A new machine learning based approach for temporomandibular joint disorders

The temporomandibular joint disorders, which can affect basic functions such as opening, closing, and chewing of the jaw, are a common ailment. The etiology of temporomandibular joint disorders remains incompletely elucidated; nevertheless, they are postulated to be linked with a range of risk factors. Presently, diagnosis is typically performed through clinical examination by a dentist or maxillofacial surgeon, without the use of an automated system. Within the scope of this thesis, an original dataset composed of MR scans has been created. This dataset was prepared in collaboration with Ankara University Faculty of Dentistry and obtained ethical approval. Within the dataset, study has been conducted on four different diseases: anterior disc displacement with reduction, anterior disc displacement without reduction, increased synovial fluid, and condylar bone degeneration. In this context, convolutional neural network models were employed for disease diagnosis. Popular deep learning architectures, including GoogleNet, VGG, Xception, ResNet, MobileNet, Inception, DenseNet, ConvNeXt, and the ViT approach, were applied. The performance of the methods was examined using various metrics, and the consistency between the models was interpreted using the ICC value. As a result of the experiments, popular deep learning approaches yielded the best results, with acceptable consistency observed among the models. On average, an accuracy rate of 80% or higher was achieved. Furthermore, when comparing the areas marked by dentists on Grad-CAM images, it was observed that the networks generally focused on regions related to the disease. Moreover, computer-based approaches on TMJ disorders were addressed, and TMD publications were individually elucidated. Thus, the aim was to enhance the understanding of studies in this field and provide guidance for future implementations.