Automatic classification of temporomandibular joint disorders by magnetic resonance imaging and convolutional neural networks

Background/purpose: In this study, we utilized magnetic resonance imaging data of the temporomandibular joint, collected from the Division of Oral and Maxillofacial Surgery at Taipei Veterans General Hospital. Our research focuses on the classification and severity analysis of temporomandibular joint disease using convolutional neural networks. Materials and methods: In gray-scale image series, the most critical features often lie within the articular disc cartilage, situated at the junction of the temporal bone and the condyles. To identify this region efficiently, we harnessed the power of You Only Look Once deep learning technology. This technology allowed us to pinpoint and crop the articular disc cartilage area. Subsequently, we processed the image by converting it into the HSV format, eliminating surrounding noise, and storing essential image information in the V value. To simplify age and left-right ear information, we employed linear discriminant analysis and condensed this data into the S and H values. Results: We developed the convolutional neural network with six categories to identify severe stages in patients with temporomandibular joint (TMJ) disease. Our model achieved an impressive prediction accuracy of 84.73%. Conclusion: This technology has the potential to significantly reduce the time required for clinical imaging diagnosis, ultimately improving the quality of patient care. Furthermore, it can aid clinical specialists by automating the identification of TMJ disorders.