Orthographic Gaussian Splatting from Axially Stacked Radiographs via SfM-Guided Novel ViewSynthesis

Abstract

We present a Gaussian-splatting pipeline tailored to axially stacked radiographs that breaks with the perspective and alpha-blending assumptions of standard 3DGS. Clini cal CT data are orthographic and exhibit little cross-slice feature continuity, causing classical SfM and vanilla splat renderers to fail (edge-biased splats, central collapse). Our method first converts ordered CT slices into a metric volume, then renders **ra diographic** projections that obey Beer–Lambert attenuation to synthesize multi-view images with controlled overlap. From these, we obtain poses and train a **rectified radiative Gaussian** model that replaces alpha compositing with additive X-ray ac cumulation and includes a density-rectification term so each Gaussian’s parameter encodes true 3D density rather than view-integrated mass. We initialize Gaussians directly from the slice volume with a lightweight sampler that seeds positions, scales, and densities, and stabilize optimization with gentle TV-on-voxel readouts and adap tive clone/split densification. On a 1.8k-slice brain CT (HiP-CT family), using 1.2k synthetic projections, our system reconstructs coherent anatomy and produces faithful novel projections after 45 minutes of training, visibly restoring central structures and suppressing edge-only artifacts compared to a cinematic-GS baseline. The approach is code-practical (no reliance on FDK/TIGRE), data-efficient, and compatible with DICOM geometry, offering a reproducible path to fast, radiography-consistent 3D reconstructions from orthographic medical stacks. We discuss remaining limits (pose accuracy, scatter, regularization strength) and outline ablations and metrics (L1/SSIM on held-out views, orthogonal-slice fidelity) to guide future clinical validation.