Risk Prediction for Root Resorption of Lateral Incisor Induced by Canine Impaction based on Deep Learning

Objective This study aimed to develop and validate a Mask RCNN–Self-Attention (Mask RCNN-SA) framework for automated 3D segmentation of maxillary structures and to investigate volumetric biomarkers of lateral incisor RR associated with impacted maxillary canines. Methods CBCT scans from 42 patients with impacted maxillary canines were randomly divided into training, validation, and test sets (7:2:1) and used to train a Mask RCNN-SA network. Segmentation performance was evaluated using the Dice similarity coefficient (DSC), sensitivity, specificity, and 95% Hausdorff distance. Volumetric measurements of the maxilla, lateral incisors, and first premolars were derived from the segmentation masks to quantify RR-related volume loss. Associations among canine impaction, sex, RR grade, and volumetric changes were analyzed using logistic regression, hierarchical regression, and support vector machine (SVM) classification. Statistical significance was set at P  < 0.05. Results The Mask RCNN-SA achieved high segmentation accuracy across anatomical regions (DSC: 93.6% for maxilla; 92.5% for lateral incisor; sensitivity ≥ 94%; accuracy ≥ 97%). Automated segmentation required 0.08 min per scan, significantly faster than manual segmentation. Volumetric analysis revealed significantly reduced volumes on impacted sides for the maxilla (8.767 vs. 9.669 cm³, P  = 0.0047) and lateral incisors (3.028 vs. 3.705 cm³, P  = 0.0043). RR predominantly affected lateral incisors adjacent to impacted canines, with more pronounced root volume loss in females (21.1% smaller than males, P  = 0.0202). SVM classification achieved an AUC of 0.956 for RR prediction. Hierarchical regression identified lateral incisor RR as the strongest determinant of volume reduction (β = 141.244, P  < 0.001), whereas maxillary volume showed no added explanatory power. Conclusion The Mask RCNN-SA framework provides accurate, rapid, and clinically interpretable segmentation and volumetric assessment of maxillary structures in cases with impacted canines. This deep learning approach offers a clinical tool for early RR prediction and supports timely interceptive orthodontic interventions.