PENGWIN 2026 Task 3: Peripelvic Fracture Reduction Planning Dataset

The fracture reduction planning task (Task 3) of the PENGWIN 2026 challenge is designed to advance the development of automated pelvic fracture reduction planning methods, with a focus on predicting rigid transformation matrices that restore fractured bone fragments to their anatomical configuration.

This repository contains two training datasets:

Simulated Pelvic Fracture Dataset: This dataset covers 480 subjects with 1,440 bones, each provided with 20 fracture variants, enabling ~4.4M combinatorial pelvic fracture configurations. The data is synthetically generated from CT scans sourced from CTPelvic1K and TotalSegmentator. Each pelvis is decomposed into three bones — sacrum (SA), left ilium (LI), and right ilium (RI) — with fracture modes simulated using DFGM, BDRD, break, and voro methods within a framework based on FracFormer. Realistic pelvic fracture configurations are constructed by freely combining fracture modes or intact templates across the three bones of the same subject, followed by the application of fragment displacement poses. Each bone is provided in two representations: voxel segmentation masks (NIfTI, .nii.gz) and pre-sampled point clouds (NumPy .npz, 5,000 points per bone). All samples are provided in the pre-displacement state.

Clinical Pelvic Fracture Dataset: This dataset comprises 170 anonymized clinical cases of real pelvic fractures. Each case includes a multi-fragment pelvis mesh and the corresponding ground-truth rigid transformation matrices for fracture reduction. Fragments are organized within a single mesh file by bone type using integer group IDs (1–100: sacrum; 101–200: left ilium; 201–300: right ilium). A reference I/O script is provided for loading and visualizing the data. This mesh format serves as the input for the test phase, and participant algorithms are expected to output reduction matrices in the same JSON structure.

For more information, please visit the challenge webpage. For the PENGWIN 2026 Task 1 and Task 2 datasets (Automatic and Interactive Peripelvic Fracture Segmentation), please visit the separate repository (10.5281/zenodo.19732767).