Detecting upper extremity native joint dislocations using deep learning: A multicenter study.

Area of Science:

• Orthopedic imaging
• Artificial intelligence in medicine
• Radiology

Background:

• Joint dislocations are orthopedic emergencies requiring prompt reduction.
• Diagnostic delays in identifying dislocations can complicate treatment.
• Automated detection could improve patient care timelines.

Purpose of the Study:

• To develop and test convolutional neural networks (CNNs) for detecting elbow and shoulder dislocations.
• To evaluate the generalizability of these CNNs on external datasets.
• To visualize CNN decision-making processes using class activation maps (CAMs).

Main Methods:

• Trained CNNs on 106 elbow and 140 shoulder radiographs from a trauma center.
• Utilized 24x data augmentation for training and validation.
• Evaluated model performance on internal and external datasets, including those from an external hospital and online repositories.
• Generated CAMs to interpret CNN focus areas.

Main Results:

• CNNs achieved Area Under the Curve (AUC) values greater than 0.99 on internal test sets.
• CNNs achieved AUCs greater than 0.97 on external test sets.
• CAMs confirmed CNNs focused on relevant anatomical joint regions for accurate decision-making.

Conclusions:

• CNNs demonstrated high accuracy and excellent generalizability in identifying elbow and shoulder dislocations.
• The findings suggest CNNs can assist in the rapid diagnosis of joint dislocations.
• Automated detection by CNNs may expedite patient access to necessary orthopedic interventions.