LUS-classification

Model Card

Overview

This repository provides the trained models from:

Hierarchy-Aware and Anatomy-Guided Learning for Lung Ultrasound Video Classification

• Paper: https://arxiv.org/abs/2607.17551
• Code: https://github.com/Alya-Almsouti/LUS-video-classification

The models classify lung ultrasound videos into four classes:

1. Healthy
2. B-lines
3. Consolidations
4. B-lines with consolidations

Available Models

The implementation uses zero-based block indices. Therefore, code blocks 7 and 11 correspond to Blocks 8 and 12 in the paper.

Model Files

Each model is distributed as a compressed .zip archive containing checkpoints and training outputs for all five patient-level cross-validation folds.

A typical archive contains:

model-name/
├── config.json
├── final_metrics.json
├── fold_0/
│   ├── best.pt
│   └── best_val_metrics.json
├── fold_1/
├── fold_2/
├── fold_3/
└── fold_4/

Use best.pt as the selected checkpoint for each fold.

Intended Use

These models are intended for research on lung ultrasound video classification and anatomy-guided deep learning.

They are not validated for clinical diagnosis or independent medical decision-making.

Citation

Please cite the associated publication when using these models:

@article{almsouti2026hierarchy,
  title   = {Hierarchy-Aware and Anatomy-Guided Learning for Lung Ultrasound Video Classification},
  author  = {Almsouti, Alya and Mecharbat, Lotfi and Aboukhater, Noha and Alabrach, Yousef and Anwar, Siddiq and Kumar, Andre and Almakky, Ibrahim and Yaqub, Mohammad},
  journal = {arXiv preprint arXiv:2607.17551},
  year    = {2026},
  doi     = {10.48550/arXiv.2607.17551}
}

Contact

For questions, problems, or reproducibility issues, please open an issue in the GitHub repository.