needle-like artifacts and enforce local density consistency.
• Empirical results show that GR-Gaussian improves PSNR and SSIM compared to prior methods while effectively balancing reconstruction quality and computational cost.

Searching arXiv for the primary paper and closely related Gaussian-representation reconstruction work. GR-Gaussian is a graph-based 3D Gaussian Splatting framework for sparse-view CT reconstruction that models a CT volume as a set of radiative Gaussians connected by a spatial graph, and optimizes that representation directly from sparse X-ray projections. In this formulation, the object is represented by Gaussian primitives with centers, densities, and covariances, while reconstruction quality is improved through two specific mechanisms: a Denoised Point Cloud Initialization Strategy and a Pixel-Graph-Aware Gradient Strategy. The method is designed to address a failure mode of earlier radiative Gaussian methods under sparse views, namely needle-like artifacts caused by inaccurate initialization and gradient-driven densification that ignores inter-Gaussian spatial relationships (Yuluo et al., 4 Aug 2025).

1. Definition and representational framework

GR-Gaussian represents the attenuation field as a set of radiative Gaussians

$\mathbb{G}^3 = \{G_i^3\}_{i=1}^M,$

where each Gaussian is parameterized by a center $\mathbf{p}_i$, a central density $\rho_i$, and a covariance $\Sigma_i$. The individual primitive is written as

$$G_i^3(\mathbf{x} \mid \rho_i, \mathbf{p}_i, \Sigma_i) = \rho_i \cdot \exp\left(-\frac{1}{2} (\mathbf{x} - \mathbf{p}_i)^\top \Sigma_i^{-1} (\mathbf{x} - \mathbf{p}_i)\right),$$

and the full density field is

$$\sigma(\mathbf{x}) = \sum_{i=1}^M G_i^3(\mathbf{x} \mid \rho_i, \mathbf{p}_i, \Sigma_i).$$

The covariance is parameterized as

$$\Sigma_i = \mathbf{R}_i \mathbf{S}_i \mathbf{S}_i^\top \mathbf{R}_i^\top,$$

with $\mathbf{R}_i$ a rotation matrix and $\mathbf{S}_i$ a diagonal scale matrix (Yuluo et al., 4 Aug 2025).

The framework is explicitly radiative rather than photometric. Unlike RGB 3D Gaussian Splatting, the primitives encode attenuation density, and rendering approximates X-ray projection formation rather than visible-light appearance. A plausible implication is that GR-Gaussian belongs to the same family as radiative Gaussian CT reconstruction methods, but specializes that family to sparse-view settings by introducing graph-aware regularization and densification.

2. Motivation in sparse-view CT

Sparse-view CT reconstruction is ill-posed because only a small number of projections are available. The paper studies the case of 25 views over $0^\circ\!-\!360^\circ$, where analytic methods such as FDK produce strong streak artifacts and iterative methods can be computationally expensive or overly smoothing (Yuluo et al., 4 Aug 2025).

Within Gaussian-based tomographic reconstruction, the paper identifies two limitations of prior radiative 3D Gaussian Splatting approaches. The first is initialization from noisy FDK volumes, which introduces inaccurate Gaussian centers and densities. The second is densification based only on the average gradient magnitude of projected positions,

$\mathbf{p}_i$0

which can fail to split large kernels that are inconsistent with local 3D structure. Under sparse views, such kernels can persist as elongated blobs aligned with projection directions, leading to needle-like artifacts in the reconstructed volume (Yuluo et al., 4 Aug 2025).

GR-Gaussian addresses these issues by combining a denoised initialization procedure with graph-based density reasoning. This suggests that the method’s central contribution is not a new forward model, but a refinement of how Gaussian primitives are initialized, connected, regularized, and selected for splitting.

3. Gaussian graph construction

The method defines a graph-based Gaussian system

$\mathbf{p}_i$1

with vertices

$\mathbf{p}_i$2

Edges are constructed by symmetric $\mathbf{p}_i$3-nearest neighbors: $\mathbf{p}_i$4

$\mathbf{p}_i$5

$\mathbf{p}_i$6

In the reported configuration, $\mathbf{p}_i$7 (Yuluo et al., 4 Aug 2025).

The graph is not ornamental. It is used in two places: first, as a regularizer through a graph Laplacian term on Gaussian densities; second, as an input to the Pixel-Graph-Aware Gradient Strategy that modifies the splitting criterion. The edge weights are

$\mathbf{p}_i$8

and the graph Laplacian regularization is

$\mathbf{p}_i$9

This term enforces local density smoothness in the Gaussian graph (Yuluo et al., 4 Aug 2025).

A common misconception would be to interpret GR-Gaussian as merely a graph neural network on Gaussians. The reported method does not introduce learned message passing over node features; instead, the graph serves as a geometric prior over density consistency and gradient augmentation.

4. Denoised point cloud initialization

Initialization proceeds from a sparse-view FDK reconstruction, but the FDK volume is first smoothed by a 3D Gaussian filter: $\rho_i$0 with

$\rho_i$1

The paper uses $\rho_i$2 (Yuluo et al., 4 Aug 2025).

After filtering, voxels below a density threshold are discarded, and $\rho_i$3 voxels are sampled as initial Gaussian centers. Initial scales are set from nearest-neighbor distances, and each density $\rho_i$4 is initialized from the denoised volume value at the sampled point (Yuluo et al., 4 Aug 2025).

This initialization strategy reduces noise-induced placement errors. In the ablation study, adding De-Init alone improves X-3D performance from PSNR 35.20 and SSIM 0.923 to PSNR 35.61 and SSIM 0.931, and improves real-world performance from PSNR 35.28 and SSIM 0.836 to PSNR 35.90 and SSIM 0.857 (Yuluo et al., 4 Aug 2025).

The role of De-Init is therefore not simply faster convergence. It changes the geometry of the starting Gaussian set, which in turn affects later graph construction and densification. A plausible implication is that graph-aware refinement depends on the initial neighborhood structure being physically meaningful.

5. Pixel-Graph-Aware Gradient Strategy

The method starts from the CT-adapted pixel-aware gradient

$\rho_i$5

where $\rho_i$6 is the number of influenced pixels and $\rho_i$7 is the Gaussian contribution in view $\rho_i$8 (Yuluo et al., 4 Aug 2025).

GR-Gaussian augments this with graph-based density discrepancies. For each neighbor $\rho_i$9, it defines

$\Sigma_i$0

The Pixel-Graph-Aware gradient becomes

$\Sigma_i$1

with $\Sigma_i$2 in the reported experiments (Yuluo et al., 4 Aug 2025).

The densification criterion is then applied to the graph-aware gradient magnitude: $\Sigma_i$3 This favors splitting Gaussians whose densities are inconsistent with nearby neighbors, even if their projection-based gradient is small. The paper states that

$\Sigma_i$4

so the graph term acts as a structural correction to purely pixel-driven splitting (Yuluo et al., 4 Aug 2025).

This addresses a specific pathology of sparse-view Gaussian CT reconstruction: large kernels at density boundaries may not receive sufficient pixel-only gradient to split, yet their mismatch with neighboring densities reveals that they should be refined. The graph term operationalizes that mismatch.

6. Objective function and optimization

The training loss combines image-space fidelity with volumetric and graph regularization. The regularization term is

$\Sigma_i$5

where $\Sigma_i$6 is 3D total variation on a voxelized density grid (Yuluo et al., 4 Aug 2025).

The total loss is

$\Sigma_i$7

The reported hyperparameters are

$\Sigma_i$8

Optimization uses Adam with initial learning rates:

• position $\Sigma_i$9: $$G_i^3(\mathbf{x} \mid \rho_i, \mathbf{p}_i, \Sigma_i) = \rho_i \cdot \exp\left(-\frac{1}{2} (\mathbf{x} - \mathbf{p}_i)^\top \Sigma_i^{-1} (\mathbf{x} - \mathbf{p}_i)\right),$$0,
• density $$G_i^3(\mathbf{x} \mid \rho_i, \mathbf{p}_i, \Sigma_i) = \rho_i \cdot \exp\left(-\frac{1}{2} (\mathbf{x} - \mathbf{p}_i)^\top \Sigma_i^{-1} (\mathbf{x} - \mathbf{p}_i)\right),$$1: $$G_i^3(\mathbf{x} \mid \rho_i, \mathbf{p}_i, \Sigma_i) = \rho_i \cdot \exp\left(-\frac{1}{2} (\mathbf{x} - \mathbf{p}_i)^\top \Sigma_i^{-1} (\mathbf{x} - \mathbf{p}_i)\right),$$2,
• scale: $$G_i^3(\mathbf{x} \mid \rho_i, \mathbf{p}_i, \Sigma_i) = \rho_i \cdot \exp\left(-\frac{1}{2} (\mathbf{x} - \mathbf{p}_i)^\top \Sigma_i^{-1} (\mathbf{x} - \mathbf{p}_i)\right),$$3,
• rotation: $$G_i^3(\mathbf{x} \mid \rho_i, \mathbf{p}_i, \Sigma_i) = \rho_i \cdot \exp\left(-\frac{1}{2} (\mathbf{x} - \mathbf{p}_i)^\top \Sigma_i^{-1} (\mathbf{x} - \mathbf{p}_i)\right),$$4,

all decayed exponentially to 10% of their initial values (Yuluo et al., 4 Aug 2025).

The method also employs dynamic stopping: PSNR is checked every 500 iterations, and optimization stops if PSNR drops by more than 0.5%. The paper reports that this avoids overfitting sparse projections, especially on real data (Yuluo et al., 4 Aug 2025).

7. Empirical results and ablations

On the X-3D dataset under 25-view reconstruction, GR-Gaussian reports average performance of PSNR 35.86 and SSIM 0.933, compared with R$$G_i^3(\mathbf{x} \mid \rho_i, \mathbf{p}_i, \Sigma_i) = \rho_i \cdot \exp\left(-\frac{1}{2} (\mathbf{x} - \mathbf{p}_i)^\top \Sigma_i^{-1} (\mathbf{x} - \mathbf{p}_i)\right),$$5-Gaussian at PSNR 35.19 and SSIM 0.922. On the real-world dataset, GR-Gaussian reports PSNR 35.95 and SSIM 0.858, compared with R$$G_i^3(\mathbf{x} \mid \rho_i, \mathbf{p}_i, \Sigma_i) = \rho_i \cdot \exp\left(-\frac{1}{2} (\mathbf{x} - \mathbf{p}_i)^\top \Sigma_i^{-1} (\mathbf{x} - \mathbf{p}_i)\right),$$6-Gaussian at PSNR 35.03 and SSIM 0.837 (Yuluo et al., 4 Aug 2025).

The abstract emphasizes gains of 0.67 dB and 0.92 dB in PSNR, and 0.011 and 0.021 in SSIM, respectively, on X-3D and real-world data (Yuluo et al., 4 Aug 2025).

The ablation study separates the two principal innovations:

Variant	X-3D PSNR / SSIM	Real PSNR / SSIM
Baseline	35.20 / 0.923	35.28 / 0.836
+ De-Init	35.61 / 0.931	35.90 / 0.857
+ PGA	35.77 / 0.932	35.62 / 0.848
Full method	35.86 / 0.932	35.95 / 0.858

These results indicate that De-Init contributes strongly on real noisy scans, whereas PGA contributes consistently across both synthetic and real settings (Yuluo et al., 4 Aug 2025).

The parameter study further reports that $$G_i^3(\mathbf{x} \mid \rho_i, \mathbf{p}_i, \Sigma_i) = \rho_i \cdot \exp\left(-\frac{1}{2} (\mathbf{x} - \mathbf{p}_i)^\top \Sigma_i^{-1} (\mathbf{x} - \mathbf{p}_i)\right),$$7 offers a quality-time trade-off superior to nearby choices, and $$G_i^3(\mathbf{x} \mid \rho_i, \mathbf{p}_i, \Sigma_i) = \rho_i \cdot \exp\left(-\frac{1}{2} (\mathbf{x} - \mathbf{p}_i)^\top \Sigma_i^{-1} (\mathbf{x} - \mathbf{p}_i)\right),$$8 outperforms weaker or stronger Gaussian smoothing, with larger $$G_i^3(\mathbf{x} \mid \rho_i, \mathbf{p}_i, \Sigma_i) = \rho_i \cdot \exp\left(-\frac{1}{2} (\mathbf{x} - \mathbf{p}_i)^\top \Sigma_i^{-1} (\mathbf{x} - \mathbf{p}_i)\right),$$9 causing oversmoothing (Yuluo et al., 4 Aug 2025).

8. Relation to neighboring Gaussian-reconstruction work

GR-Gaussian is situated within a broader line of Gaussian-based inverse problems, but it is technically distinct. It extends radiative 3D Gaussian Splatting for CT, whereas GR-Diffusion applies a 3D discrete Gaussian representation to PET reconstruction and then uses that reconstruction as a structural prior for diffusion-based refinement (Geng et al., 12 Feb 2026). The PET work explicitly states that it builds on “radiative/GR Gaussians for tomographic reconstruction” and adopts related “discretize and align” ideas (Geng et al., 12 Feb 2026). This suggests that GR-Gaussian is part of an emerging class of Gaussian-primitive reconstruction methods in medical imaging, though the sparse-view CT formulation remains specific to graph-based densification and radiative splatting.

A further misconception would be to conflate GR-Gaussian with general graph-based 3D Gaussian representations for view synthesis or scene modeling. In the present usage, the term denotes a CT reconstruction framework whose “graph-based” aspect refers to local density relationships among radiative Gaussian primitives, not a generic graph neural architecture.

9. Significance, limitations, and interpretation

GR-Gaussian’s significance lies in showing that sparse-view CT artifacts in Gaussian splatting are not solely a rendering or forward-model issue. They are also a representation-management issue: how primitives are initialized, how neighborhoods are defined, and how splitting decisions are made. By adding graph-aware density reasoning, the method changes when and where model capacity is allocated (Yuluo et al., 4 Aug 2025).

The method’s stated limitations are local in character. It relies on FDK being structurally informative enough to support denoised initialization, and its graph priors remain local KNN smoothness constraints rather than learned or nonlocal priors. The paper notes that larger $$\sigma(\mathbf{x}) = \sum_{i=1}^M G_i^3(\mathbf{x} \mid \rho_i, \mathbf{p}_i, \Sigma_i).$$0 improves reconstruction at the cost of runtime, and suggests that more sophisticated graph priors or learned components are possible future directions (Yuluo et al., 4 Aug 2025).

In encyclopedic terms, GR-Gaussian can therefore be defined as a sparse-view CT reconstruction method that combines radiative Gaussian splatting with graph-structured density regularization and graph-aware densification. Its core novelty is the use of neighborhood density discrepancies to guide Gaussian splitting, thereby suppressing needle-like artifacts that arise when pixel-only gradients fail to refine structurally inconsistent primitives (Yuluo et al., 4 Aug 2025).