VictorButoi/flexray-data

FleXray data release

• —Website: FleXray project page
• —Paper: FleXray: Universal Clinical X-ray Segmentation
• —Code: github.com/VictorButoi/FleXray

Training and evaluation data for FleXray, a pan-anatomy X-ray segmentation model. This repository holds every real X-ray source whose license permits redistribution, repackaged as image/mask pairs with fxr-dataset manifests and dataset-native label names that FleXray maps into its common protocol (CC BY 4.0), plus two more parts of the data described in the paper:

• —FluXray (synthetic, CC BY-NC 4.0): FluXray/ — 138,063 generatively edited digitally reconstructed radiographs rendered from the 1,597 MOOSE CTs at 90 poses each, quality-filtered, with exact overlapping masks in 63 channels (62 structures plus background). Shipped as the training database itself (FluXray/thunder_dbs/1.0/data.mdb, an LMDB of float16 256 x 256 images and 63-channel binary masks that flexray reads directly) together with samples.csv (pose, split, MOOSE subject and per-sample license for every image), protocol.yml (label order and mask thresholds), and the filter_*.csv quality-control scores and thresholds. FluXray/README.md documents every file; FluXray/LICENSE summarizes the license.
• —MURA forearm/humerus annotations (masks only, CC BY 4.0): bundled with the FleXray GitHub repository and mirrored here under mura_forearm_humerus_annotations/

Licenses therefore differ by folder: the redistributed real X-ray folders, splits/, and the MURA annotations are CC BY 4.0; FluXray/ is CC BY-NC 4.0 as a collection, with each image inheriting the license of its MOOSE source site.

Data access

This release contains 4,232 real image/mask pairs across seven datasets, 138,063 FluXray samples, and 100 MURA annotation masks whose source images must be obtained separately. Browse the dataset folders below or use the download and packaging instructions.

The automatic Hugging Face image-folder viewer is disabled because it only recognizes a subset of the images and does not represent the paired masks or the FluXray database. Use each dataset.yml for image/mask paths and split assignments; FluXray's samples.csv records its samples and splits.

• —Real image/mask datasets: PNG images, PNG or NPY masks, and a packaging manifest per dataset.
• —FluXray: a ThunderDB database plus protocol and sample metadata. The database download is about 18.8 GB.
• —MURA annotations: masks and a script that joins them to your own MURA download.
• —Splits and exclusions: source-relative records for reconstructing the published partitions.

Redistributed datasets (CC BY 4.0)

Each <Dataset>/ folder contains dataset.yml, images/ (16-bit PNG), labels/ (PNG index maps or NPY channel masks), a README.md with preprocessing and label details, and a LICENSE with attribution. Images were min-max normalized per image, zero-padded to a square and resized to 256 x 256; splits are the ones used in the paper.

Datasets referenced by pointer only

These sources are used by FleXray but not redistributed here. Their FleXray label specifications (native label names, protocol aliases and drops) are shipped with the flexray package under fxr/configs/datasets/<Name>.yml.

Splits and exclusions

splits/<Dataset>/splits.csv lists the train/val/test assignment of every image FleXray trained or evaluated on, and splits/<Dataset>/exclusions.csv lists every image removed during quality control together with the reason, for all fifteen real X-ray sources above (redistributed or not). Paths are relative to each source's original download, so the paper's partitions can be rebuilt exactly; see splits/README.md for the schema.

Usage

Dataset packaging and training require the training extra:

python -m pip install "flexray[train]"

Download only the dataset you need. For example, fetch HipRay and its split records without downloading the FluXray database:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="VictorButoi/flexray-data",
    repo_type="dataset",
    local_dir="./flexray-data",
    allow_patterns=["README.md", "HipRay/*", "splits/HipRay/*"],
)

Validate and pack it into the layout consumed by FleXray:

fxr-dataset validate ./flexray-data/HipRay/dataset.yml
fxr-dataset pack ./flexray-data/HipRay/dataset.yml /data/flexray/HipRay
export XRAY_DATAPATH=/data/flexray

Repeat with another folder name to package another real X-ray dataset. To include FluXray, download its approximately 18.8 GB database and sidecars into the same download directory:

snapshot_download(
    repo_id="VictorButoi/flexray-data",
    repo_type="dataset",
    local_dir="./flexray-data",
    allow_patterns=["FluXray/*"],
)

export GENERATED_DATAPATH="$PWD/flexray-data"

FluXray is already packaged at FluXray/thunder_dbs/1.0/. Its historical GENERATED_DATAPATH variable selects the storage root; it is configured as an Xray source:

# training config excerpt
data:
  Xray:
    HipRay: {}
    FluXray: {version: "1.0"}

See dataset documentation and training configuration for complete examples. Source datasets retain their roles in the paper; adding one to a training config does not change its published evaluation split.

Citation

If you use either FluXray or our MURA annotations, please cite the FleXray paper:

@software{butoi2026flexray,
  title = {FleXray: Universal Clinical X-ray Segmentation},
  author = {Butoi, Victor Ion and Gopalakrishnan, Vivek and
            Guttag, John V. and Dalca, Adrian V. and Dey, Neel},
  year = {2026},
  license = {MIT},
  url = {https://github.com/VictorButoi/FleXray}
}

If you use any of the other datasets, please cite the original dataset sources and comply with their copyright and license terms. Citations and licensing details are listed in the README.md and LICENSE files within each dataset folder.