Volumetric Plane-Based Gaussian Selection for Internal View Synthesis With 3-D Gaussian Splatting

Abstract

3D Gaussian Splatting (3DGS) represents scenes with anisotropic Gaussian primitives optimized from multi-view images and rendered through differentiable rasterization and alpha compositing. Conventional 3DGS training primarily constrains camera-visible surfaces, whereas occluded interior regions remain insufficiently supervised. In medical imaging, internal view synthesis requires cross-sectional images from volumetric CT and MRI datasets at arbitrary plane orientations and spatial positions. We investigate whether Gaussian primitives can reconstruct internal volumetric structure under cross-sectional supervision. We propose GS-IR, a plane-based internal rendering framework that replaces exterior-view supervision with cross-sectional supervision sampled from the volume. Virtual cameras define slicing planes, and a binary mask selects the Gaussians whose centers lie within the plane thickness. The selected Gaussian subset is rendered with the original Gaussian rasterizer and optimized against ground-truth cross-sections generated through trilinear resampling. This mechanism concentrates the optimization gradients on interior Gaussians. We further evaluate distance-aware soft weighting, which continuously attenuates Gaussian opacity as Gaussian-to-plane distance increases. Experiments on KiTS23 CT and IXI MRI apply GS-IR to 3DGS, 2DGS, and Mip-Splatting without modifying their rasterizers. Experiments across the x-, y-, and z-axis configurations and spatially translated slicing planes evaluate generalization to unseen orientations and off-center positions. GS-IR improves PSNR by up to 21.52 dB and SSIM by up to 0.54; the largest gains occur for 3DGS on MRI and 2DGS on CT, respectively. These results demonstrate internal reconstruction across Gaussian representations. Our source code for GS-IR is publicly available at https://github.com/Choishroy/GS-IR