A Combined Deep Learning Model with Attention Mechanism for Detection of Implant Manufacturer Using X-Ray Images

Abstract

Shoulder replacement surgery is one of the invasive techniques in orthopedic disciplines that replaces dead shoulder joints with prostheses made from polyethylene and metal components. To perform the surgery, there is a need to know the implant accessories and manufacturer of the implant. Some problems arise in a situation where the patient experiences pain and shoulder malfunctions that need replacement, and the accessories and implant manufacturer are mysterious to the doctor or the patient. In such a case, the solution to the problem depends on the accuracy of the identification of the manufacturer of the prosthesis. This research study proposes a novel detection and classification approach that integrates models based on deep learning with an attention mechanism to identify the implant manufacturers prior to surgery. The ensemble deep learning model utilizes the more sophisticated Long Short Term Memory architecture (LSTM) and the traditional multi-layer Convolution Neural Networks for extraction of features and predicting the implant manufacturer. The model employs an attention mechanism to focus on the critical part of the prosthesis that is crucial in the detection of the prosthesis manufacturer. The features map from the attention layer is finally fed into the LSTM for prediction by the implant manufacturer. Collection implant images of 597 from different implant manufacturers, which include 294 images generated by the Deputy manufacturer, 83 images generated by the Cofield manufacturers, 149 generated by the Zimmer manufacturer, and 71generated by the Tornier manufacturer, are utilized as a dataset not only for training but also for testing the model. The results show that the combined Deep Learning (DL) model with attention mechanism performs better than the Convolution Neural Network model, Convolution Neural Network +Attention, and Convolution Neural Network + LSTM models. Depending on the accomplishment of the model, it is concluded that this model could become an important tool for planning the preoperative procedure and this can be implemented for identifying and classifying the implants from different manufacturers.