# HandBones (hand radiographs with per-bone masks)

Part of the FleXray data release. Repackaged from the original distribution at
https://universe.roboflow.com/boneage-x90qt/-hand-bones-mdjkr under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/).

**Citation:** Roboflow Universe project `boneage-x90qt/-hand-bones-mdjkr` (uploader-declared CC BY 4.0).

## Contents

- 93 samples (test: 13, train: 65, val: 15); one image and one label per sample.
- `images/*.png`: 16-bit grayscale PNG, intensities in [0, 65535] map linearly to [0, 1].
- `labels/`: binary channel masks (NPY, C x H x W). Label ids/channels:
- channel `0`: background
- channel `1`: carpals
- channel `2`: phalange_distal
- channel `3`: phalange_intermediate
- channel `4`: metacarpals
- channel `5`: phalange_proximal
- channel `6`: radii
- channel `7`: ulnae
- `dataset.yml`: `fxr-dataset` manifest (schema version 1). Splits are the
  ones used in the FleXray paper; `metadata.original_shape` is the source image shape.

## Preprocessing

Every image was min-max normalized to [0, 1] per image, padded to a square with zeros
(centered) and resized to 256 x 256 (area interpolation for images, nearest for labels). Labels were remapped from the
source annotation to the dataset-native names listed above; the FleXray package maps
those names into its label protocol at training time.

## Usage

```bash
pip install flexray
fxr-dataset validate dataset.yml
fxr-dataset pack dataset.yml /data/flexray/HandBones
```