VictorButoi / FleXray

# FleXray: Universal Clinical X-ray Segmentation 💪
### Paper · Project Page
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.
## What FleXray Does
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.
## Install FleXray
FleXray targets Python 3.10+. The PyPI package name is `flexray` (the import
package is `fxr`) and the default install is inference-focused:
 [code] 
## Predict From The Command Line
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:
 [code] 
`--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:
 [code] 
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`.
## Understand The Output Files
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 n