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

Area of Science:

• Computer-assisted surgery
• Medical imaging
• 3D reconstruction

Background:

• Accurate 3D reconstruction is vital for computer-assisted minimally invasive surgery.
• Dynamic surgical scenes with instrument occlusions present significant reconstruction challenges.
• Existing 3D Gaussian Splatting (3DGS) methods struggle with incomplete surfaces and error propagation in occluded areas.

Purpose of the Study:

• To enhance 3D reconstruction accuracy in dynamically occluded surgical environments.
• To address limitations of current 3DGS approaches in handling occlusions and motion errors.

Main Methods:

• Proposed a diffusion-guided Gaussian Splatting (DiGS) framework.
• Developed a diffusion-guided surface completion network using surgical scene priors for occluded regions.
• Implemented a lightweight annealed smoothing mechanism to correct endoscope motion estimation errors and stabilize optimization.

Main Results:

• DiGS demonstrated superiority over state-of-the-art methods on EndoNeRF and StereoMIS datasets.
• Achieved a 61.75% LPIPS improvement on EndoNeRF for better perceptual alignment in occluded scenes.
• On StereoMIS, obtained a 7.03% PSNR gain, 40.79% LPIPS improvement, and higher SSIM scores, indicating superior structural detail preservation.

Conclusions:

• The DiGS framework effectively improves 3D model accuracy and temporal coherence in challenging surgical scenes.
• Successfully addresses dynamic occlusions and motion-induced errors in surgical scene reconstruction.
• The DiGS code is publicly available for further research and development.