hrnet-cephalometric-landmark-detection

HRNet Cephalometric Landmark Detection

This model performs automatic detection of 19 anatomical landmarks in lateral cephalometric radiographs using HRNet-W32 architecture.

🦷 Model Description

• Architecture: HRNet-W32 (High-Resolution Network)
• Task: 19-point cephalometric landmark detection
• Dataset: ISBI Lateral Cephalograms
• Input Size: 768×768 pixels
• Output: 19 landmark coordinates (x, y)
• Model Size: 331.1 MB

📍 Landmarks Detected

1. Sella turcica - Center of pituitary fossa
2. Nasion - Frontonasal suture
3. Orbitale - Lowest point of orbital cavity
4. Porion - Highest point of acoustic meatus
5. Subspinale (Point A) - Deepest midline point on maxilla
6. Supramentale (Point B) - Deepest midline point on mandible
7. Pogonion - Most prominent midline point of chin
8. Menton - Lowest point of mandibular symphysis
9. Gnathion - Midpoint between Pogonion and Menton
10. Gonion - Corner of the jaw angle
11. Lower Incisor Tip - Tip of lower central incisor
12. Upper Incisor Tip - Tip of upper central incisor
13. Upper Lip - Most prominent point of upper lip
14. Lower Lip - Most prominent point of lower lip
15. Subnasale - Junction between nose and upper lip
16. Soft Tissue Pogonion - Most prominent point of chin in profile
17. Posterior Nasal Spine - Tip of posterior nasal spine
18. Anterior Nasal Spine - Tip of anterior nasal spine
19. Articulare - Junction of temporal bone and mandible

🚀 Usage

Quick Start with Streamlit

import streamlit as st
import torch
from huggingface_hub import hf_hub_download

# Download model
@st.cache_resource
def load_model():
    model_path = hf_hub_download(
        repo_id="cwlachap/hrnet-cephalometric-landmark-detection",
        filename="best_model.pth"
    )
    
    # Load your HRNet model here
    model = get_hrnet_w32(config)
    checkpoint = torch.load(model_path, map_location='cpu')
    model.load_state_dict(checkpoint['model_state_dict'])
    model.eval()
    return model

model = load_model()

Python API

import torch
from huggingface_hub import hf_hub_download

# Download model
model_path = hf_hub_download(
    repo_id="cwlachap/hrnet-cephalometric-landmark-detection",
    filename="best_model.pth"
)

# Load model
model = get_hrnet_w32(config)
checkpoint = torch.load(model_path, map_location='cpu')
model.load_state_dict(checkpoint['model_state_dict'])
model.eval()

# Perform inference
with torch.no_grad():
    landmarks = model(input_image)

📊 Performance

• Mean Radial Error (MRE): ~1.2-1.6mm
• Successful Detection Rate (SDR@2mm): ~80-85%
• Successful Detection Rate ([email protected]): ~88-92%
• Training Time: ~15-20 hours on RTX 4070 Ti SUPER

🏥 Applications

• Orthodontic Treatment Planning: Automated cephalometric analysis
• Research: Large-scale cephalometric studies
• Education: Teaching cephalometric landmark identification
• Clinical Decision Support: Assisting radiological assessment

⚠️ Limitations

• Designed for lateral cephalometric radiographs only
• Performance may vary on images with different acquisition parameters
• Intended for research and educational purposes
• Clinical use requires validation by qualified professionals

📝 Citation

If you use this model in your research, please cite:

@misc{hrnet-cephalometric-2024,
  title={HRNet for Cephalometric Landmark Detection},
  author={cwlachap},
  year={2024},
  url={https://huggingface.co/cwlachap/hrnet-cephalometric-landmark-detection}
}

📄 License

This model is released under the MIT License, making it free for both academic and commercial use.

🤝 Contributing

This is an open-source project! Contributions, issues, and feature requests are welcome.

• Repository: [GitHub Repository URL]
• Issues: [GitHub Issues URL]
• Discussions: Use the Community tab above

🙏 Acknowledgments

• ISBI Challenge for providing the cephalometric dataset
• HRNet authors for the excellent architecture
• The medical imaging community for advancing automated analysis techniques

---

Built with ❤️ for the medical imaging community