TKA-AID: An Uncertainty-Aware Deep Learning Classifier to Identify Total Knee Arthroplasty Implants

Area of Science:

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

Background:

• Increasing primary total knee arthroplasties (TKAs) necessitate more revision TKAs.
• Accurate preoperative planning for revision TKA relies on timely implant identification.
• Current implant identification methods are time-consuming and challenging for surgeons.

Purpose of the Study:

• To develop a deep learning algorithm for automated identification of common primary TKA implant models.
• To improve the efficiency and accuracy of implant identification in preoperative planning.

Main Methods:

• Utilized a total joint registry with 9,651 patients and 111,519 images.
• Trained an EfficientNet-based deep learning model to classify nine TKA systems across radiographic views.
• Implemented conformal prediction for uncertainty estimates and an outlier detection system.

Main Results:

• Achieved 99.7% average accuracy on an internal test set.
• The outlier detection system identified 93% of manual outliers.
• Demonstrated high performance on an external test set with one error in 301 images.
• The model processes approximately 30 images per second.

Conclusions:

• Developed an automated tool for classifying nine TKA implant designs.
• The tool is effective across multiple radiographic views.
• Incorporated uncertainty quantification and outlier detection as safety mechanisms for clinical use.