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

Abstract

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