README.md

---
license: cc-by-nc-sa-4.0
task_categories:
- image-segmentation
tags:
- medical
- x-ray
- pelvis
- fracture
- synthetic
- multi-label
pretty_name: 'PENGWIN Task 2: Pelvic Fragment Segmentation on Synthetic X-ray Images'
dataset_info:
  features:
  - name: image_display
    dtype: image
  - name: overlay
    dtype: image
  - name: image
    dtype: image
  - name: mask
    dtype: image
  - name: image_id
    dtype: string
  - name: case_id
    dtype: int32
  - name: projection_index
    dtype: int32
  - name: has_hardware
    dtype: bool
  - name: fragment_labels
    list: int32
  - name: n_fragments
    dtype: int32
  - name: n_sa
    dtype: int32
  - name: n_li
    dtype: int32
  - name: n_ri
    dtype: int32
  splits:
  - name: train
    num_bytes: 37941083347
    num_examples: 50000
  download_size: 37948398318
  dataset_size: 37941083347
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

# PENGWIN Task 2: Pelvic Fragment Segmentation on Synthetic X-ray Images

Mirror of the **training split** of Task 2 of the MICCAI 2024 PENGWIN challenge
(https://pengwin.grand-challenge.org/), from the official Zenodo record
[10913196](https://zenodo.org/records/10913196) (`train.zip`, md5 `9c90215dae54d8f494a85cfc7b19bc96`).

**These are SYNTHETIC X-rays, not real radiographs**: DeepDRR renders of the 100 PENGWIN
Task 1 training CTs simulating intraoperative C-arm fluoroscopy, 500 random poses per CT
= **50,000 image/mask pairs**. Projections `0000-0249` show clean anatomy; `0250-0499`
additionally contain up to 10 simulated K-wires/orthopaedic screws (`has_hardware`).
The challenge validation (8,000) and test (600) X-rays were never publicly released.
No real radiographs exist anywhere in PENGWIN 2024. (Zenodo's description writes the
hardware range as `0250-0500`; indices verifiably end at `0499`.)

## Columns

| Column | Content |
|---|---|
| `image_display` | uint8 JPEG, official `visualize_drr` rendering (neg-log -> CLAHE -> invert). Browsing aid, lossy. |
| `overlay` | RGB JPEG, per-fragment color fill + contour on `image_display`. Browsing aid, lossy. |
| `image` | **Raw float32 448x448 DRR** (lossless deflate TIFF, pixel-identical to Zenodo). Intensities are pre-neg-log; apply `-log` + windowing before use (see below). |
| `mask` | **uint32 bit-encoded multi-label segmentation** (lossless deflate TIFF; stored int32, values < 2^31, pixel-identical to Zenodo). NOT a plain label map. |
| `image_id` | `{case:03d}_{projection:04d}` |
| `case_id` | Source CT case 1-100 == the same patient's `PENGWIN_Task1` volume (see Overlap) |
| `projection_index` | 0-499 |
| `has_hardware` | `projection_index >= 250` |
| `fragment_labels` | Fragment labels present (set bit positions 1-30) |
| `n_fragments`, `n_sa`, `n_li`, `n_ri` | Fragment counts (total / sacrum / left hipbone / right hipbone) |

## Mask encoding

A pixel's uint32 value has bit `b = 10*(category-1) + fragment` set iff that fragment
projects onto the pixel (categories: 1 sacrum SA, 2 left hipbone LI, 3 right hipbone RI;
fragments 1-10, fragment 1 = main). **Overlapping fragments are the norm** (X-ray
projection superimposes bone), so decode to per-fragment binary masks - do not treat the
value as a class ID. Bit 0 is never set. Bit `b` corresponds exactly to label value `b`
in `MedOtter/PENGWIN_Task1`.

```python
import numpy as np
masks = [((seg >> b) & 1).astype(bool) for b in range(1, 31) if ((seg >> b) & 1).any()]
```

The official `pengwin_utils.py` (this repo's root, from the Zenodo record) provides
`seg_to_masks` / `masks_to_seg`, the DRR renderer, and the challenge augmentation
pipeline. Official deterministic test-time input: `neglog` then quantile window
`(0.01, 0.95)` (see `build_augmentation(train=False)`).

## Overlap warning

Derived from **exactly the 100 CTs in `MedOtter/PENGWIN_Task1`** (same case numbering,
same patients) - never treat the two as independent benchmarks. Task 1 in turn likely
shares patients with CTPelvic1K's CLINIC subset (no ID crosswalk exists), and its GT was
seeded by a CTPelvic1K-pretrained nnU-Net. Do not confuse with the separate PENGWIN 2026
challenge, whose "Task 2" is a different task on different data.

## License

The Zenodo record metadata declares CC BY 4.0, while the challenge summary paper's Data
Availability statement says the X-ray training set is released under **CC BY-NC-SA** - the
same record-vs-paper conflict as PENGWIN Task 1. As with our Task 1 mirror, this mirror
adopts the stricter author-stated **CC BY-NC-SA 4.0**.

## Citation

Sang Y. et al., "Benchmark of Segmentation Techniques for Pelvic Fracture in CT and
X-Ray: Summary of the PENGWIN 2024 Challenge," IEEE TMI, doi:10.1109/TMI.2025.3650126
(arXiv:2504.02382). Data: doi:10.5281/zenodo.10913196. Lineage: Liu Y. et al., MICCAI
2023, doi:10.1007/978-3-031-43996-4_30.