Knee Implant Identification by Fine-Tuning Deep Learning Models.

Area of Science:

• Orthopedic surgery
• Medical imaging analysis
• Artificial intelligence in healthcare

Background:

• Identifying knee arthroplasty implant models pre-operatively for revision surgery is challenging and causes delays.
• Failure to identify implants promptly increases surgical complexity.
• Deep learning shows potential for improving diagnostic accuracy in medicine.

Purpose of the Study:

• To develop an automated deep learning solution for identifying the make and model of knee arthroplasty prostheses.
• To optimize the process of pre-operative planning for revision knee surgeries.

Main Methods:

• Deep learning algorithms were employed to classify 6 different knee arthroplasty implant models using 1078 radiographs (AP and lateral views).
• Model performance was evaluated using accuracy, sensitivity, and Area Under the Receiver-Operating Characteristic Curve (AUC).
• Saliency maps were utilized for visualization and comparative analysis.

Main Results:

• The best-performing deep learning model achieved 96.38% accuracy, 97.2% sensitivity, and an AUC of 0.985 on an external test set of 162 radiographs.
• The model successfully and uniquely identified implant models, with results visualized through saliency maps.
• Performance was compared against multiple trained models.

Conclusions:

• Deep learning models are effective in differentiating between 6 knee arthroplasty implant models.
• Saliency maps enhance understanding of the model's predictive focus, aiding result interpretation.