MediSplat-ICP: A Hybrid Geometric-Photometric SLAM Method for High-Fidelity Surgical Scene Reconstruction

Addressing the urgent need for 3D reconstruction and navigation in robotic surgery, this paper introduces a novel hybrid SLAM framework named MediSplat-ICP to tackle challenges such as weak textures, specular highlights, and tissue deformation under endoscopic views. The framework innovatively combines a fast and robust G-ICP geometric tracking front-end with a high-fidelity 3D Gaussian Splatting mapping back-end. Its core contribution is the design of a novel coupling mechanism that utilizes the geometric covariance information generated during GICP tracking to intelligently guide the initialization of 3D Gaussian primitives, thereby significantly improving the speed and quality of the mapping process. Extensive experiments on standard datasets demonstrate that MediSplat-ICP achieves state-of-the-art performance in tracking accuracy, reconstruction quality, and robustness, laying a solid technical foundation for future highprecision computer-assisted surgery systems.