Automatic Detection of Perilunate and Lunate Dislocations on Wrist Radiographs Using Deep Learning

Area of Science:

• Orthopedic imaging
• Artificial intelligence in radiology
• Computer vision for medical diagnosis

Background:

• Delayed diagnosis of perilunate/lunate dislocations leads to poor patient outcomes.
• Computer vision offers potential for improved diagnostic accuracy in radiology.

Purpose of the Study:

• To develop and evaluate a deep learning algorithm for detecting perilunate and lunate dislocations on lateral wrist radiographs.

Main Methods:

• A dataset of 435 lateral wrist radiographs was used, with images partitioned into training and test sets.
• A two-stage deep learning model, including an object detector for lunate localization and a classifier for dislocation detection, was employed.
• Performance was assessed using accuracy, sensitivity, specificity, and receiver operating characteristic analysis.

Main Results:

• The lunate object detector achieved 97.0% accuracy.
• The dislocation classifier demonstrated 93.8% sensitivity and 93.3% specificity.
• The overall system achieved an area under the curve of 0.986.

Conclusions:

• A proof-of-concept computer vision system for diagnosing wrist dislocations was developed.
• This deep learning algorithm shows potential to enhance clinical sensitivity and reduce diagnostic delays.