Abstract: Background　Temporomandibular joint disorder (TMD) is a common oral and maxillofacial disease with an incidence ranging from 20% to 60%, and its prevalence has been increasing year by year. Accurate diagnosis of this condition remains a major clinical challenge.Objective　To develop a deep learning model based on magnetic resonance imaging (MRI) for the automatic identification of anterior temporomandibular joint disc displacement, so as to improve the efficiency of clinical image reading. Methods　A total of 249 patients diagnosed with anterior temporomandibular joint disc displacement via MRI from October 2020 to October 2024 at the Department of Stomatology, the First Medical Center of PLA General Hospital were retrospectively enrolled, yielding a total of 2 002 oblique sagittal MRI scans (both open-mouth and closed-mouth positions). Two specialists in temporomandibular joint disorders manually segmented and annotated the temporomandibular joint regions using Materialise Mimics Medical software, and all MRI images were grouped according to the presence or absence of disc displacement. A dense convolutional attention U-Net framework was established, including a segmentation model for temporomandibular joint discs and a classification model for anterior disc displacement. The segmentation performance was evaluated using the intersection over union (IoU) and Dice coefficient, while the classification performance was assessed via accuracy, precision and recall.Results　 For the deep learning segmentation model of temporomandibular joint discs, the mean IoU on the validation set was 0.912 ± 0.004 (95% CI: 0.906 - 0.918), and the mean Dice coefficient was 0.925 ± 0.005 (95% CI: 0.920 - 0.930). For the classification model of anterior disc displacement, the mean recall, precision, F1-score and area under the receiver operating characteristic curve (AUC) were 0.918 ± 0.017 (95% CI: 0.894 - 0.937), 0.819 ± 0.007 (95% CI: 0.810 - 0.833), 0.866 ± 0.009 (95% CI: 0.854 - 0.878) and 0.721 ± 0.015 (95% CI: 0.702 - 0.740), respectively. The mean true positive rate (TPR) was 0.916 and the mean false positive rate (FPR) was 0.465. Inter-group differences were statistically significant (P=0.015, P=0.033). Conclusion　The constructed deep learning model shows favorable segmentation and classification performance, enabling automatic identification of anterior disc displacement of the temporomandibular joint (binary classification: presence or absence of displacement). It provides a feasible tool for the intelligent auxiliary diagnosis of anterior disc displacement of the temporomandibular joint.