Knee arthritis detection
Teaching an image model to sort knee X-rays by how severe the arthritis looks.
- Status
- Research experiment
- Year
- 2024
- Role
- Solo
- Area
- Machine learning
In plain English
Doctors grade knee arthritis on X-rays on a scale from normal to severe. This project asked how well a general-purpose image model could learn that grading from a labeled set of knee images. It is a learning experiment and is not meant for diagnosis.
How it works
- Start with knee X-rays labeled in five grades, from normal through doubtful, mild and moderate to severe.
- Balance the grades with augmentation (making extra training images by zooming, rotating and adding noise).
- Adapt ResNet-50 (a well-known pretrained image model) after an earlier MedViT model gave weaker accuracy.
- Check results with a confusion matrix (a table showing which grades get mixed up).
What I did
I built it solo in PyTorch: the data pipeline, the augmentation and the model training. The pretrained ResNet-50 model and the X-ray dataset come from others.
What I learned
Transfer learning (reusing a pretrained model) gets you a working classifier quickly. Proving one would hold up in a real clinic is a far bigger study, and I now plan for that difference from the start.