SAM-KDNet: A Segmentation and Knowledge Distillation Framework for Automated CVM Stages Classification from CBCT

Objective . Automated skeletal maturity assessment is essential in orthodontic diagnosis and treatment planning, guiding the optimal timing of interventions such as myofunctional appliances and orthognathic surgery. This study aims to develop, test, and validate automated interpretable deep learning algorithms for the assessment and classification of cervical vertebrae maturation (CVM) stages from cone beam computed tomography (CBCT) scans. Methods. The sample consisted of 364 CBCT scans of orthodontic patients from private practices in the midwestern United States. The CVM stages were classified by two orthodontists and an oral and maxillofacial radiologist. A deep learning pipeline was designed that incorporated four innovations: (1) prompt-guided segmentation using SAM 1 for anatomical region localization, (2) demographic-aware modeling with age embeddings and gender conditioning, (3) knowledge distillation from a teacher network pretrained on a larger skeletal age dataset, and (4) tri-modal input fusion to emphasize clinically salient regions. Results. The best-performing model (88.54% accuracy, 5-fold cross-validation average) was the ResNet50 architecture with knowledge distillation, trained on a subset of 96 samples using our previously developed spheno-occipital synchondrosis staging model as the teacher network. Conclusion. Our findings demonstrate that combining anatomical cues, patient-specific information, and knowledge distillation enhances deep learning–based CVM staging and enables the models to effectively address the challenges of CBCT-based skeletal maturity assessment in orthodontics. This approach provides a foundation for more accurate and clinically meaningful integration of CBCT into orthodontic diagnostic workflows.