Automated classification of total knee replacement prosthesis on plain film radiograph using a deep convolutional neural network

The identification of the make and model a total knee replacement (TKR) is necessary step prior to revision surgery for periprosthetic fracture, loosening, wear or infection. Current methods may fail correctly identify implant up 10% time. This study presents training Convolutional Neural Network (CNN) automatically seven TKR implants absence on plain-film radiographs. Our dataset consists 588 anteroposterior (AP) X-rays knee. They were randomly divided into train, validation testing sets with 50:25:25 split. A CNN based ResNet-18 architecture was trained best selected using results. final tested hold-out test dataset. network demonstrated perfect accuracy in classifying one eight labelled classes. Saliency maps outlines are key given prediction. Further research will benefit from larger datasets more complete coverage possible implants. ability recognize that outside networks distribution essential such an algorithm operating safely clinical practice. With these issues limitations addressed there potential could save clinicians time reduce instances where not identified pre-operatively, simplifying re-operative cases improving outcomes.