Deep learning-based identification and maturation assessment of the zygomaticomaxillary suture in cone-beam computed tomography images.

OBJECTIVE

This study addresses maxillary deficiency management by developing a deep learning-based system for zygomaticomaxillary suture assessment. As the maturation status serves as a pivotal indicator for optimal maxillary protraction timing, its accurate evaluation is crucial for treatment success. The proposed cone-beam computed tomography-based system overcomes current limitations of subjective manual staging, offering quantitative and reproducible maturation analysis to guide clinically critical decisions in maxillary protraction therapy.

METHODS

A dataset of 505 cone-beam computed tomography scans was collected and annotated by three orthodontists to locate and stage the suture. Data augmentation and various loss functions were applied to optimize the model. The framework employs dual-network architecture: (1) 3D U-Net for precise suture identification, followed by (2) 3D ResNet50-based classification network for maturation stage prediction. Model performance was validated using multi-dimensional metrics including classification accuracy, precision, recall, F1-score, and area under the curve values from the receiver operating characteristic curve, with benchmark comparisons against expert evaluators' assessments.

RESULTS

The model assessment took about 46 s for two tasks. 3D U-Net showed an average Euclidean distance of 2.31 ± 1.57 mm between the located and annotated zygomaticomaxillary suture centers. For classification tasks, the two-stage (ABC/DE) task had an accuracy of 0.922 and an F1-score of 0.921; the three-stage (A/BC/DE) task had an accuracy of 0.833 and an F1-score of 0.838; and the five-stage (A/B/C/D/E) task had an accuracy of 0.611 and an F1-score of 0.604.

CONCLUSION

The dual-network deep learning system can automatically locate and assess the maturation of the zygomaticomaxillary suture in three dimensions from CBCT files, providing a critical foundation for clinical decision-making in maxillary protraction therapy. This deep learning-model improves diagnostic workflow by reducing assessment time and minimizing inter-observer variability, providing clinicians with a reliable and reproducible reference for optimal maxillary protraction timing-avoiding over-treatment of the very young (stage A) while ensuring timely intervention before sutural closure (stages D and E).