High-Quality Gaussian Surface Mesh Reconstruction for Endoscopic Deformable Tissue

Reconstructing 3D scenes in endoscopic videos is a core technology in the fields of robot-assisted surgery, virtual reality and augmented reality. Compared with Neural Radiation Field (NeRF), the emergence of 3D Gaussian Splatting (3DGS) significantly reduces the time of 3D reconstruction and rendering while maintaining high rendering quality to a certain extent, which provides new possibilities for real-time 3D rendering during endoscopic procedures. However, current research on this technique in endoscopy focuses on point cloud reconstruction of tissues, and it is not yet possible to directly generate meshed surfaces, limiting the accuracy of depth perception. To resolve this issue, this study proposes a surface reconstruction of endoscopic videos based on two-dimensional Gaussian Splatting (2DGS). The method removes global highlights and corrects chromatic aberration in endoscopic images by incorporating a fine-tuned visual foundation model, followed by a projected 2D rasterised Gaussian splatting method to complete the fast reconstruction of deformable tissues and achieve the final surface meshing effect. The results of the experiments indicate that, compared with the classical NeRF -based method, this method greatly reduces the requirement of computational equipment and the time needed, and successfully solves the problems of poor quality and inability to generate continuous surfaces, which currently exist in the field of mesh reconstruction of deformable tissues of endoscopes with Gaussian splatting method. This method greatly improves the reconstruction quality, provides a more reliable solution for 3D rendering in endoscopic surgery, and provides a new direction for subsequent related research.