Detecting total hip arthroplasty dislocations using deep learning: clinical and Internet validation.

Area of Science:

• Orthopedic Surgery
• Artificial Intelligence
• Medical Imaging

Background:

• Periprosthetic dislocations of total hip arthroplasty (THA) are urgent injuries requiring prompt diagnosis and treatment.
• Delays in diagnosis and treatment of THA dislocations can complicate reduction and negatively impact patient outcomes.
• Automated radiographic triage systems could significantly reduce diagnostic delays for THA dislocations.

Purpose of the Study:

• To develop and evaluate convolutional neural networks (CNNs) for the automated detection of THA dislocations.
• To assess the generalizability of trained CNNs on external, independent datasets.
• To explore the potential clinical utility of CNNs in emergency department triage for THA dislocations.

Main Methods:

• A dataset of 357 THA radiographs was used to train and internally test various CNNs.
• External validation was performed using two independent datasets to evaluate model generalizability.
• CNN performance was quantified using the area under the receiver operating characteristic curve (AUROC).
• Class activation mapping (CAM) was employed to visualize regions of interest highlighted by the CNNs.

Main Results:

• Several CNN models achieved an AUROC of 1.0 on both internal and external test sets.
• The high AUROC values indicate excellent diagnostic performance and robust generalizability.
• CAM visualizations confirmed that CNNs consistently focused on the relevant anatomical structures of the THA.

Conclusions:

• CNNs demonstrate high diagnostic accuracy for identifying THA dislocations.
• The strong performance on external datasets supports the clinical utility of these AI models.
• Trained CNNs show significant potential for improving triage efficiency in emergency departments for THA dislocations.