DatasetsAugust 2026
ClearWrist: children's wrist X-rays with patient-level splits
- 0
- wrist X-rays
- 0
- children, each in one split
- 0
- kinds of finding boxed
1Summary
ClearWrist is 20,327 wrist X-rays from 6,091 children, rebuilt from the GRAZPEDWRI-DX release into one dataset. Every image carries its fracture label, its radiologist bounding boxes and flags for anything that could confuse a model, and the splits are made by patient, so no child appears in more than one of train, validation and test.
2Motivation
GRAZPEDWRI-DX is one of the largest public sets of children's wrist X-rays, but it ships as four zip archives (16 GB), a separate table of labels and a separate folder of boxes, joined by file name.
Copies on Hugging Face stripped the labels, leaving images with no fracture label, no patient ID and no boxes, and no way to build a fair test split. A model can then be tested on X-rays of a child it has already seen in training, which makes it look better than it is.
3Composition
One row per X-ray: the image, a preview with its boxes drawn, the boxes, the patient ID, age, sex, wrist side and projection, the fracture label and its AO classification, and flags for casts, metal, osteopenia and diagnostic uncertainty.
| Split | X-rays | Children | With a fracture |
|---|---|---|---|
| Train | 14,234 | 4,264 | 66.8% |
| Validation | 3,127 | 914 | 66.7% |
| Test | 2,966 | 913 | 66.1% |
About two thirds of the X-rays show a fracture, and the share is nearly the same in all three splits.
4Collection
The X-rays were taken at the Department for Pediatric Surgery of the University Hospital Graz between 2008 and 2018, and released as GRAZPEDWRI-DX on Figshare (Nagy et al., Scientific Data, 2022). Images, the label table and the YOLO box files were joined by file name, and each patient was assigned to one split with a fixed seed.
5Cleaning and labels
Nothing was re-annotated. The fracture labels and boxes are the original pediatric radiologists' annotations.
- Fracture (
fracture_visible) is the primary label. A finer AO fracture code is included where the source has one. - Confounders. Casts, metal, osteopenia and diagnostic uncertainty each have their own column, and
has_confoundmarks any of them, so a clean test subset is one filter away. - Boxes cover nine kinds of finding, from fractures to soft-tissue swelling. An image without boxes is not necessarily normal: only findings the radiologists marked are boxed.
- A fixed bug. An early version stored the training images as nearly blank 8-bit files. All 14,234 were pulled again from the source and checked pixel by pixel against the other splits.
6Uses
For training and testing fracture detection, classification and localization models, and for comparing general multimodal models with detection models such as YOLO on children's X-rays.
Report results on the test split, and on its confounder-free subset as well, since casts and metal make fractures easier or harder to see.
It is a research dataset. It is not for diagnosis or for decisions about real patients.
7Distribution
On Hugging Face as Layered-Labs/clearwrist-pediatric-wrist-xray, with the images stored inside the files:
from datasets import load_dataset
ds = load_dataset("Layered-Labs/clearwrist-pediatric-wrist-xray")
clean_test = ds["test"].filter(lambda r: r["has_confound"] == 0)
Released under CC BY 4.0, the same license as the source. All images are de-identified.
8Maintenance
Layered Labs maintains this release, and every change is recorded in the dataset's changelog. Questions and corrections go to the dataset's Community tab on Hugging Face.
- v1.0 First release.
- v1.0.1 Fixed the training images, which had been saved nearly blank, and one empty image.
- v1.1 Added the bounding boxes and a preview with the boxes drawn.
- v1.2 Renamed to ClearWrist; the old address redirects.
Living draft, updated November 2026. This page can change before formal publication.