From Geometric Initialization to Gaussian Refinement: A Depth-Prior-Guided Framework for Monocular Endoscopic 3D Gaussian Reconstruction

Abstract Monocular endoscopic 3D reconstruction is fundamental to computer-assisted minimally invasive surgery. Recent 3D Gaussian Splatting (3DGS)-based methods enable efficient reconstruction, but their performance strongly depends on the initial geometry, which is commonly obtained through Structure-from-Motion (SfM). In endoscopic scenes, weak textures, specular reflections, and tissue deformation limit stable feature correspondences, often resulting in sparse or incomplete SfM point clouds for Gaussian initialization. Depth back-projection provides an alternative means of constructing dense initial geometry, but estimates from a single depth model may introduce model-specific biases that propagate through subsequent optimization. To alleviate this issue, we propose a collaborative expert depth-guided Gaussian reconstruction framework that improves Gaussian reconstruction from both initialization and refinement perspectives. Specifically, Multi-Model Depth Fusion Initialization robustly fuses depth predictions from multiple models to guide dense Gaussian initialization. During subsequent optimization, Tri-Domain Pyramid Regularization augments spatial-domain supervision with multi-scale pyramid constraints in the wavelet and gradient domains to refine the initialized Gaussian representation. Extensive experiments on two real surgical datasets, SCARED and EndoNeRF, demonstrate that MDGGR achieves state-of-the-art reconstruction quality, with PSNR improvements of 4.4\% and 7.8\% and LPIPS reductions of 35.7\% and 59.2\% over prior methods, respectively, particularly in terms of rendering fidelity and fine-grained detail reconstruction.