MedOtter/PENGWIN_Task1

PENGWIN Task 1 — Pelvic Fracture Segmentation on CT

The CT task of the PENGWIN 2024 challenge (PElvic bone fraGment (WIN)dow, MICCAI 2024): segment the sacrum, left hipbone and right hipbone, and the individual fracture fragments of each, in preoperative pelvic trauma CT.

This is an instance segmentation task, not a 3-class semantic one — the label value identifies which fragment of which bone, and the fragment count varies per case.

What this mirror contains — read first

This is the 100-case public training split, not the full 150-case cohort. PENGWIN 2024 used 100 train / 20 validation / 30 test. Only the training split was ever released; validation and test were withheld for the leaderboard and have not appeared on Zenodo. Any "PENGWIN CT" number quoted as n=150 refers to the paper's cohort, not to available data.

Name collision — pin to the 2024 challenge. A separate PENGWIN 2026 challenge ("Peripelvic Fracture Segmentation and Reduction Planning") exists with its own Task 1/2/3 and different Zenodo records. This mirror is Zenodo 10927452, MICCAI 2024.

Not raw scans. The volumes are de-identified DICOM→MHA conversions, and 36 of 100 were cropped to the pelvic region — which is why the geometry varies per case (see Two processing batches below).

Dataset Details

Field	Value
Modality	CT (preoperative, before fracture reduction surgery)
Body part	Pelvis — sacrum, left hipbone, right hipbone + fracture fragments
Task	3D instance segmentation of bone fragments
Cases	100 (public training split of a 150-case cohort)
Cohort	6 Chinese hospitals, 2017–2023
Format	.mha (MetaImage), flat NNN.mha, 001–100 contiguous
Size	8.08 GB (losslessly compressed; 33.77 GB uncompressed)
Slices per case	193–414
In-plane	322×154 to 512×512 (71 distinct shapes)
Spacing	0.658–1.22 mm in-plane, 0.625–1.25 mm slice (75 distinct)
License	CC BY-NC-SA 4.0 — see the discrepancy note below
DOI	10.5281/zenodo.10927452

There is no official validation or test split in the release, and no patient/center metadata of any kind. Splitting is left to the consumer; see Two processing batches for the one grouping variable that is recoverable.

Label encoding

0 = background. Foreground encodes anatomy and fragment index:

Range	Anatomy
1–10	Sacrum fragments
11–20	Left hipbone fragments
21–30	Right hipbone fragments

anatomy      = (label - 1) // 10   # 0 sacrum, 1 left hipbone, 2 right hipbone
fragment_idx = (label - 1) %  10   # 0 = main fragment

Verified properties (checked on all 100 label volumes)

These were measured, not taken from the documentation, and several are easy to get wrong:

• Observed maximum label is 24, not 30. Values actually present across the release: 1–4, 11–16, 21–24. Do not size a one-hot buffer at 30 and assume the tail is populated.
• All three anatomies are present in all 100 cases — labels 1, 11 and 21 never missing.
• Groups are contiguous and always start at their base (1/11/21); no gaps in any of the 300 anatomy-groups.
• The base label is always the largest fragment in its group (300/300). However, the remaining fragments are not reliably size-ordered — 45 of 300 groups violate descending order (e.g. 006.mha sacrum: 1→76023, 2→52875, 3→33603, 4→68271). Do not infer size rank from the fragment index beyond the main fragment.
• Fragments per case: 3–9, mean 5.75.
• Label dtype is inconsistent: 98 int16, 2 uint8 (022.mha, 072.mha). Do not assume uint8.

⚠️ Mixed orientation — 34 cases are RAS, 66 are LPS

This is the single easiest thing to get wrong with this dataset.

Direction cosines	n	Orientation
diag(+1, +1, +1)	66	LPS
diag(−1, −1, +1)	34	RAS

No case is genuinely oblique — it is a clean ±1 flip on x and y.

Image and label share identical direction in every case, so per-case overlap metrics stay correct even if you ignore this. But a loader that calls GetArrayFromImage() without consulting the direction cosines will get 34 cases left–right and anterior–posterior flipped relative to the other 66. The consequence is semantic: labels 11–20 are the left hipbone anatomically, but land on opposite sides of the array depending on the case. Any model with a left/right prior, and any evaluation that treats 11–20 as a consistent class, is silently corrupted.

Canonicalize before use:

import SimpleITK as sitk
img = sitk.DICOMOrient(sitk.ReadImage("images/001.mha"), "LPS")
msk = sitk.DICOMOrient(sitk.ReadImage("labels/001.mha"), "LPS")

The per-case orientation column in train.jsonl records which is which.

Two processing batches

Orientation is a near-perfect proxy for whether a volume was cropped:

	512×512 in-plane	Cropped in-plane
LPS (66)	64	2
RAS (34)	0	34

Every RAS case is cropped (each to a distinct matrix size); 64 of 66 LPS cases are untouched 512×512. Image dtype correlates too — 79% of RAS cases are int32 versus 39% of LPS. This matches the Zenodo note that volumes containing extra anatomy "were cropped to contain the pelvic region": that second pass evidently also rewrote orientation.

So the 100 cases are two sub-populations produced by different pipelines. This is the only grouping variable the release exposes and is worth stratifying on. It is not a recovery of the 6-hospital split — PENGWIN publishes no center labels, and this correlation identifies processing batch, nothing more.

Image properties

• Image dtype is inconsistent: 53 int32, 47 int16. HU values fit comfortably in int16; the int32 cases are simply stored wider. This mirror preserves the original dtype rather than downcasting.
• Intensity ranges are wide (down to −6152, up to +24970 HU in some cases), consistent with trauma cohorts containing implants and metal.
• Image and label share an identical grid (size, spacing, origin, direction) in all 100 cases — verified — so no resampling is needed to pair them.

Ground truth — single gold tier

Two independent annotators (5+ years' experience) segmented each case in 3D Slicer, seeded by an nnU-Net pretrained on CTPelvic1K, after which a senior expert (15+ years) selected the better of the two annotations — they were not merged, and no STAPLE was applied. Fragments below 500 mm³ were omitted. Reported inter-annotator agreement: IoU 0.984, ARI 0.993.

Only one mask per case ships, so there is no multi-rater tier in this release and no rater ambiguity to resolve.

⚠️ Cross-dataset overlap — CTPelvic1K

Treat PENGWIN Task 1 and CTPelvic1K as potentially patient-overlapping.

CTPelvic1K's CLINIC subset is n=103 pelvic-fracture CT "collected from preoperative images without metal artifact" at a collaborating orthopedic hospital. PENGWIN's Beijing Jishuitan center contributed n=103 scans "acquired in high quality before fracture reduction surgery". Chunpeng Zhao and Xinbao Wu co-author both papers. Identical count, identical hospital, identical inclusion criteria.

Against exact identity: the scanner mix differs (CTPelvic1K's CLINIC is roughly 86 Toshiba + ~17 other; PENGWIN's JST is 58 Toshiba + 45 United Imaging), and PENGWIN spans 2017–2023, past CTPelvic1K's 2020 curation. Neither paper acknowledges any overlap.

Conclusion: not identical, but drawn from the same archive over an overlapping window. Partial patient overlap is likely and cannot be excluded from published metadata. There is no cross-reference ID — both releases use anonymized sequential IDs (001.mha–100.mha vs dataset6_CLINIC_0001–0103) and PENGWIN ships no patient, center or scanner fields. Deduplication would have to be content-based (match on spacing and slice count, then cross-correlate mid-axial slices within the overlapping FOV).

Two further leakage notes:

1. The ground truth is partly a function of CTPelvic1K. PENGWIN's annotations were seeded by an nnU-Net trained on CTPelvic1K, so the two label sets are not statistically independent even where the patients differ.
2. PENGWIN Task 2 X-rays are DeepDRR renderings of these same CT volumes. Using both tasks together creates internal patient overlap by construction.

No overlap with TotalSegmentator (Basel, routine whole-body CT) or VerSe (European multi-center spine CT). PENGWIN CT is newly collected Chinese hospital trauma data and shares nothing with CTPelvic1K's public-archive lineage (COLONOG / KITS19 / MSD-T10 / ABDOMEN / CERVIX).

⚠️ License discrepancy

Source	States
Zenodo record 10927452 metadata	CC BY 4.0 (cc-by-4.0, open access)
PENGWIN challenge report text	CC BY-NC-SA

These contradict. The same team has the mirror-image discrepancy on CTPelvic1K (paper says CC BY-NC-SA 4.0, Zenodo 4588403 says CC BY 4.0), so it appears systematic rather than a typo.

This mirror declares the more restrictive, author-stated CC BY-NC-SA 4.0 so that use is safe under either reading. Both licenses permit redistribution. If you need commercial or non-ShareAlike terms, consult the Zenodo record and contact the organizers rather than relying on this choice.

Structure

images/NNN.mha        # 100 CT volumes   (001-100)
labels/NNN.mha        # 100 instance masks, same grid as the image
train.jsonl           # per-case metadata, one JSON object per line
README.md
LICENSE.txt

train.jsonl columns:

Column	Meaning
case_id	"001" … "100"
image, mask	repo-relative paths
split	always "train" (no official val/test released)
shape_zyx, spacing_xyz, origin_xyz	geometry
orientation	"LPS" or "RAS" — see the orientation warning
is_cropped	true if in-plane is not 512×512
image_dtype, label_dtype	original dtypes (both are mixed)
hu_min, hu_max	intensity range
label_values	sorted foreground labels present
n_fragments	total fragments
n_sacrum_fragments, n_left_hip_fragments, n_right_hip_fragments	per-anatomy counts
fragment_voxels	{label: voxel_count}

Storage note

The .mha files are rewritten with lossless zlib compression (33.77 GB → 8.08 GB, 4.18×). Voxel arrays, dtype, spacing, origin and direction were verified bit-identical to the Zenodo originals on all 200 files (np.array_equal, exact, after a fresh re-read from disk). .mha compression is transparent to ITK/SimpleITK — no change to how you read the files.

Source & Citation

• Zenodo: https://doi.org/10.5281/zenodo.10927452 (open, no registration, no DUA)
• Challenge: https://pengwin.grand-challenge.org/ (an account is needed only for leaderboard submission, not for the data)

@article{sang2026pengwin,
  author  = {Sang, Yudi and Liu, Yanzhen and Yibulayimu, Sutuke and others},
  title   = {Benchmark of Segmentation Techniques for Pelvic Fracture in CT and
             X-Ray: Summary of the PENGWIN 2024 Challenge},
  journal = {IEEE Transactions on Medical Imaging},
  year    = {2026},
  doi     = {10.1109/TMI.2025.3650126}
}

@inproceedings{liu2023pelvic,
  author    = {Liu, Yanzhen and Yibulayimu, Sutuke and Sang, Yudi and Zhu, Gang
               and Wang, Yu and Zhao, Chunpeng and Wu, Xinbao},
  title     = {Pelvic Fracture Segmentation Using a Multi-scale Distance-Weighted
               Neural Network},
  booktitle = {MICCAI 2023},
  pages     = {312--321},
  year      = {2023},
  doi       = {10.1007/978-3-031-43996-4_30}
}

@article{liu2025automatic,
  author  = {Liu, Yanzhen and Yibulayimu, Sutuke and Zhu, Gang and others},
  title   = {Automatic pelvic fracture segmentation: a deep learning approach
             and benchmark dataset},
  journal = {Frontiers in Medicine},
  volume  = {12},
  pages   = {1511487},
  year    = {2025},
  doi     = {10.3389/fmed.2025.1511487}
}