Automated multi-orientation (plane-wise) segmentation of TMJ structures using deep learning: a dual-center study.

OBJECTIVE

To develop and validate a deep learning framework for automated multi-view (across planes) MRI segmentation of temporomandibular joint (TMJ) structures, reducing subjectivity and improving efficiency, and supporting future quantitative analyses (e.g., geometry/volume measurements) pending dedicated clinical validation.

STUDY

DESIGN

Two-center MRI data from 121 (242 TMJs) included oblique coronal/sagittal sequences (closed/open-mouth). Four models (2D U-Net, 3D U-Net, nnU-Net, Swin-Unet) were trained on 65 patients and tested on two independent sets of 28 patients each for segmenting the temporal bone articular surface, articular disc, and condyle. Performance was evaluated using Dice similarity coefficient, intraclass correlation coefficients (ICCs), and Wilcoxon tests.

RESULTS

All models showed high agreement with manual annotations (ICC > 0.90, P < .05). nnU-Net achieved superior performance (ICC > 0.99 across structures), with condyle Dice = 0.949 ± 0.020 (95% CI: 0.942-0.956) and disc Dice = 0.865 ± 0.042 (95% CI: 0.850-0.880). Performance was consistent across imaging planes (error<5%).

CONCLUSIONS

The nnU-Net-based framework enables accurate, computationally feasible, and robust multi-view (across-plane) TMJ segmentation. Dual-center evaluation across institutions with similar protocols supports technical reproducibility, while broader generalizability and clinical utility require further multi-scanner and outcome-based validation.