# FleXray manual annotations for MURA forearm and humerus radiographs

Bone segmentation masks for 100 radiographs from the MURA dataset
(Rajpurkar et al., 2017; https://stanfordmlgroup.github.io/competitions/mura/), one image
per patient, annotated manually with polygon tools in CVAT for the FleXray paper.

- `MURA_FOREARM`: 50 radiographs
- `MURA_HUMERUS`: 50 radiographs

The MURA images are **not** included: the MURA Research Use Agreement does not permit
redistribution. Download MURA yourself and join the masks with the manifest below.

## Files

- `manifest.csv`: one row per annotated image. `mura_relative_path` is the image path
  relative to the MURA download (starting with `MURA-v1.1/`); `mask_path` is the matching
  mask; `flexray_split` is the train/val/test assignment used in the paper.
- `<dataset>/masks/*.png`: 8-bit indexed PNG at the native MURA image resolution.
- `labels.json`: mask value to structure name (0: background, 1: humeri, 2: radii, 3: ulnae).
- `mura_to_fxr_manifest.py`: builds an `fxr-dataset` package manifest from your MURA copy.

## Build a FleXray training package

```bash
pip install flexray
python mura_to_fxr_manifest.py --mura-root /path/containing/MURA-v1.1 \
    --annotations . --dataset MURA_FOREARM --output-dir /data/mura_forearm_pkg
fxr-dataset pack /data/mura_forearm_pkg/dataset.yml /data/flexray/MURA_FOREARM
```

The join script reproduces the preprocessing used to train FleXray: per-image min-max
normalization, zero padding to a square, and resizing to 256 x 256 (area averaging for
images, nearest for masks).

## License

The annotations are released under CC BY 4.0. The MURA images remain subject to the
Stanford MURA Research Use Agreement.