CheXmask-U: Uncertainty-aware landmark-based anatomical segmentation for chest X-rays

📄 Paper | 💻 Code | 🎛️ Interactive Demo | 📦 Dataset

---

Model Description

CheXmask-U is a landmark-based chest X-ray segmentation model providing node-wise uncertainty estimation. It outputs:

• Anatomical landmarks for lung and heart structures
• Latent uncertainty from the learned variational latent space
• Predictive uncertainty via stochastic output sampling

The model is implemented using a hybrid graph-convolutional architecture (HybridGNet), combining convolutional encoders with graph-based decoders. For full usage and code, see the GitHub repository.

---

Usage

from models.HybridGNet2IGSC import HybridGNetHF

device = "cuda"  # or "cpu"
model = HybridGNetHF.from_pretrained(
    "mcosarinsky/CheXmask-U",
    subfolder="v1_skip",
    device=device
)

# xray_image: tensor or suitable input
landmarks, _, _ = model(xray_image)

Citation

If you use this model, please cite:

@misc{cosarinsky2025chexmaskuquantifyinguncertaintylandmarkbased,
      title={CheXmask-U: Quantifying uncertainty in landmark-based anatomical segmentation for X-ray images}, 
      author={Matias Cosarinsky and Nicolas Gaggion and Rodrigo Echeveste and Enzo Ferrante},
      year={2025},
      eprint={2512.10715},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2512.10715}, 
}