PelviSet

PelviSet: A unified pediatric pelvic X-ray dataset for developmental dysplasia of the hip · version 1.0.0 · 2026-10-24

PelviSet unifies four public pediatric anteroposterior pelvic radiograph datasets (14,622 cases, three institutions, two countries) under one annotation scheme: 12-class anatomical segmentation masks, 8 keypoints, and geometry-derived clinical measurements. Annotation gaps in the sources were filled by quality-gated pseudo-labeling; the *_source columns in master.csv record which annotations are source-provided and which are pipeline-derived.

The canonical archive is on Zenodo: https://doi.org/10.5281/zenodo.20615290. This repository mirrors it as browsable files.

Sources

Folder	Source	Institution	Cases	Images included	Masks	Image licence
av_ddh/	AV-DDH	Jordan University Hospital, Jordan	2,417	2,417	2,417	CC BY 4.0
mtddh_ds1/	MTDDH Dataset 1	Children's Hospital, Zhejiang University School of Medicine, China	1,300	1,300	1,300	CC BY 4.0
mtddh_ds2/	MTDDH Dataset 2	Children's Hospital, Zhejiang University School of Medicine, China	905	905	905	CC BY 4.0
ustc/	USTC-DDH	Anhui Provincial Children's Hospital, China	10,000	0	10,000	images not redistributed
Total			14,622	4,622	14,622

USTC-DDH images

The USTC-DDH terms do not allow redistribution of its images, so ustc/ contains masks only. Its annotations (masks, keypoints, measurements) are released here; this was confirmed in correspondence with the USTC-DDH authors. To obtain the images, send a signed commitment letter confirming permitted use to the USTC-DDH authors (Liu et al., IEEE TMI 2020); they grant access. Place the images in ustc/images/ as ustc_XXXXXX_image.jpg, matching case_id.

Structure

pelviset/
├── master.csv          one row per case: ids, split, keypoints, measurements, provenance
├── SHA256SUMS.txt      checksum of every file
├── README.md
├── av_ddh/
│   ├── images/       av_ddh_XXXXXX_image.{png|jpg}
│   └── masks/        av_ddh_XXXXXX_mask.png
├── mtddh_ds1/
│   ├── images/       mtddh_ds1_XXXXXX_image.{png|jpg}
│   └── masks/        mtddh_ds1_XXXXXX_mask.png
├── mtddh_ds2/
│   ├── images/       mtddh_ds2_XXXXXX_image.{png|jpg}
│   └── masks/        mtddh_ds2_XXXXXX_mask.png
└── ustc/
    ├── images/       ustc_XXXXXX_image.{png|jpg}   (empty: see USTC-DDH images)
    └── masks/        ustc_XXXXXX_mask.png

Annotations

Masks. Single-channel PNG, one label per pixel, same size as the image.

Value	Structure
0	background
1	right ilium
2	left ilium
3	right obturator foramen
4	left obturator foramen
5	right pubis
6	left pubis
7	right ischium
8	left ischium
9	right femoral head
10	left femoral head
11	right proximal femur
12	left proximal femur

Keypoints. Absolute pixel coordinates in master.csv, columns {R|L}{landmark}_{x|y}.

Landmark	Definition
TCC	tri-radiate cartilage center
ASM	acetabulum superolateral margin
FHC	femoral head center
MOFM	midpoint of the superior margin of the ossified femoral metaphysis

Measurements. Computed deterministically from the landmarks with the open-source Pelvitect engine: r_ai, l_ai, r_ce, l_ce, r_ihdi, l_ihdi, r_tonnis, l_tonnis.

All master.csv columns: case_id, source, original_split, split, image_path, width, height, bbox_x, bbox_y, bbox_w, bbox_h, RTCC_x, RTCC_y, LTCC_x, LTCC_y, RASM_x, RASM_y, LASM_x, LASM_y, RFHC_x, RFHC_y, LFHC_x, LFHC_y, RMOFM_x, RMOFM_y, LMOFM_x, LMOFM_y, mask_path, age_months, age_group, sex, kp_source, mask_source, original_id, r_ai, l_ai, r_ce, l_ce, r_ihdi, l_ihdi, r_tonnis, l_tonnis

Splits

Source	golden_kp	golden_seg	test	train	val
av_ddh	0	0	242	1,933	242
mtddh_ds1	63	0	150	940	147
mtddh_ds2	87	100	0	629	89
ustc	0	0	0	9,500	500
Total	150	100	392	13,002	978

golden_seg and golden_kp are frozen evaluation sets carrying source-provided labels only. Do not train on them, and report them separately from test.

Loading

from huggingface_hub import snapshot_download
import numpy as np, pandas as pd
from PIL import Image

root = snapshot_download("amirmmahdavikia/pelviset", repo_type="dataset")
df = pd.read_csv(f"{root}/master.csv")
row = df.iloc[0]
src = row.case_id.rsplit("_", 1)[0]
mask = np.array(Image.open(f"{root}/{src}/masks/{row.case_id}_mask.png"))  # labels 0..12
kps = row[["RTCC_x", "RTCC_y", "LTCC_x", "LTCC_y"]]

Limitations

• Most annotations are pipeline-derived. Two clinicians reviewed 200 stratified cases; in the random stratum, both rated segmentation acceptable in 98/100 and keypoints in 97/100.
• The femoral head center is ill-defined before the ossific nucleus appears. Treat FHC-dependent measurements (center-edge angle, Tönnis grade) with caution in infants aged 0–5 months.
• Recurring failure modes: orthopedic hardware and spica casts, gonadal shields (common in AV-DDH), severe dislocation, and limited field of view.

Corrections

Found a wrong mask or landmark? Open a Discussion on this repository with the case_id and what is wrong. Corrections are released as new versions.

License

Annotations, and the MTDDH and AV-DDH images, are released under CC BY 4.0. USTC-DDH images are not included. Please also cite the source datasets.

Citation

@dataset{mahdavikia2026pelviset,
  author    = {Mahdavikia, Amir M. and Nabian, Mohammad Hossein and Aghaghazvini, Leila and McArthur, Adam J. M. and Rouzrokh, Pouria and Jaremko, Jacob L.},
  title     = {PelviSet: A unified pediatric pelvic X-ray dataset for developmental dysplasia of the hip},
  year      = {2026},
  version   = {1.0.0},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.20615290}
}

Source datasets:

• Liu C, Xie H, Zhang S, Mao Z, Sun J, Zhang Y. Misshapen pelvis landmark detection with local-global feature learning for diagnosing developmental dysplasia of the hip. IEEE Trans Med Imaging 2020;39(12):3944–3954. doi:10.1109/TMI.2020.3008382
• Qi G, Jiao X, Li J, et al. A dataset for quality evaluation of pelvic X-ray and diagnosis of developmental dysplasia of the hip. Sci Data 2025;12:865. doi:10.1038/s41597-025-05146-x
• Haddad B, AlHosanie T, Dweidari A, et al. Acetabular-vision hip developmental dysplasia's (AV-DDH). Mendeley Data, 2025. doi:10.17632/4gvcb6gmh2.1

Authors

• Amir M. Mahdavikia, School of Medicine, Tehran University of Medical Sciences, Tehran, Iran · ORCID 0009-0000-5522-1415
• Mohammad Hossein Nabian, Center for Orthopedic Trans-Disciplinary Applied Research, Tehran University of Medical Sciences, Tehran, Iran
• Leila Aghaghazvini, Department of Radiology, Tehran University of Medical Sciences, Tehran, Iran
• Adam J. M. McArthur, Department of Radiology and Diagnostic Imaging, University of Alberta, Edmonton, AB, Canada · ORCID 0009-0001-1371-5289
• Pouria Rouzrokh, Diagnostic Radiology, Yale University, New Haven, CT, USA
• Jacob L. Jaremko, Department of Radiology and Diagnostic Imaging, University of Alberta, Edmonton, AB, Canada · ORCID 0000-0001-5314-2297