Physics-Informed Dynamic Gaussian Splatting
for Surgical Scene Reconstruction
Existing 4D Gaussian Splatting treats a surgical scene as one big bag of moving blobs — ignoring that it is governed by physics. PhysioSplat injects three physical priors into the splatting process, so reconstruction stops hallucinating and starts respecting reality: rigid tools are separated from soft tissue, deformations stay biologically plausible, and wet-surface highlights no longer explode into floaters.
Hover a pillar — each addresses a failure mode that purely-visual 4DGS cannot.
Splits Gaussians into a rigid tool field and a deformable tissue field. A learnable semantic logit routes each primitive to rigid SO(3) motion or a non-rigid hex-plane field — enabling tool removal and recovery of occluded anatomy.
An As-Rigid-As-Possible strain energy, a tool–tissue collision repulsion, and a volume-preservation term keep the deformation field on the manifold of biologically plausible tissue motion — no collapse, no interpenetration.
Endoscope light is co-located with the camera. PhysioSplat decomposes radiance into diffuse + view-dependent specular, with normals read from the rendered depth gradient — stabilizing highlights on wet, reflective tissue.
Same case, same viewpoint. Drag the handle — prior work shatters into floaters where ours stays a clean, continuous surface.
Tool removal & anatomy recovery. Each card pairs an input frame (top, with occluding instruments) and PhysioSplat's reconstructed 3D model (bottom) — instruments erased, the tissue behind them recovered.
input3D model
input3D model
input3D model
input3D model
The green pyramid is the moving camera frustum. Because tool and tissue are disentangled, the full geometry is recovered free of camera occlusion — and can be re-rendered from novel views.
PSNR ↑ on the four evaluation splits. Bars animate when in view.
Component ablation (PSNR, EndoNeRF-Cutting) — each physical model contributes.
# clone & install
git clone https://github.com/hritam-98/physiosplat.git
cd physiosplat && pip install -r requirements.txt && pip install -e .
# run the built-in synthetic demo — no data needed
python examples/run_synthetic.py
# train on your own EndoNeRF-style scene
python scripts/train.py --data /path/to/scene --iterations 30000 --device cuda@inproceedings{basak2026physiosplat,
title = {PhysioSplat: Physics-Informed Dynamic Gaussian Splatting
for Surgical Scene Reconstruction},
author = {Basak, Hritam and Yin, Zhaozheng},
booktitle = {Medical Image Computing and Computer Assisted Intervention (MICCAI)},
year = {2026}
}