Case study
AutoKnee — knee X-ray processing
NECTEC web tool that crops knee X-rays in AP and LP views, finds femur and tibia contours, and runs a U-Net segmentation model.
FIG.1Research
- Role
- Main developer (2020) — 17 of 18 commits
- Stack
- PythonFlaskTensorFlowKerasOpenCVReactKonva
- Result
- Contours come from the model instead of hand tracing.
Problem
Measuring knee X-rays by hand is slow and varies between readers.
What I owned
Main developer (2020) — 17 of 18 commits
What I built
- Built AP/LP cropping and contour calculation.
- Integrated the U-Net model behind the API.
- Built the React + Konva viewer.
System shape
Flask API (JWT, SQLAlchemy) running OpenCV contouring and a TensorFlow/Keras U-Net, with DICOM input. React + Redux client with a Konva canvas viewer.
What changed
Contours come from the model instead of hand tracing.