Fracture Localization
Finds where a bone is broken on an X‑ray and draws a box around it, using a YOLOv8 detector trained on FracAtlas.
- Client
- Self directed
- Duration
- 4 days
- Stack
- Python, Ultralytics YOLOv8, PyTorch, Jupyter

Knowing that a bone is fractured is half the job; a clinician also needs to know where. This project trains an object detector that boxes the fracture on the image.
It uses the FracAtlas dataset, which ships YOLO, COCO and VGG annotations, and a small YOLOv8 model so inference stays fast enough for interactive use. The longer term goal is to label unannotated scans with weak supervision.
How it works
Prepare the dataset
FracAtlas images and YOLO format labels are split into train, validation and test folders and described in a one class config (`fracture`).
Train a small detector
YOLOv8 nano, 3 million parameters, trained for 50 epochs on Apple Silicon in about an hour.
Measure it
The best checkpoint is validated on 403 held out images with 92 marked fractures, reporting precision, recall and mAP.