Automated Detection of Shoulder Arthroplasty in X-rays Using Machine Learning

Area of Science:

• Orthopedic surgery
• Medical imaging
• Machine learning

Background:

• Rising demand for shoulder arthroplasty due to an aging, active population.
• Joint registries are crucial for tracking long-term outcomes but face data encoding errors.
• Current classification methods rely on non-medical encoders, introducing inaccuracies.

Purpose of the Study:

• To investigate machine learning (ML) for classifying shoulder arthroplasty techniques from postoperative X-rays.
• To assess the performance of neural network models in this classification task.
• To improve the accuracy of data within joint registries for shoulder arthroplasty.

Main Methods:

• Utilized a balanced dataset of 1000 samples from the Scottish Arthroplasty Project.
• Trained four common neural network models using 10-fold cross-validation.
• Classified four broad categories of shoulder arthroplasty techniques from X-ray images.

Main Results:

• The InceptionV3 model achieved the highest overall accuracy (93.85%) after cross-validation.
• The EfficientNet model demonstrated the highest individual classifier accuracy (99%).
• These results indicate significant potential for ML in improving data accuracy.

Conclusions:

• Machine learning shows strong potential to enhance the accuracy of joint registry data encoding for shoulder arthroplasty.
• More accurate data can facilitate evidence-based improvements in implant design and surgical techniques.
• Improved data accuracy will lead to better insights into implant survival and revision rates.