MRI-Based Deep Learning Image Classification for Screening Temporomandibular Joint Degenerative Joint Disease

Background: This study aimed to develop and evaluate a deep learning (DL) algorithm based on magnetic resonance imaging (MRI) for the classification of temporomandibular joint degenerative joint disease (TMJDJD), thereby exploring its potential role in assisting cone-beam computed tomography (CBCT) examinations and minimizing patient radiation exposure. Methods: A retrospective analysis was conducted on 104 patients who had undergone both MRI and CBCT examinations of the temporomandibular joint. A total of 1769 sagittal and 1448 coronal MRI images were collected. After image preprocessing, the datasets were grouped according to different imaging features. Three DL frameworks—ResNet101, DenseNet201, and MobileNetV2—were developed to classify TMJDJD using various MRI image categories. The classification performance of these models was evaluated using accuracy, precision, recall, F1 score, Matthews correlation coefficient (MCC), and the results were compared with classifications made by two senior clinical experts in a blinded image-level reader experiment. Results: DenseNet201 achieved the best performance in the sagittal (Sg) group, with an accuracy of 80.11%, precision of 80.26%, recall of 75.31%, and an MCC of 0.60. DenseNet201 achieved the best performance among the evaluated DL models, with higher overall performance metrics than ResNet101 and MobileNetV2. In a constrained image-level reader comparison, DenseNet201 demonstrated comparable overall agreement metrics to the two experts, although one expert achieved higher recall and a lower false-negative rate. t-Distributed Stochastic Neighbor Embedding (t-SNE) and training curve analyses confirmed that DenseNet201 demonstrated superior feature extraction capability and a more stable training process. Grouping MRI images, however, did not improve model accuracy. Conclusions: DenseNet201 showed preliminary potential for MRI-based TMJDJD classification, particularly on sagittal images. Based on image-level classification results, the model may provide preliminary support for identifying MRI findings associated with TMJDJD and assisting decisions regarding further CBCT evaluation.