Digs: diffusion-guided Gaussian Splatting for dynamic occlusion surgical scene reconstruction

Accurate 3D reconstruction from endoscopic videos is crucial for advancing computer-assisted minimally invasive surgery. However, existing approaches struggle with dynamic surgical scenes where instrument occlusions cause significant reconstruction artifacts. Although 3D Gaussian Splatting (3DGS) enables rapid reconstruction, it often suffers from incomplete surface recovery due to occlusion-induced missing regions and error propagation from suboptimal initial point clouds during radiance field optimization. This study aims to enhance reconstruction accuracy in dynamically occluded surgical environments.

We propose a diffusion-guided Gaussian Splatting (DiGS) framework comprising two key components: (1) a diffusion-guided surface completion network that incorporates surgical scene priors to restore high-fidelity textures in occluded regions, improving surface completeness; and (2) a lightweight annealed smoothing mechanism designed to mitigate endoscope motion estimation errors, ensuring temporal coherence during continuous frame interpolation and stabilizing radiance field optimization.

Extensive experiments on the EndoNeRF and StereoMIS datasets demonstrate the superiority of DiGS over state-of-the-art baselines. On EndoNeRF, DiGS achieves a 61.75% improvement in LPIPS, indicating stronger perceptual alignment in dynamically occluded scenes. On StereoMIS, DiGS delivers an 7.03% PSNR gain and a 40.79% LPIPS improvement, along with consistently higher SSIM scores confirming superior preservation of structural details.

The proposed DiGS framework effectively addresses the challenges of dynamic occlusions and motion-induced errors in surgical scene reconstruction, producing more accurate and temporally coherent 3D models. The code is publicly available at https://github.com/IGSResearch/DiGS .