FleXray: Universal Clinical X-ray Segmentation 💪

Victor Ion Butoi, Vivek Gopalakrishnan, John V. Guttag, Adrian V. Dalca, Neel Dey

FleXray (pip package flexray, imported as fxr) is a command-line tool and Python package for segmenting anatomy in X-ray images. The quickest path is to install the package, run flexify on an image or folder of images, and inspect the saved NumPy arrays.

FleXray can also train new full-body X-ray segmentation models. Training combines 2D X-ray segmentation data (including generated/FluXray samples) with CT segmentation volumes rendered to DRRs during training.

FleXray is a segmentation model that outputs anatomical segmentations for arbitrary X-rays in common image formats such as PNG, JPEG, TIFF, and BMP, plus DICOM (pip install flexray[dicom]). The pretrained weights live on our Hugging Face page, VictorButoi/flexray.

For every input image, FleXray writes three .npy arrays:

• Thresholded segmentation masks for each anatomical label.
• Per-channel probabilities that show model confidence.
• Raw logits for developers who need the unprocessed model scores.

It also writes one label_names.json naming those channels in order.

FleXray targets Python 3.10+. The PyPI package name is flexray (the import package is fxr) and the default install is inference-focused:

python -m pip install flexray

From the terminal, run prediction at one of four quality levels. These are the same Low / Normal / High / X-High modes as the browser demo:

# Low: flagship model, one forward pass
flexify --input ./image.png --output-dir ./predictions

# Normal: flagship model with 8-pass test-time augmentation
flexify --tta-samples 8 --input ./image.png --output-dir ./predictions

# High: five-model FleXray ensemble, one pass per model
flexify --ensemble --input ./image.png --output-dir ./predictions

# X-High: five-model ensemble with 8-pass test-time augmentation
flexify --ensemble --tta-samples 8 --input ./image.png --output-dir ./predictions

--input also accepts a directory, in which case every supported image in it is predicted. --model-id defaults to VictorButoi/flexray, one Hugging Face repository that holds the flagship and its four ensemble members under members/ (listed in its ensemble.json). Pass --subfolder to run one member on its own, --revision to pin a repository commit, or repeat --model-id to average bundles from several repositories:

flexify --subfolder members/flux000 --input ./image.png --output-dir ./predictions

By default, every model label is written. Pass --binary LABEL to restrict the output to one anatomical label, for example --binary femurs. Masks use --threshold 0.5 by default. For more documentation around command-line usage, see docs/inference.md.

For an input named image.png, flexify writes:

• image_masks.npy: thresholded segmentation masks. These are uint8 arrays where each channel is a predicted anatomical mask.
• image_probabilities.npy: model confidence per channel. These are float32 arrays with values after the selected probability conversion mode.
• image_logits.npy: raw model scores before probability conversion. These are mostly useful for developers and debugging.

All three arrays are channel-first with shape CxHxW, where C is the number of model output labels and H/W come from the bundle's preprocessing size. With --binary LABEL, C is 1. The CLI writes .npy arrays and does not write PNG overlays.

We also support Python. The API runs the same Hugging Face-backed workflow, with the same quality levels, and returns logits, probabilities, and masks in memory:

from fxr.inference import FleXraySegmenter

segmenter = FleXraySegmenter.from_pretrained()                   # flagship model
segmenter = FleXraySegmenter.from_pretrained(ensemble=True)      # five-model ensemble
prediction = segmenter.predict("./image.png", tta_samples=16)    # tta_samples=1: one pass

logits = prediction.logits
probabilities = prediction.probabilities
masks = prediction.masks

For more details on the inference code, see docs/inference.md.

FleXray ships an MCP server, so AI clients such as Claude Code and Claude Desktop can segment X-rays and inspect label protocols through natural language:

python -m pip install "flexray[mcp]"
claude mcp add flexray -- fxr-mcp

The server exposes segment_image (with per-label statistics), model listing and description, and protocol/dataset introspection tools over local stdio. See docs/mcp.md for the full tool reference and client configuration.

The training release has three parts: the model CLI, the data engine CLI, and the label harmonizer and config system. Install them with:

python -m pip install "flexray[train]"

For repository development on Linux, uv sync --extra train --extra test uses PyTorch's official CUDA 12.6 index so Volta/V100 cards remain supported. The packaged base recipe reproduces the training run behind the released FleXray weights; see docs/training.md.

• flexify segments images with a published bundle (above).
• fxr-train validates and runs training; fxr-submit submits config-driven sweeps to a cluster. Local training accepts --device {auto,cpu,cuda} and --gpu N.

Initialize from the released model with --init-from VictorButoi/flexray; add --replace-head (and optionally --freeze-backbone) to fine-tune onto a different label protocol. --init-from-run RUN_DIR seeds a new run from a trusted local run and --resume RUN_DIR continues an interrupted run with its optimizer, scheduler, EMA, and random state. Training is single-process with at most one GPU.

fxr-dataset turns images/masks or CT volumes into the ThunderDB packages training reads, and fxr-render renders DRR training samples from packed CT:

fxr-dataset scaffold xray-seg --images ./images --masks ./masks --name MyXrays --output dataset.yml
fxr-dataset validate dataset.yml
fxr-dataset pack dataset.yml /data/MyXrays        # optional preprocessing/crops blocks
fxr-dataset check /data/MyXrays                   # or `inspect` for any ThunderDB
fxr-render /data/MyCT --dataset-name MyCT --profile MOOSE --output ./drr

CT manifests accept NIfTI volumes, an HU window, and the 512x512x256 axial crop layout the released model trains on; X-ray manifests can reproduce its pad-to-square/area-resize preprocessing. See docs/datasets.md and docs/camera.md.

A dataset spec maps a dataset's native mask ids into the FleXray protocol (merging finer labels, dropping structures the protocol lacks); fxr-protocol inspects protocols and compiles those mappings, and data.<modality>.<name>.dataset_spec wires a spec into a training config:

fxr-protocol list
fxr-protocol show all_structures_flexray_v4
fxr-protocol compile --dataset CustomHips.yml

The runnable walkthrough in examples/custom_dataset builds a tiny synthetic dataset, harmonizes its labels, packs it, and dry-runs training on CPU in seconds:

cd examples/custom_dataset
python make_synthetic_dataset.py --output work/data
fxr-protocol compile --dataset CustomHips.yml
fxr-dataset validate work/data/dataset.yml
fxr-dataset pack work/data/dataset.yml "$PWD/work/packed/CustomHips"
fxr-train train_custom.yml \
  --set data.Xray.CustomHips.path="$PWD/work/packed/CustomHips" \
  --set data.Xray.CustomHips.dataset_spec="$PWD/CustomHips.yml" \
  --set log.root="$PWD/work/runs" --device cpu --dry-run --smoke-data

Training configs are packaged YAML (fxr/configs/training/base.yml) that inherit with _base_: and take --set KEY=VALUE overrides; see docs/protocols.md and docs/config.md.

If you find FleXray or any of its materials useful, please cite our paper. See CITATION.cff for citation metadata.

@misc{butoi2026flexray,
      title={FleXray: Universal Clinical X-ray Segmentation},
      author={Victor Ion Butoi and Vivek Gopalakrishnan and John V. Guttag and Adrian V. Dalca and Neel Dey},
      year={2026},
      eprint={2609.26756},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2609.26756},
}

• Code is released under the MIT License.
• Public model weights are released under CC-BY-NC-4.0 unless a model card says otherwise.