Related Experiment Video
Updated: Jun 27, 2026
06:54
Photorealistic Learned Landscapes for Augmented Reality
Published on: June 27, 2025
EndoLRMGS: Combining Large Reconstruction Modelling and Gaussian Splatting for Complete Endoscopic Scene
IEEE Transactions on Medical Imaging
|June 25, 2026
Summary
This study introduces EndoLRMGS, a novel framework for 3D reconstruction of surgical scenes, accurately modeling both deformable tissues and articulated instruments from endoscopic videos.
Area of Science:
- Computer Vision
- Medical Robotics
- 3D Reconstruction
Background:
- Reconstructing dynamic surgical scenes from endoscopic videos is challenging.
- Existing methods often neglect articulated instruments, focusing only on deformable tissues.
Purpose of the Study:
- To present EndoLRMGS, the first unified framework for reconstructing both deformable tissue and articulated instruments from monocular video and depth priors.
- To improve the accuracy and completeness of 3D surgical scene reconstruction.
Main Methods:
- Frequency-Modulated Gaussian Splatting (FMGS) for high-fidelity deformable tissue reconstruction.
- Large Reconstruction Model (LRM) combined with Orthographic and Perspective joint Projection Optimization (OPjPO) for instrument reconstruction, scale recovery, and spatial alignment.
Main Results:
- EndoLRMGS achieves high PSNR and low Chamfer distance for tissue reconstruction (PSNR: 28.4981–38.4179, CD: 1.43–4.71 mm).
- Instrument reconstruction shows competitive PSNR values and high IoU (71.56%–85.82%) for spatial alignment.
- Demonstrates accurate, complete, and photorealistic 3D reconstruction of surgical scenes.
Conclusions:
- EndoLRMGS offers a powerful and versatile solution for comprehensive 3D surgical scene reconstruction.
- The modular approach effectively handles both deformable tissues and articulated instruments.
- Enables more accurate and detailed analysis and planning in robot-assisted surgery.
