Deep Learning for Automated Diagnosis of Temporomandibular Joint Degenerative Disease Using MRI

Objectives The purpose of this study is to create and evaluate a deep learning (DL) algorithm to automatically detect temporomandibular joint degenerative joint disease (TMJDJD) using magnetic resonance imaging (MRI) to reduce dependence on conical beam computed tomography (CBCT) and reduce patients' radiation exposure. Methods This study conducted a retrospective analysis on 104 patients who had undergone both MRI of the temporomandibular joint (TMJ) and CBCT examinations. We obtained a total of 1769 sagittal MRI images and 1448 coronal MRI images. After cropping, standardization processing, and manual annotation of the condyle, these images were divided into multiple groups based on different image features. We developed three deep learning frameworks: ResNet101, DenseNet, and MobileNet V2, to automatically diagnose TMJDJD in different MRI image categories. These diagnostic performers were evaluated using accuracy, precision, recall, F1 score, Matthew correlation coefficient (MCC), and classification speed, and compared with the explanatory results obtained from professional radiologists. Results DenseNet performed the best in the Sg group, achieving an accuracy rate of 80.11%, precision of 80.26%, recall rate of 75.31%, and MCC of 0.60. Its classification performance surpassed that of ResNet101, MobileNet V2, and human experts. The t-SNE image visualization and training curve analysis confirmed that DenseNet has a better feature extraction ability and a stable training process. Moreover, the grouping of MRI images did not improve the model's accuracy. Conclusion The DL model can automatically diagnose TMJDJD in MRI images as an extremely promising AI-assisted tool. It can reduce the reliance of physicians and patients on CBCT and minimize radiation exposure.