CXAS
Chest X-ray anatomy segmentation (Mask2Former models: 187 frontal and 177 lateral classes) plus 132 named landmarks and landmark-based measurements such as cardiothoracic ratio, costophrenic angle, Cobb angle and the Haller index, in mm for DICOM input; cxas_idc writes one multi-segment DICOM SEG and a TID 1500 SR per radiograph; trained on PAXRay++ CT-projected radiographs
- What it is
- Segmentation (Segmentation & frameworks)
- Modality
- Radiology: chest radiographs (frontal and lateral)
- Maintainer
- Constantin Seibold et al.
- License
- MIT
- Commercial use
- Permissive
- Access
- Free, public source repository (see links below)
- Last update
- 2026-11-14 (last repository push, UTC; Maintained)
- Last checked
- 2026-11-16
- Limitations
- Listed for research and development reference. Hounsfield Lab has not tested its clinical performance; check the project’s own documentation for intended use and regulatory status.
Links
License review: Repo LICENSE.txt is MIT, and the README states MIT for code and model weights. Weights are public, not gated, on Hugging Face (cmseibold/cxas-models and cmseibold/cxr-landmarks, both tagged MIT) and are tag-pinned and SHA-256 checked on download. v0.1.0 (Nov 13, 2026) added the new models, landmarks and DICOM SEG/SR output; v0.1.3 (Nov 14) fixed landmark measurements under NumPy 2. As of Nov 19, PyPI lists 0.1.2 as the latest version; the 0.1.3 Docker images pin NumPy 1.26. The README states it is a research tool, not a medical device. Papers: BMVC 2022 (PAXRay) and arXiv 2306.03934.
Related tools in Segmentation & frameworks
- MONAI: PyTorch-based AI toolkit for healthcare imaging
- MONAI Label: AI-assisted image labeling and active learning; integrates with 3D Slicer and OHIF
- nnU-Net: Self-configuring segmentation framework
- TotalSegmentator: Segments more than 100 anatomical structures in CT and MR
- TotalSpineSeg: Two-stage nnU-Net tool that segments and individually labels every vertebra and intervertebral disc plus the spinal cord and canal, using an odd/even vertebra scheme and landmark discs (C2-C3, C7-T1, T12-L1, L5-S) to avoid fused or split vertebrae; works on localizers and short-FOV scans; peer-reviewed in European Radiology (Nov 7, 2026)
- NV-Segment-CT: CT segmentation built from MONAI VISTA3D: 117 structures automatically, or 132 promptable by point or class. Research use only, not for clinical care. Needs a CUDA GPU; weights are about 872 MB
Related public imaging datasets
- LIDC-IDRI: CT, CR, DX, Chest · CC BY 3.0
- MIDRC-RICORD (1A, 1B, 1C): CT, DX, CR, Chest · CC BY-NC 4.0
- NIH ChestX-ray14: DX, Chest · NIH Clinical Center terms (no use restrictions; attribution required)
- CheXpert Plus: DX, CR, Chest · Stanford AIMI research use agreement (non-commercial; commercial license sold separately)
- RSNA Pneumonia Detection Challenge: DX, CR, Chest · Kaggle competition rules (non-commercial)
- SIIM-ACR Pneumothorax Segmentation: DX, CR, Chest · Kaggle competition rules (non-commercial)
Browse all DICOM datasets with license and access details.