Work

Endo4DGSLAM: Multi-level deformable modeling with redundancy reduction for endoscopic non-rigid SLAM

Year: 2026

Type: article

Abstract: Medical endoscopic SLAM enables real-time pose estimation and 3D mapping inside the human body, holding significant value for surgical navigation. Although emerging 3DGS improves SLAM through fast, high-quality rendering, mainstream methods remain constrained by the static-scene assumption. Recent dynamic GS-SLAM methods struggle with irregular tissue motion and suffer from high computational and deformation modeling overhead, hindering real-time online optimization. To address these challenges, we present a dynamic GS-SLAM framework specifically tailored for endoscopic environments. To handle complex tissue deformation in endoscopic scenes, we adopt a multi-scale motion representation with timestamp-aware encoding, which captures tissue motion and viewpoint variations across different temporal scales in continuous observations. In addition, we incorporate several acceleration strategies to reduce the computational overhead of the GS-SLAM pipeline. Specifically, the method reuses deformation information to construct a motion mask, which mitigates the influence of locally dynamic tissues on camera tracking and reduces unnecessary backpropagation during optimization. Meanwhile, the dynamic resolution downsampling and history-aware pruning strategies adaptively adjust the resolution and evaluate Gaussian importance across frames, effectively removing redundant pixels and Gaussian primitives while preserving spatiotemporal consistency. We evaluate our method on four public endoscopic datasets and achieve competitive results, with consistent improvements observed in several pose estimation and novel view synthesis metrics. In addition, our system achieves an end-to-end throughput of 26.3 FPS. At the algorithmic level, our method delivers approximately a 5× reduction in per-iteration tracking time compared to the leading dynamic SLAM approach.

Source: Medical Image Analysis

DOI: 10.1016/j.media.2026.104270