Gastroendoscopy View Synthesis: A New Real Dataset and Evaluation

Novel view synthesis (NVS) holds significant promise for medical applications, such as creating digital twins for training or extending the field of view in endoscopic procedures. However, existing datasets for NVS in gastroendoscopy are either synthetic, limited in viewpoint variety, or designed for different tasks like surgical tool occlusion removal. This paper addresses this gap by introducing the first real-world dataset specifically constructed for NVS in gastroendoscopy.

The authors present the GastroNVS dataset, which includes endoscopic image sequences from five human subjects. To ensure high-quality data, they used indigo carmine blue dye to enhance surface features, facilitating robust structure-from-motion (SfM) reconstruction. The dataset provides camera poses, intrinsic parameters, and 3D point clouds. The researchers evaluated several 3D Gaussian splatting (3DGS) methods—including standard 3DGS, 3DGS+depth, 2DGS, PGSR, and GSDF—using two different data splits: one with regular intervals (Split-Reg) and one with consecutive segments (Split-Con) to test performance on truly novel viewpoints.

The study reveals that standard 3DGS often struggles with the complex, curved, and textureless surfaces of the stomach, leading to artifacts and blurred regions. Methods that explicitly optimize for geometry—specifically 3DGS+depth and GSDF—consistently produce higher-quality renderings and more accurate surface representations. The performance gap between methods becomes more pronounced in the more challenging Split-Con scenario, where GSDF demonstrated superior robustness. The authors also identified a critical challenge specific to endoscopy: inter-view illumination inconsistency, where variations in lighting intensity and direction across frames significantly impact reconstruction quality.