Wrist AI

Wrist AI

Bao Do MD

Wrist AI predicts fracture on pediatric wrist X-rays and is developed as an educational tool for hands-on exploration of a fracture AI model. Although AI for fracture detection has transitioned from academic research to clinical implementation and has potential to reduce wait time to intervention and improve accuracy, significant challenges remain including algorithm performance, detection scope, workflow integration, training, legal liability, and economic justification.

Wrist AI was trained on 18,294 cases and tested on 2,033 hold out images. Study-level sensitivity was 93.1%, precision 95.4%, FP rate 2.9%, F1 94.2%, specificity 92.1%, and accuracy 92.7%; at the per detection level, it achieved 88.9% sensitivity, 10.9% FP, 88.0% precision, and 88.4% F1. Wrist AI is made possible by the publicly available dataset from Nagy et al's outstanding work in "A pediatric wrist trauma X-ray dataset (GRAZPEDWRI-DX) for machine learning."

Try cases below, on the web, or your own DICOM to see how an object based fracture AI model works. Test yourself: 1 of the cases below is normal. 4 are fractures. Would AI have changed your diagnosis ? Alter images to explore how factors such as rotation, exposure, or artifacts might affect performance. Would an adult x-ray work ?

Drag and drop next images without reloading.

Education or research, not for clinical use.

Sample cases: [1], [2], [3], [4], [5]

Model loading

Wrist AI operates locally on your device. Images are not uploaded, remaining strictly on your computer to ensure complete data privacy. Each image takes ~20 seconds to analyze on a modern PC or Mac.