Diagnosis of temporomandibular joint disorder using one-dimensional convolutional neural networks: A comparative study

Abstract

Temporomandibular Joint (TMJ) is a joint located on both sides of the cranium that connects the mandible to the cranium. Temporomandibular Joint Disorder (TMD) is generally defined as pathological conditions resulting from abnormal movement of the TMJ. Symptoms of TMD usually occur in the form of pain in the mandible and the muscles that control mandibular movement. One of the clinical diagnostic methods is the auscultation of sounds coming from the joint during the opening and closing of the mandible. In this study, previously recorded TMJ sounds were analyzed for TMD diagnosis using one-dimensional (1-D) convolutional neural networks (CNN), a sub-branch of deep learning algorithms. The obtained results were compared with the results of previous outcomes of studies which were using deep learning algorithms such as two-dimensional CNN, which is generally used for image processing, and LSTM network, which is widely used for time series analysis. Comparison results indicate that 1-D CNNs are less effective than image-based CNN algorithms. Results show that 1-D CNN classification of Type-1 algorithm and Type 2 algorithm are 75% and 65% accuracy respectively. These figures are significantly lower than the 94% accuracy achieved by the 2-D image-based algorithm. When compared to LSTM networks, which have an accuracy of 70%, 1-D CNNs yield comparable results. For a more comprehensive analysis, precision, recall, specificity, and F1-Score metrics were evaluated, and the findings were interpreted accordingly.