Diagnosing Knee Osteoarthritis Using Artificial Neural Networks and Deep Learning

Among various medical diagnostic tests performed to identify osteoarthritis in the knee, most of them are invasive and expensive. Therefore, in this study, another methodology for diagnosing osteoarthritis in the knee in a more qwick, non-invasive and cheap manner was proposed. For that purpose, surface electromyography signals recorded from the four muscles surrounding the knee, the recording of the flexion degree in the knee and pattern recognition algorithms were used. The datasets of this experiment comprised 22 subjects among whom 11 subjects had normal knee and other 11 Subjects had an osteoarthritis-affected knee. The total sample size was 1, 048, 576 samples and were processed using segments of overlapping-windows of 5000 samples. Time-series features were then extracted from each segment and were used to train, test and validate 7 different learning classifiers and 7 variants of deep learning networks. In this study, the best performance measure of 99.5% was achieved by multilayer perceptron. Quadratic support vector machine and complex tree performed as well with accuracy of 99.4% and 98.3% respectively. In contrast, the use of deep learning networks which were investigated over a wide range of hidden size of the sparse autoencoders, showed accuracy of 86.6% with final softmax layer and accuracy of 91.3% by replacing the final softmax layer with k-nearest neighbour. By comparison, artificial neural networks outperformed deep learning networks and it is therefore concluded that the knee pathology can be diagnosed more effeciently and automatically using surface electromyography signals and artificial neural network algorithms.