MedPelvis3D: a 3D pelvic anatomy dataset for anatomical landmark detection

MedPelvis3D is a curated 3D pelvic anatomy dataset designed to support the development and evaluation of automated anatomical landmark detection, anatomical measurement, pose estimation, geometric learning, registration, and surgical-planning research methods. The dataset comprises 99 de-identified pelvic CT cases (59 males, 40 females) retrospectively collected at the Chinese PLA General Hospital. All coordinates are expressed in millimetres in the original patient-level CT physical coordinate frame, with the LPS coordinate convention preserved. Each case is provided with: * de-identified CT volume in NIfTI format (.nii.gz) * pelvic segmentation mask in NIfTI format, including LeftHipBone, RightHipBone, and Sacrum labels * surface meshes in STL format, named after the corresponding anatomical component (LeftHipBone.stl, RightHipBone.stl, Sacrum.stl) * 57 expert-annotated 3D anatomical landmarks in CSV format, with legacy F-id labels and standardized anatomical short_name / anatomical_name_en columns * 50,000-point reference point cloud sampled from the pelvic surface mesh (.npy) The release also includes a patient-level metadata file containing anonymized case_id and sex labels, together with a repository-level landmark anatomical mapping table. Quantitative validation reported in the accompanying Data Descriptor includes: * mean inter-ASIS distance: 228.5 ± 18.2 mm * mean cortical landmark-to-surface distance: 0.04 mm * mean intra-pelvic landmark-to-surface distance: 1.51 mm * inter-operator agreement on a 10-case subset including all 57 landmarks per case: mean landmark-level Euclidean distance of 4.56 ± 2.04 mm (95% CI 4.02-5.10 mm), with inter-operator ICC(2,1) values exceeding 0.997 along the x, y, and z axes Three example cases (600001, 600003, and 600052) are provided uncompressed under samples/ for quick preview without downloading the full archives. Files included: ct_nifti_part{1,2,3,4}of4.zip 99 .nii.gz CT volumes split into 4 archives (each part contains approximately 25 cases; extract all four archives into the same folder) masks_nifti.zip 99 segmentation masks in NIfTI format stl_models.zip 99 cases × 3 anatomical STL files point_clouds_50000.zip 99 case-level 50,000-point .npy files annotations.zip 99 landmark CSV files, each containing 57 landmarks patient_metadata.csv anonymized case_id and sex metadata landmark_anatomical_mapping.csv landmark id to anatomical name mapping samples.zip 3 fully formed example cases, provided uncompressed Citation: Guo, N., Pan, J., Zhang, G., Wang, J., Su, X., Qi, Y., Liu, M., Zhang, Q. & Hao, M. (2026).MedPelvis3D: a 3D pelvic anatomy dataset for anatomical landmark detection.Zenodo. https://doi.org/10.5281/zenodo.20473747 Companion Data Descriptor: Guo et al., Scientific Data (under review, 2026).