Calibrated Closed-Form Uncertainty for Radiative Gaussian Splatting in Sparse-View CT

Technical Summary: Calibrated Closed-Form Uncertainty for Radiative Gaussian Splatting in Sparse-View CT

Problem Statement
Sparse-view computed tomography (CT) reduces patient radiation dose by acquiring fewer X-ray projections, resulting in an ill-posed inverse problem where large regions of the reconstruction are interpolated rather than measured. While Radiative 3D Gaussian Splatting (specifically R2-Gaussian) has enabled fast and accurate reconstruction in this regime, existing methods output only single point estimates. They lack a mechanism to indicate where the reconstruction is trustworthy versus where it is a confident guess. Current uncertainty quantification efforts in this space are limited: concurrent work (Wu et al.) uses ensemble disagreement purely as an acquisition score without validating its correspondence to reconstruction error or producing per-voxel uncertainty maps. Furthermore, uncertainty literature in Gaussian splatting is largely confined to RGB rendering, where non-linear alpha compositing forces approximate variance propagation, whereas X-ray imaging involves strictly linear transmissive physics that remains unexploited for exact uncertainty propagation.

Methodology
The authors propose a framework that leverages the strict linearity of X-ray rendering to derive exact, closed-form uncertainty propagation for radiative Gaussian splatting.

1. Variational Density Posterior: The method lifts the central density of each Gaussian, $\rho_i$, from a point estimate to a Gaussian posterior distribution $\rho_i \sim \mathcal{N}(m_i, s_i^2)$, while keeping positions and covariances as point estimates. This "partial Bayes" approach allows the uncertainty parameter $s_i$ to absorb geometric ambiguity.
2. Training Objective: The model is trained using a reparameterized sampling strategy where densities are sampled as $\tilde{\rho}_i = m_i + \epsilon_i s_i$ (clamped to be non-negative). The loss function includes a photometric term (L1 + SSIM), total variation regularization, and a Kullback-Leibler (KL) prior. The KL prior is critical: for Gaussians unconstrained by any ray (the null space of sparse-view tomography), the data term provides no gradient, so the prior anchors the uncertainty $s_i$ to a baseline value, ensuring the model is "uncertain where unobserved."
3. Closed-Form Propagation: Due to the linearity of the projection and voxelization operations (unlike the non-linear alpha compositing in RGB), the predictive variance of the reconstruction is the variance of a weighted sum of independent variables. This yields exact closed-form expressions for both volume space ($\sigma^2(x) = \sum g_i(x)^2 s_i^2$) and projection space ($\text{Var}[I_p] = \sum w_{i,p}^2 s_i^2$).
4. Implementation Efficiency: The authors demonstrate that these variance maps can be computed in a single forward pass using the existing rasterizer and voxelizer without CUDA modifications. By exploiting the identity that squaring a Gaussian kernel halves its covariance (scaling $S_i \to S_i/\sqrt{2}$), the system computes the exact variance. This "B2" (closed-form) approach is 2.5–10.4× faster than a Monte Carlo (B1) reference, though B1 is used for calibration evaluation as it better approximates the truncated distribution used during training.
5. Calibration Protocol: The paper introduces a rigorous calibration protocol using three metrics: Spearman rank correlation (spatial ordering), AUSE (Area Under the Sparsification Curve), and temperature-scaled Expected Calibration Error (ECE). A split-holdout strategy is used to fit the temperature scalar $\tau$ on a subset of voxels to prevent overfitting.

Key Contributions

1. Exact Closed-Form Uncertainty: The derivation of exact predictive variance for radiative Gaussians in both volume and projection space, computable in a single pass. This is exact due to the linear nature of X-ray rendering, contrasting with the approximate methods required for RGB splatting.
2. Systematic Calibration Study: The first systematic evaluation of Gaussian-splatting CT uncertainty against ground-truth error across 15 scenes and three view budgets. The study benchmarks the variational posterior against deep ensembles and a perturbation-based heuristic transplanted to voxel space.
3. Mechanism Analysis of View Selection: A controlled dissection revealing that uncertainty scores earn a measurable premium over free geometric rules only in specific regimes (anisotropic content). In degenerate coverage regimes, the heuristic functions merely as a coverage detector (replaceable by geometric rules), and in isotropic regimes, view selection is a low-stakes decision where any candidate is nearly as good as the best.
4. Coverage-Gated, Calibrated Acquisition (CGCA): A proposed acquisition policy that combines a model-free geometric coverage gate (to prevent degenerate clustering) with a maturity schedule (scoring only on converged models) and the closed-form variance score. This policy matches the acquisition gains of concurrent heuristics at roughly one-tenth the scoring cost.

Results

• Calibration Performance: On the official 15-scene benchmark, the variational posterior ranks true reconstruction error (Spearman > 0.6) in 14 of 15 scenes across three view budgets. Nine of these scenes also meet the magnitude-calibration target (temperature-scaled ECE < 0.1) after a single temperature scaling.
• Comparison to Heuristics: The concurrent perturbation-ensemble heuristic, when transplanted to voxel space, fails to correlate with error in several scenes (rank correlation as low as -0.08 on the "head" scene), despite having acceptable ECE scores. This highlights that magnitude calibration alone is insufficient; rank correlation is essential.
• Efficiency: The closed-form (B2) variance is highly correlated with the Monte Carlo (B1) reference (Pearson 0.80–0.88 in volume, 0.94–0.95 in projection) but offers significant speedups.
• Acquisition Policy: In the CGCA evaluation, the geometric coverage gate alone prevents catastrophic failure (degenerate clustering). In anisotropic regimes (abdomen), adding the uncertainty score improves reconstruction quality, whereas in isotropic regimes (chest), the score provides negligible benefit and may even slightly degrade performance due to the cost of scoring immature models.

Significance and Claims
The paper claims that calibrated, closed-form, per-voxel uncertainty is a first-class deliverable for Gaussian-splatting CT, distinct from and superior to using uncertainty merely as a proxy for view selection. The authors argue that acquisition is a "structurally forgiving" consumer of uncertainty signals, often masking poor calibration, whereas clinical or downstream decision-making requires reliable per-voxel error localization.

The work establishes that:

• Uncertainty quantification and view selection are separable products; a score useful for acquisition is not necessarily a good uncertainty map.
• The strict linearity of X-ray physics allows for exact uncertainty propagation, a property not available in RGB rendering.
• A principled acquisition policy (CGCA) should rely on geometric coverage to prevent failure modes and use calibrated uncertainty scores only where they provide a measurable premium (anisotropic content).
• The calibrated posterior points toward a dose-adaptive stopping rule (stopping when uncertainty falls below a threshold), though this is proposed as a design implication for future work rather than experimentally validated in this paper.

The authors remain modest regarding generalization, noting that all evaluations use synthetic benchmarks with exact ground truth, and transfer to real clinical data remains untested. They also acknowledge that the current factorized mean-field posterior may miss fine-grained correlations in balanced isotropic states, suggesting correlation-aware extensions as a future direction.