Risk Prediction of Root Resorption Induced by Impacted Canine based on Deep Learning (Preprint)

UNSTRUCTURED Root resorption (RR) of adjacent teeth secondary to impacted maxillary canines represents a significant orthodontic challenge, characterized by high morbidity and compromised treatment outcomes. This study introduced the Mask RCNN-SA network, which integrating volumetric features, to overcome traditional algorithmic limitations in anatomical modeling and establish an interpretable RR risk prediction model. Using 4200 CBCT slices augmented 42 CBCT patients, we trained the model for multiclass segmentation of the maxilla, lateral incisors, canines, first premolars, and background. The CBCT scans segmented by the Mask RCNN-SA network demonstrated a large overlap with the gold standard segmentations. The Mask RCNN-SA network–based 3D models of the maxilla and the teeth showed minor surface deviations when compared with the corresponding gold standard 3D models. The Mask RCNN-SA framework reduced segmentation time compared with manual segmentation, confirming volumetric parameters as key biomechanical RR indicators for quantitative preventive strategies.