关键词: 颞下颌关节盘锚固术, 不可复性盘前移位, 多模态深度学习, 三维卷积神经网络, Transformer, 疗效评估

Abstract: Objective    To develop a model of 3D convolutional neural network（3D CNN）combined with a Transformer architecture（hybrid 3D CNN-Transformer model）and investigate its effectiveness in evaluating postoperative outcomes of temporomandibular joint（TMJ）disc anchorage. Methods    A retrospective analysis was conducted on 31 patients（62 TMJs）with bilateral anterior disc displacement without reduction who underwent TMJ disc anchorage at the Department of Oral and Maxillofacial Surgery，Hospital of Stomatology，China Medical University，from January 2024 to May 2025. Preoperative and 3‑month postoperative MRI data and clinical characteristics - including visual analog scale（VAS）scores for pain and maximum interincisal opening（MIO）- were collected for comprehensive therapeutic evaluation. The hybrid 3D CNN-Transformer model was constructed, including dual-branch feature extraction（MRI data and clinical features），multimodal fusion，and multitask output（category of evaluation being excellent，good and poor, and changes in VAS scores and MIO）. Model performance was evaluated using stratified five‑fold cross‑validation combined with an out-of-fold predictions（OOF）strategy. Classification metrics included accuracy，F1‑score，precision，recall rate，and area under the receiver operating characteristic curve（AUC）. Regression metrics included mean absolute error（MAE），root mean square error（RMSE），and coefficient of determination（R2）. Reliability metrics included confidence score，confidence distribution，and sample coverage rate. The correlation between the predicted probability for ＂excellent＂ therapeutic outcome［P（excellent）］and the decrease in VAS score，increase in MIO and the comprehensive clinical improvement index was further analyzed. Results    In the OOF samples，consisting of 31 patients，the overall accuracy was 0.936 （95%CI：0.786 - 0.982），overall precision was 0.944，and recall rate was 0.936. The weighted-average F1-score and macro-average F1-score were 0.937 and 0.936，respectively. The AUC values for the categories excellent，good，and poor were 1.00，0.97，and 0.96. For regression tasks，the prediction of changes in VAS achieved MAE = 0.803 points，RMSE = 0.977 points，R2 = 0.755；prediction of changes in MIO achieved MAE = 2.026 mm，RMSE = 2.412 mm，R2 = 0.665. Model confidence ranged from 0.441 to 0.980 with a median of 0.919；54.8%（17/31）of cases exhibited high confidence（> 0.9）. As the confidence threshold increased，sample coverage decreased while accuracy and macro‑average F1‑score increased correspondingly. P（excellent）was positively correlated with decrease in VAS score（r = 0.747，P < 0.001），increase in MIO（r = 0.813，P < 0.001），and the comprehensive clinical improvement index（r = 0.773，P < 0.001）. Conclusion     The hybrid 3D CNN-Transformer model shows very good accuracy and reliability in assessing postoperative conditions at 3 months after TMJ anchorage. The model reasonably represents the reconstruction of structures as well as improvements in pain and maxillary opening functions. It can also be applied in predicting treatment effects before surgeries. Nonetheless，optimization of the model is still needed.

Key words: temporomandibular joint disc anchorage, anterior disc displacement without reduction, multimodal deep learning, 3D convolutional neural network, Transformer, efficacy evaluation