anycxr 0.1.1

Inference-only public package for AnyChest anatomical segmentation.

AnyCXR Public Inference

This repository is the inference-only public release for AnyCXR / AnyChest.

It is intended for reviewer and downstream use with:

• a local image file or folder of image files
• a local checkpoint bundle downloaded from Hugging Face
• no local training code
• no dependency on the original training workspace layout

What is included

• cxas/: inference package
• seg.py: CLI entry point
• cxas/data/anychest_reference.json: packaged reference JSON with class order, view metadata, and the original Dataset003_Full folder map
• pyproject.toml, setup.py: installation files

Current Scope

This public package currently exposes inference for:

• PA chest radiographs
• LA chest radiographs
• oblique chest radiographs at the packaged AnyChest oblique angles

The helper CLI creates only these PA, LA, and OB input folders for now. Additional workflows will be added gradually.

Installation

Create a Python environment and install this folder:

pip install -e .

After publication, install from PyPI with:

pip install anycxr

This installs two CLI commands:

anycxr --help
anychest-infer --help

anycxr is the user-facing helper. It creates the expected data folders, downloads the Hugging Face checkpoint bundle on first inference, and runs PA, LA, then oblique inference by default.

anychest-infer is the lower-level inference entry point. Use it when you want to pass a checkpoint and profile manually.

Docker

This repository includes a containerized inference environment:

• Dockerfile

Build the image:

docker build -t anycxr-infer .

Run the CLI inside the container:

docker run --rm -it \
  -v $(pwd):/workspace/AnyCXR \
  anycxr-infer --help

Initialize a Data Folder

On first interactive use, run:

anycxr init

The CLI asks you to choose a data folder, creates the PA, LA, and OB structure, and prints where to place images:

AnyCXR_data/
  PA/
    imagesTr/
  LA/
    imagesTr/
  OB/
    225/imagesTr/
    45/imagesTr/
    675/imagesTr/
    1125/imagesTr/
    135/imagesTr/
    1575/imagesTr/
  outputs/
  weights/

Use a non-interactive path when running on a server:

anycxr init --data-root /path/to/AnyCXR_data

Put PA images in PA/imagesTr, LA images in LA/imagesTr, and oblique images in the matching angle folder under OB.

Run Inference

Run every currently supported view in the fixed order PA -> LA -> OB:

anycxr infer --data-root /path/to/AnyCXR_data

Run only one view:

anycxr infer --data-root /path/to/AnyCXR_data --view pa
anycxr infer --data-root /path/to/AnyCXR_data --view la
anycxr infer --data-root /path/to/AnyCXR_data --view ob
anycxr infer --data-root /path/to/AnyCXR_data --view oblique_45

If the checkpoint bundle is missing, anycxr infer downloads it from Hugging Face into DATA_ROOT/weights before inference. Empty view folders are skipped.

Manual Checkpoint Download

Download the merged inference bundle from Hugging Face with the hf CLI:

hf download agaresd/anychest-inference anychest_inference_bundle.pt --local-dir ./weights

Model bundle repository:

• agaresd/anychest-inference

The merged bundle contains three slimmed inference checkpoints:

• la
• pa
• oblique

Oblique profiles share one checkpoint and expose the following view-specific bundle profiles:

• oblique_22_5
• oblique_45
• oblique_67_5
• oblique_112_5
• oblique_135
• oblique_157_5

The Hugging Face release also includes:

• a model card
• anychest_reference.json

Quick Start

The commands below use the lower-level anychest-infer entry point. Most users should start with anycxr init and anycxr infer.

Single LA radiograph:

anychest-infer \
  --input-path /path/to/image.png \
  --output-dir /path/to/output \
  --checkpoint ./weights/anychest_inference_bundle.pt \
  --profile la

Single PA DICOM:

anychest-infer \
  --input-path /path/to/image.dcm \
  --output-dir /path/to/output \
  --checkpoint ./weights/anychest_inference_bundle.pt \
  --profile pa

Single oblique image:

anychest-infer \
  --input-path /path/to/image.jpg \
  --output-dir /path/to/output \
  --checkpoint ./weights/anychest_inference_bundle.pt \
  --profile oblique_45

Flat folder of PA images:

anychest-infer \
  --input-path /path/to/folder \
  --output-dir /path/to/output \
  --checkpoint ./weights/anychest_inference_bundle.pt \
  --profile pa

Dataset-style folder tree:

anychest-infer \
  --input-path /path/to/Dataset003_Full \
  --output-dir /path/to/output \
  --checkpoint ./weights/anychest_inference_bundle.pt \
  --profile oblique

Example Cases

The repository includes example input/output cases derived from Dataset003_Full/test_case:

• examples/inputs/
• examples/outputs/
• examples/example_cases.json
• examples/README.md

To regenerate the bundled examples after downloading the checkpoint bundle:

PYTHON_BIN=python \
CHECKPOINT_PATH=./weights/anychest_inference_bundle.pt \
bash scripts/build_example_cases.sh

Inputs

The public CLI accepts:

• .png
• .jpg
• .jpeg
• .dcm
• .dicom

--input-path can point to:

• a single file
• a flat folder of supported files
• a dataset-style folder containing view subfolders such as LA/imagesTr or 45/imagesTr

If the input is flat or ambiguous, pass --view-name.

Outputs

For each processed image the package writes:

• per-class segmentation masks under labelsTr/
• a color overlay under overlays/

Mask layout is controlled by:

• --save-option one|sep|total
• --save-format img|npy

Reference JSON

The packaged reference JSON is cxas/data/anychest_reference.json.

It records:

• the 54 output classes in inference order
• the AnyChest view-to-angle mapping
• the original Dataset003_Full folder names such as 225, 45, LA, and PA

Reproducing Evaluation

The repository includes lightweight scripts for reproducing the public test-case evaluation workflow:

• scripts/evaluate_test_case.py
• scripts/reproduce_main_eval.sh

Example:

PYTHON_BIN=python \
CHECKPOINT_PATH=./weights/anychest_inference_bundle.pt \
DATASET_DIR=/path/to/Dataset003_Full/test_case \
bash scripts/reproduce_main_eval.sh

This generates:

• prediction folders
• per_image_dice.csv
• per_class_summary.csv
• summary.json

Notes

• This release is inference-only. Training scripts and trainer code are intentionally excluded.
• The bundled reference JSON replaces the need for a local dataset.json during public inference.
• The merged HF bundle is slimmed to inference weights only; optimizer and scheduler state are removed.