Revisiting Pose Sensitivity in Splat-based Computed Tomography under Sparse-view Reconstruction

Recent advances in splat-based tomography—which represent 3D volumes as distributions of anisotropic 3D Gaussians—have shown great promise for fast, high-quality CT reconstruction. However, when applied to real-world sparse-view data, these models often suffer from severe streak and strip artifacts. This paper investigates whether these artifacts are an inherent limitation of sparse-view reconstruction or if they stem from other factors in the acquisition pipeline.

To isolate the cause of the artifacts, the authors performed a controlled experiment comparing real-world sparse-view CT data against synthetic data generated from a pseudo-ground-truth volume. By analyzing the error distribution in the projection space, they identified that real-world artifacts exhibit a directional bias around edges, which is characteristic of geometric misalignment.

Building on this, the authors reformulated the splatting projection model to include camera pose parameters (rotation and translation) as learnable variables. They derived a stable, differentiable gradient-based framework that allows for the joint optimization of the volumetric Gaussian parameters and the camera geometry. Unlike previous methods that rely on external calibration or heuristic smoothness constraints, this approach incorporates the Jacobian of the projection function directly into the backward pass, enabling the system to self-calibrate during the reconstruction process.

This work provides a critical insight into the robustness of modern differentiable rendering techniques in medical and industrial imaging. By demonstrating that geometric calibration is the primary bottleneck for splat-based CT, the authors provide a lightweight, integrable solution that significantly improves reconstruction fidelity. This approach removes the need for expensive, time-consuming offline calibration phantoms, making high-quality sparse-view CT more accessible for real-world deployment.