C3VD-Raycasting-10k: A Clinical Point Cloud Registration Dataset for Image-Guided Colonoscopy

<b>C3VD-Raycasting-10k</b> is a clinically grounded benchmark dataset for 3D point cloud registration in image-guided colonoscopy. It contains <b>10,014 geometrically aligned point-cloud pairs</b> that simulate the cross-modal alignment problem between preoperative CT anatomy and intraoperative endoscopic observations.The dataset is derived from clinical CT and endoscopy data provided by the Colonoscopy 3D Video Dataset (C3VD)<b> </b>[Bobrow et al., MedIA 2023]. Starting from complete CT-based colon meshes and recorded endoscope trajectories, we use <b>physics-based ray casting</b> to generate realistic intraoperative viewpoints. For each recorded camera pose, we cast rays from the endoscopic viewpoint onto the CT-derived surface to obtain a <b>partial target point cloud</b> that mimics what is observable during colonoscopy. The corresponding <b>source point cloud</b> is sampled from the <b>dense CT mesh</b> representing the underlying preoperative anatomy.Each sample in C3VD-Raycasting-10k therefore consists of:A <b>dense source point cloud</b> derived from the preoperative CT colon mesh.A <b>partial target point cloud</b> generated by ray casting from an endoscopic viewpoint, with occlusions and visibility constraints that reflect realistic intraoperative conditions.By construction, the dataset emphasizes challenging but clinically relevant cases, including:<b>Partial-to-partial alignment</b> with varying field-of-view, coverage, and missing regions.<b>Locally homogeneous geometry</b> and repetitive structures that cause feature degeneracy on tubular organ surfaces.<b>Cross-modal variability</b> between CT-derived anatomy and endoscopic appearance, while still providing precise geometric ground truth.C3VD-Raycasting-10k is designed to support <b>rigorous and reproducible benchmarking</b> of 3D registration algorithms for image-guided colonoscopy and related minimally invasive procedures.<b>Citing the Dataset</b>Cite [Linzhe:arXiv2025] whenever research making use of this dataset is reported in any academic publication or research report.<b>Declaration</b>This point cloud dataset is derived from the Colonoscopy 3D Video Dataset (C3VD) (https://durrlab.github.io/C3VD/).<br><br>Original data: Bobrow et al., "Colonoscopy 3D video dataset with paired depth from 2D-3D registration", Medical Image Analysis, 2023.<br><br>In accordance with the original C3VD dataset license, our derived point cloud dataset is also released under the CC BY-NC-SA 4.0 license and may only be used for non-commercial purposes.