Self-configuring 3D Segmentation of Pediatric Dentition

Robust 3D segmentation of primary and permanent teeth in cone-beam CT (CBCT) is critical for pediatric and orthodontic care. We propose a fully automatic deep-learning pipeline built on the self-configuring nnU-Net v2 framework, tailored for high-fidelity dental shape modeling. Our approach learns fine-scale tooth geometries directly from volumetric data, eliminating manual tuning. On a pediatric CBCT cohort (369 training, 93 validation, 55 test scans), our model attains a mean Dice score of 0.87 across 55 dental and supporting anatomical structures. Key components include adaptive preprocessing (isotropic resampling, automatic craniofacial cropping, intensity normalization), on-the-fly 3D augmentations, and lightweight postprocessing to remove spurious segment. The resulting segmentations are consistent and clinically actionable, supporting advanced 3D morphometric analysis and digital treatment planning. By extending state-of-the-art volumetric segmentation to mixed dentition CBCT data, our work facilitates integration of AI-driven geometric learning into routine pediatric dentistry workflows.