Deep Learning for Trauma Radiographs

Plain radiographs are the first-line imaging study for injured patients, but subtle fractures are easily missed in a busy emergency department. We develop deep learning models that detect and localize traumatic findings on plain films.

Our flagship work, PelviXNet, is a scalable, physician-level algorithm that detects universal trauma findings on pelvic radiographs, trained with weak point-level supervision on over 5,000 clinically annotated images (Nature Communications, 2021). Earlier work demonstrated automated detection and visualization of hip fractures on plain pelvic radiographs with accuracy comparable to experienced clinicians (European Radiology, 2019), recognized with a National Innovation Award.

The line of research has since extended to simultaneous rib and clavicle fracture detection on trauma chest radiographs (Trauma Surgery & Acute Care Open, 2024) and multi-view ankle fracture detection with ensemble strategies (British Journal of Radiology, 2023), with ongoing projects on trauma chest X-ray interpretation.