---
title: "FleXray: Universal Clinical X-ray Segmentation"
emoji: 🦴
colorFrom: red
colorTo: green
sdk: gradio
sdk_version: 6.28.0
app_file: app.py
short_description: Segment 60 anatomical structures from any clinical X-ray
python_version: "3.10"
startup_duration_timeout: 30m
---

# FleXray: Universal Clinical X-ray Segmentation

FleXray is a single 2D UNet that segments **60 anatomical structures** (plus background)
from standard radiographs across body regions, projections, and acquisition settings —
chest, spine, pelvis, hands, knees, and more — with one set of weights.

Upload an X-ray (or pick one of the examples), choose a quality mode, and the model
returns a color-coded anatomical overlay. Use the label checkboxes to focus on the
structures you care about and hover the overlay legend to identify each color.

**Quality modes** (matching the [official browser demo](https://flexray.csail.mit.edu/)):

| Mode | What runs |
| --- | --- |
| Low | flagship model, single pass — fastest |
| Normal | flagship model with 8-pass test-time augmentation |
| High | 5-model ensemble, single pass |
| X-High | 5-model ensemble with 8-pass TTA — best quality |

## Model

- Weights: [`VictorButoi/flexray`](https://huggingface.co/VictorButoi/flexray) (flagship + 4-member ensemble)
- Code: [github.com/VictorButoi/FleXray](https://github.com/VictorButoi/FleXray) (this Space uses the official `flexray` PyPI package 1:1)
- Paper: [FleXray: Universal Clinical X-ray Segmentation](https://arxiv.org/abs/2609.26756)

## Example images

The bundled examples (`examples/chest_lateral.png`, `examples/pelvis_ap.png`) are the two
radiographs the authors use in their own tutorial materials, courtesy of
[Radiopaedia.org](https://radiopaedia.org/), used with permission (see the
[FleXray repo](https://github.com/VictorButoi/FleXray/blob/main/examples/assets/README.md)).

## License & disclaimer

Model weights are released under CC-BY-NC-4.0 (code is MIT). This demo is for research
and educational purposes only — it is **not** a medical device and must not be used for
clinical decision-making.