Practical computer vision application to detect hip fractures on pelvic X-rays: a bi-institutional study

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Orthopedic Surgery

Background:

• Pelvic X-ray (PXR) is crucial for diagnosing hip fractures.
• Radiologist availability and image quality can impact PXR diagnostic accuracy.
• Computer vision algorithms offer potential for improved fracture detection.

Purpose of the Study:

• To validate a computer vision algorithm for detecting hip fractures on PXRs.
• To assess the algorithm's performance across diverse patient populations.
• To develop a web application for bedside adoption of the algorithm.

Main Methods:

• A convolutional neural network (Xception) was trained on 4235 PXRs.
• Image augmentation and gradient-weighted class activation mapping were used.
• Validation was performed on 500 PXRs from two institutions.

Main Results:

• The algorithm achieved high performance metrics (AUC 0.98 and 0.97) in validation cohorts.
• Sensitivity, specificity, accuracy, and F1 score exceeded 0.90 for both cohorts.
• The web application provides fracture probability and heatmap localization.

Conclusions:

• A high-performing computer vision algorithm for hip fracture detection on PXRs was developed and validated.
• A web application facilitates bedside use of the algorithm.
• Further studies are needed to confirm clinical validity and utility.