Deepmsm: Multimodal Deep Learning For Midpalatal Suture Maturation Staging

Accurate assessment of midpalatal suture (MPS) maturation is crucial for maxillary expansion treatment planning. This study developed a multimodal deep learning model (DeepMSM) integrating cone-beam computed tomography (CBCT) images, age, gender, cervical vertebral maturation (CVM), and mandibular third molar (MTM) calcification to predict MPS maturation stages, aiming to enhance prediction accuracy and clinical applicability. CBCT scans, lateral cephalograms, age, and gender were collected from patients aged 7–36 years. MPS, cervical vertebrae, and mandibular third molars were segmented and annotated using Angelieri’s (MPS), Baccetti’s (CVM), and Demirjian’s (MTM) methods. A CNN-based multimodal deep learning model was developed in collaboration with Aalborg University, incorporating data augmentation. Model performance was evaluated using multiclass metrics (accuracy, precision, recall, F1-score, AUC-ROC) and compared with clinicians’ assessments. MPS maturation stages showed statistically significant correlations with age, gender, CVM, and MTM stages. The multimodal model achieved 85% accuracy (F1-score: 84.31%), with high precision (89.16%) and recall (85.00%). Classes C (precision: 91%, recall: 100%) and E (precision: 76%, recall: 100%) exhibited outstanding performance, surpassing the average accuracy of less-experienced clinicians. DeepMSM demonstrates that integrating CBCT, CVM, MTM staging, age and gender significantly improves MPS maturation prediction. The model's superior accuracy enables personalized maxillary expansion treatment while enhancing clinicians' staging precision. This research advances intelligent orthodontic diagnostics, promoting precision oral healthcare through optimized staging and standardized clinical decision-making.