Automated Landmark Detection for Calcaneus.
Streamlit app for automated calcaneus landmark detection using YOLO and GU2Net deep learning. 94% success detection rate with clinical interface.
This is a machine learning research project focused on automated anatomical landmark detection for clinical applications. Using a hybrid architecture combining YOLO object detection and GU2Net segmentation, the system identifies key anatomical landmarks on foot radiographs with high precision. The application achieves a 94% Successful Detection Rate (SDR) with mean radial error of 2±2.05mm at 5mm margin. Built with Python and PyTorch, the model was trained on 3,000+ radiographs from the Foot and Ankle Innovation Lab at Harvard Medical School. Deployed via Streamlit, the application provides an interactive clinical interface for radiologists and surgeons. The web platform enables batch processing of radiographs, visualization of detected landmarks with confidence scores, and export of results for surgical planning. Over 100+ clinical users have access to the system for preliminary evaluations.
Problem
Manual landmark detection on radiographs is time-consuming and subject to inter-observer variability. Surgeons spend hours reviewing and annotating images, delaying treatment planning. Automation could reduce review time and improve consistency.
Solution
We developed a deep learning pipeline combining YOLO for object localization and GU2Net for precise boundary detection. The hybrid global/local CNN features achieve sub-pixel accuracy. Streamlit deployment makes the tool accessible to clinicians without ML expertise.
- 01Built YOLO-based GU2Net app for automated calcaneus landmark detection
- 02Reduced surgeon review time by 30% compared to manual annotation
- 94%Success Detection Rate
- 3,000+ radiographsTraining Data
frontend
- Streamlit
- Matplotlib / Plotly
backend
- Python
- PyTorch
- YOLOv8
- OpenCV
- Harvard Medical School Team