Optimal Mixture Model Distribution Alignment-Based 3D-2D Gaussian Splatting Registration for Monocular Endoscopic Ar Guidance

Registration between preoperative 3D CT model and intraoperative 2 D endoscopic images is essential to enhance the application of Augmented Reality (AR) guidance in surgery. However, restricted by the complicated intracavitary scene, it is difficult to perform effective landmark-free registration between the 3D CT model and the 2D endoscopic images. To overcome these challenges, a novel 3D-2D Gaussian Splatting registration architecture for endoscopic surgery is proposed by performing optimal distribution alignment between the CT/Endo gaussian mixture model. Firstly, derived from the Gaussian Splatting pipeline, the CT/Endo gaussian mixture model distributions are constructed from the preoperative CT model and intraoperative endoscopic images, respectively. Secondly, based on the CT/Endo gaussian mixture model distribution, a novel optimal distribution alignment coarse registration module is designed to establish the global registration. Finally, to refine the precise registration in endoscopic images, a novel 2D rastering fine registration module is built to optimize the alignment between the 2D CT rastering and the organ binary mask of endoscopic images. The evaluation is carried out on the public Scared endoscopic dataset and a clinical endoscopic dataset from the local hospital. The experimental results show that the proposed method outperforms state-of-theart methods in quantitative and qualitative comparisons.