An Ensemble Voting Approach for Shoulder Implant Classification from X-Ray Images

Total Shoulder Arthroplasty (TSA) is an effective method that involves replacing damaged joint surfaces with appropriate prosthetic components to manage pain and improve joint mobility. Over time, the prosthetic replacement or improvement process requires both the surgeon and the patient to know the prosthesis manufacturer for successful outcomes. However, inadequate medical records often necessitate additional X-rays for the patient. In this paper, a deep learning-based method is proposed to analyze prosthesis details from X-ray images automatically. The method combines predictions made by pre-trained ViT, DeiT, and Swin model versions on a dataset consisting of 597 shoulder implant X-ray images to achieve more accurate classification. The performances of hard and soft voting ensembles were analyzed respectively. During the experiments, the soft voting approach yielded the highest performance, achieving 96.15 % accuracy, 92.12 % precision, and 88.56 % recall, marking the highest classification performance observed thus far. The results demonstrate that our voting ensemble method can be utilized as a reliable and effective tool for automatically analyzing prosthesis details.