XClipGS: Exact Half-Space Clipping for Medical Volume Gaussian Splatting

arXiv cs.GR details XClipGS, a new approach to medical volume Gaussian splatting that treats clipping as a first-class render-time operation rather than a crude keep-or-drop decision. The method targets the common case where a clipping plane exposes anatomy that was never directly visible during external-view training, which is exactly where conventional splatting tends to break down.

The core idea is a closed-form half-space clip operator under the local affine model used by EWA splatting. Instead of adding learned clipping parameters or an auxiliary network, XClipGS factors the ray integral into the usual 2D footprint and a conditional Gaussian CDF, keeping the operator differentiable with respect to both the primitive and the plane. That makes it practical for training pipelines as well as interactive viewers.

The paper also introduces a paired clipped/unclipped cut-face protocol with difference-referenced cut error and culled-side leakage, which is a useful reminder that global image metrics can hide the exact failures developers care about near the plane. On eight CT and MRI volumes, XClipGS reports the best PSNR on every volume, 33.56 dB versus 32.34 dB for ClipGS, while rendering at over 650 FPS. It also improves cut-face quality and cuts leakage dramatically on arbitrary planes without retraining.

For graphics and tools teams, the broader takeaway is that exact geometric handling can matter more than another learned patch when the user is actively slicing through dense volumetric data. The technique should be especially relevant to medical visualization, but the...