R2-Gaussian Framework in Tomography

• R2-Gaussian framework is an advanced method that uses radiative Gaussian splats for bias-corrected, differentiable 3D volumetric reconstruction from sparse projections.
• It rectifies integration biases inherent in naive 3D Gaussian splatting via kernel refactoring and analytic marginalization.
• The framework achieves state-of-the-art accuracy and speed with GPU-accelerated voxelization, improving performance in sparse-view CT and PET.

The term "R²-Gaussian framework" encompasses advanced methodologies in tomography and computer vision relying on radiative Gaussian splats for differentiable 3D volume reconstruction from sparse and noisy projections. R²-Gaussian designates a family of forward and inverse models that rectify the analytic integration bias endemic to naive 3D Gaussian splatting (3DGS) via precise kernel refactoring and bias correction, enabling highly efficient, unbiased, and GPU-accelerated tomography pipelines. The approach has shown state-of-the-art accuracy and computing speed in sparse-view CT and related problems, and underlies a new standard in neural-field-based volumetric reconstruction.

1. Mathematical Foundations of 3D Gaussian Splatting in Tomography

R²-Gaussian models the volumetric density field as a finite superposition of $M$ local, anisotropic 3D Gaussian kernels: $$\sigma(\mathbf{x}) = \sum_{i=1}^M G^3_i(\mathbf{x}) = \sum_{i=1}^M \rho_i\, \exp\left(-\frac{1}{2} (\mathbf{x} - \mathbf{p}_i)^\top \Sigma_i^{-1} (\mathbf{x} - \mathbf{p}_i)\right),$$ where each $G^3_i$ is parameterized by mean $\mathbf{p}_i\in\mathbb{R}^3$, covariance $\Sigma_i\in\mathbb{R}^{3\times3}$, and amplitude $\rho_i\in\mathbb{R}$. For a detector ray $\mathbf{r}(t) = \mathbf{o} + t\mathbf{d}$, the forward operator (via the Beer–Lambert law) computes the projected line integral: $$I(\mathbf{r}) = \int_{t_n}^{t_f} \sigma(\mathbf{r}(t))\,dt = \sum_{i=1}^M \int_{t_n}^{t_f} \rho_i\,\exp\left(-\frac{1}{2}(\mathbf{r}(t) - \mathbf{p}_i)^\top \Sigma_i^{-1}(\mathbf{r}(t) - \mathbf{p}_i)\right)dt.$$ This model enables a fully differentiable mapping from density parameters to observed measurements, critical for gradient-based inversion (Zha et al., 2024, Chen et al., 1 Feb 2026).

2. Integration Bias in Standard 3DGS and Analytic Correction

The original 3DGS rendering paradigm, when adapted for tomography, introduces a geometric bias by simplifying the analytic projection of 3D Gaussians to 2D detector elements. Specifically, plain 3DGS collapses to 2D with effective image-space amplitude $\hat{\rho}_i$, neglecting the Jacobian determinant arising from marginalizing the Gaussian along the integration direction: $$I_{\text{3DGS}}(u) = \sum_i \hat{\rho}_i \exp\left(-\frac{1}{2}(u-\bar{p}_i)^\top(\Sigma_i^{2\times2})^{-1}(u - \bar{p}_i)\right),$$ with $$\sigma(\mathbf{x}) = \sum_{i=1}^M G^3_i(\mathbf{x}) = \sum_{i=1}^M \rho_i\, \exp\left(-\frac{1}{2} (\mathbf{x} - \mathbf{p}_i)^\top \Sigma_i^{-1} (\mathbf{x} - \mathbf{p}_i)\right),$$0, thereby omitting the analytic factor

$$\sigma(\mathbf{x}) = \sum_{i=1}^M G^3_i(\mathbf{x}) = \sum_{i=1}^M \rho_i\, \exp\left(-\frac{1}{2} (\mathbf{x} - \mathbf{p}_i)^\top \Sigma_i^{-1} (\mathbf{x} - \mathbf{p}_i)\right),$$1

where $$\sigma(\mathbf{x}) = \sum_{i=1}^M G^3_i(\mathbf{x}) = \sum_{i=1}^M \rho_i\, \exp\left(-\frac{1}{2} (\mathbf{x} - \mathbf{p}_i)^\top \Sigma_i^{-1} (\mathbf{x} - \mathbf{p}_i)\right),$$2 is the marginal over the 2D detector plane. This bias leads to angularly and positionally dependent amplitude errors, resulting in inconsistent reconstructed densities and an inability to recover the true volume when the projection geometry is varied (Zha et al., 2024, Chen et al., 1 Feb 2026).

R²-Gaussian rectifies this by including the full analytic correction term through explicit analytic marginalization: $$\sigma(\mathbf{x}) = \sum_{i=1}^M G^3_i(\mathbf{x}) = \sum_{i=1}^M \rho_i\, \exp\left(-\frac{1}{2} (\mathbf{x} - \mathbf{p}_i)^\top \Sigma_i^{-1} (\mathbf{x} - \mathbf{p}_i)\right),$$3 This adjustment ensures unbiased and rotation-consistent density recovery, which is particularly critical in quantitative CT and PET (Zha et al., 2024).

3. Radiative Kernel Design and GPU Voxelization

R²-Gaussian extends the isotropic splats of classical 3DGS to general, full-covariance "radiative" Gaussians, parameterized as: $$\sigma(\mathbf{x}) = \sum_{i=1}^M G^3_i(\mathbf{x}) = \sum_{i=1}^M \rho_i\, \exp\left(-\frac{1}{2} (\mathbf{x} - \mathbf{p}_i)^\top \Sigma_i^{-1} (\mathbf{x} - \mathbf{p}_i)\right),$$4 with orientation matrix $$\sigma(\mathbf{x}) = \sum_{i=1}^M G^3_i(\mathbf{x}) = \sum_{i=1}^M \rho_i\, \exp\left(-\frac{1}{2} (\mathbf{x} - \mathbf{p}_i)^\top \Sigma_i^{-1} (\mathbf{x} - \mathbf{p}_i)\right),$$5 and diagonal scale matrix $$\sigma(\mathbf{x}) = \sum_{i=1}^M G^3_i(\mathbf{x}) = \sum_{i=1}^M \rho_i\, \exp\left(-\frac{1}{2} (\mathbf{x} - \mathbf{p}_i)^\top \Sigma_i^{-1} (\mathbf{x} - \mathbf{p}_i)\right),$$6. This construction allows control over blob elongation, flattening, and orientation, enabling the representation of complex anatomical structures or imaging artifacts (Zha et al., 2024).

Efficient forward and backward passes are implemented through a CUDA-based voxelizer that iterates over 3D tiles, accumulating per-voxel contributions from all active kernels with $$\sigma(\mathbf{x}) = \sum_{i=1}^M G^3_i(\mathbf{x}) = \sum_{i=1}^M \rho_i\, \exp\left(-\frac{1}{2} (\mathbf{x} - \mathbf{p}_i)^\top \Sigma_i^{-1} (\mathbf{x} - \mathbf{p}_i)\right),$$7 support in each tile. The framework supports parallel gradient computations with fused CUDA kernels, supporting high-throughput (over 100 FPS for volume queries), and ensures all parameter updates remain differentiable (Zha et al., 2024).

4. Reconstruction Pipeline, Performance, and Quantitative Results

The R²-Gaussian algorithm proceeds as follows:

• The volume domain is partitioned into a set of anisotropic Gaussian kernels, each tracked in structure-of-arrays (SoA) format for memory throughput.
• At each training iteration, the forward operator analytically projects splats, including the bias correction, to compute synthetic measurements.
• The discrepancy between predicted and actual projections drives gradient-based updates on all kernel parameters.
• Regularization and culling reduce spurious kernels and enforce positivity and physical plausibility.

On 15 CT datasets (organs, insects, artifacts) with 25–75 sparse views, R²-Gaussian outperforms the NeRF-based IntraTomo, NAF, and SAX-NeRF by +0.93 dB PSNR and +0.014 SSIM, and surpasses classical SART/ASD-POCS by +2–3 dB PSNR. The full pipeline converges in approximately 3 minutes, delivering peak quality within 8 minutes—an order of magnitude faster than NeRF baselines and competitive with analytic iterative methods (SART: 2–4 minutes; NeRF-derivatives: 30–780 minutes) (Zha et al., 2024). RMSE improves by more than 5% over prior art. In PET (NEMA phantom), R²-Gaussian achieves SBR error < 5% in multiple spheres with reduced background bias and approximate clinical runtimes (Chen et al., 1 Feb 2026).

5. Implementation Assumptions and Algorithmic Limitations

R²-Gaussian assumes that projection mappings are well-approximated by first-order (local-affine) expansions at each kernel mean and that support truncation or pruning suffices to avoid artifacts from Gaussian tails. No occlusion or non-linear attenuation is modeled, so the method targets linear, strictly additive tomographic forward models (standard in medical CT, X-ray, PET).

Primary limitations include residual bias for large, highly anisotropic, or nonlocal Gaussians, modest robustness to calibration errors (e.g., misalignments, beam-hardening), and some "needle-like" artifacts in extreme sparsity regimes. Nonlinear geometric corrections (cylindrical detector arcs in PET) are not directly supported within the splatting paradigm; these require either explicit ray tracing, resampling, or model extension (Chen et al., 1 Feb 2026, Zha et al., 2024).

6. Extensions, Future Directions, and Comparative Context

The R²-Gaussian framework is the first 3DGS-based model for truly bias-free tomographic reconstruction and forms the basis for differentiable, interpretable, and scalable neural-field inverse solvers. The analytic kernel projection and radiative kernel design distinguish it from neural radiance field (NeRF) models and classical iterative reconstruction (SART, ASD-POCS).

Current research targets joint estimation of calibration parameters, integration with learned priors from MRI, and hybrid kernel-based surface-volume extraction routines. Substantial practical gains are expected in sparse-view clinical CT, non-destructive testing, and preclinical/small-animal PET (Zha et al., 2024, Chen et al., 1 Feb 2026).

In summary, the R²-Gaussian framework provides a rigorously corrected, physically consistent, and computationally performant foundation for state-of-the-art tomographic volume reconstruction in modern computational imaging.