Active View Selection with Perturbed Gaussian Ensemble for Tomographic Reconstruction

TL;DR

Perturbed Gaussian Ensemble is presented, an active view selection framework that integrates uncertainty modeling with sequential decision-making, tailored for X-ray Gaussian Splatting, and effectively eliminates geometric artifacts and consistently outperforms existing baselines in progressive tomographic reconstruction under unified view selection protocols.

Abstract

Sparse-view computed tomography (CT) is critical for reducing radiation exposure to patients. Recent advances in radiative 3D Gaussian Splatting (3DGS) have enabled fast and accurate sparse-view CT reconstruction. Despite these algorithmic advancements, practical reconstruction fidelity remains fundamentally bounded by the quality of the captured data, raising the crucial yet underexplored problem of X-ray active view selection. Existing active view selection methods are primarily designed for natural-light scenes and fail to capture the unique geometric ambiguities and physical attenuation properties inherent in X-ray imaging. In this paper, we present Perturbed Gaussian Ensemble, an active view selection framework that integrates uncertainty modeling with sequential decision-making, tailored for X-ray Gaussian Splatting. Specifically, we identify low-density Gaussian primitives that are likely to be uncertain and apply stochastic density scaling to construct an ensemble of plausible Gaussian density fields. For each candidate projection, we measure the structural variance of the ensemble predictions and select the one with the highest variance as the next best view. Extensive experimental results on arbitrary-trajectory CT benchmarks demonstrate that our density-guided perturbation strategy effectively eliminates geometric artifacts and consistently outperforms existing baselines in progressive tomographic reconstruction under unified view selection protocols.

Figures (4)

• Figure 1: (a) We compare our approach against the state-of-the-art 3DGS-based active view selection method, FisherRF fisherrf, evaluating both quantitative metrics (3D PSNR$\uparrow$ and SSIM$\uparrow$) and qualitative visual fidelity, including zoomed-in details. Our method achieves the highest reconstruction quality and best preserves fine structural details. (b) We plot the training iterations vs. the average 3D PSNR$\uparrow$ (dB) and SSIM$\uparrow$ on the synthetic dataset. As demonstrated, our proposed approach consistently delivers superior reconstruction fidelity throughout the progressive reconstruction process.
• Figure 2: Comparison of active view selection paradigms for X-ray Gaussian Splatting. (a) Gradient-based method fisherrf estimates the expected information gain (EIG) of candidate views by computing a Fisher Information Matrix (FIM) via backpropagation, but suffers from gradient coupling and the absence of view-dependent parameters. (b) Ensemble-based method quantifies epistemic uncertainty by rendering disagreement across multiple Gaussian representations initialized with different random seeds, resulting in prohibitively high computational burden. (c) Our Perturbed Gaussian Ensemble introduces density-guided parameter perturbations to efficiently construct an ensemble, wherein uncertain primitives exhibit pronounced behavioral randomness. By calculating the structural variance of the rendered projections, our approach achieves superior uncertainty modeling and enables more effective active view selection.
• Figure 3: Visual comparisons of the reconstructed 3D volumes using different view selection strategies. The 3D PSNR$\uparrow$ (dB) for each scene is displayed at the top-left corner of the corresponding image. Our approach consistently achieves the highest reconstruction quality and better preserves fine structural details.
• Figure 4: Visual comparisons of novel view synthesis results across different view selection strategies. The PSNR$\uparrow$ (dB) for each scene is displayed at the top-left corner of each image. Our approach achieves the highest rendering quality.