Authors (including presenting author) :
Lau LCM(1), Chui ECS(1), HO KKW(1), ONG MTY(1), YUNG PSH(1)
Affiliation :
Department of Orthopaedics and Traumatology, The Faculty of Medicine, The Chinese University of Hong Kong, The Prince of Wales Hospital, Hong Kong SAR
Introduction :
Total knee arthroplasty (TKA) loosening is the leading cause of revision among various complications, and it tends to occur many years after the initial TKA. With the summative effect of longer life expectancy, late occurrence of loosening, and increasing number of patients living with TKA, how to follow up patients with TKA is becoming a major challenge in orthopedic field as early detection of loosening is potentially beneficial to the patients. A delay in diagnosis of loosening and hence a prolonged period of walking with an unstable implant can result in loss of bone stock and deterioration of surrounding soft tissues, which may entail a larger scale of revision surgery with poorer outcome. A system that can auto-detect loosening can lessen the burden of orthopedic surgeons potentially and further safeguard their practice.
Machine learning has been successfully applied in various medical field, for instance, automatic detection of strokes, retinopathies, and cancerous histology with same level of accuracy as the relevant field experts. Actualized by advanced computational power, machine learning can self-teach and self-develop its pattern recognition by reading a vast number of relevant labelled images and/or data and does not necessarily follow clinical criteria set by the medical experts.
Therefore, aim of this study was to build and evaluate an optimized image-based machine-learning model that could detect total knee arthroplasty loosening based on radiographs alone effectively. Additional clinical-information-based machine-learning models were developed and combined with image-based machine-learning model for further evaluation and compared.
Objectives :
aim of this study was to build and evaluate an optimized image-based machine-learning model that could detect total knee arthroplasty loosening based on radiographs alone effectively. Additional clinical-information-based machine-learning models were developed and combined with image-based machine-learning model for further evaluation and compared. Class activation heatmap was generated to represent machine-learning model focus of loosening detection during radiograph analysis and probability of loosening was generated.
Methodology :
Image-based machine-learning model was developed based on ImageNet which is an open-source project that could classify an Input Image into 1000 separate object categories. The model was trained on approximately 1.2 million Images, 50000 images for validation and 100,000 images for testing. Xception model, an extension of the Inception Architecture which replaced the standard Inception modules with Depthwise Separable Convolutions, was used. The development of a deep learning-based prosthesis loosening estimating system was based on Xception pre-trained model and a TKA patient X-ray image dataset. Random forest consists of a large amount of individual decision trees that operate as an ensemble. Each individual tree in the random forest generated a class prediction. The class with the most votes became our model’s prediction. The process of random forest will be shown. A classification system based on a dataset with TKA patient clinical parameters was developed using random forest classifier.
Result & Outcome :
Evaluation was run by a single model on a single crop of input images. 75% of X-ray images in the dataset were used as the test set and 25% of X-ray images in the dataset were used as validation set. We will only report findings on validation set. As a result, precision rate and recall rate were 0.95 and 0.96 respectively. Accuracy rate of 96.3% for visualization classification was observed. Clinical-information model (Random forest classifier) was implemented for estimating the occurrence of prosthesis loosening. It resulted in precision rate of 0.71 and recall rate of 0.20. The difference between a combined model of image-based and clinical-information-based model to image-based model alone was insignificant. It was observed that using X-ray images as input and deep learning for estimation could achieve greater precision and recall rates, thus a better estimation for prosthesis loosening. Examples of loosening prediction are shown in Figures.
