Multi-Architecture deep learning for CBCT segmentation of dental hard tissues and pulp in mixed dentition

To develop and evaluate deep learning-based 3D models (CNN, Transformer, and Mamba architectures) for automated segmentation of pulp, primary, and permanent dental structures in pediatric CBCT scans with mixed dentition.

A total of 151 CBCT scans were analyzed, comprising 105 internal dataset scans(29,478 images) and 46 external scans. Expert-annotated internal data were used for model development with 78 for training/validation and 27 for testing. Fully supervised multi-task models were trained to segment pulp, primary and permanent hard tissue structures. Six architectures (ResEncM, U-Mamba Bot/Enc, WNet, UNETR, SegResNet) were evaluated using DSC, IoU, HD95, segmentation volume, and processing time.

Within the internal dataset, 3D_fullres ResEncM achieved the highest accuracy across structures and age groups. Permanent teeth were segmented most reliably (ages 6-9: DSC 0.9859 ± 0.0086, HD95 0.2035 ± 0.0925; ages 10-13: DSC 0.9825 ± 0.0100, HD95 0.2825 ± 0.0525), while primary teeth were less accurate, particularly in older children (10-13 years: DSC 0.8855 ± 0.1557, HD95 4.4201 ± 6.5116). U-Mamba Bot and U-Mamba Enc performed well, with U-Mamba Bot excelling in pulp segmentation and U-Mamba Enc in primary structures. WNet, UNETR, and SegResNet were less consistent, especially for primary teeth in older children. In external validation, ResEncM maintained the highest DSC (0.8464 ± 0.1226) and lowest HD95 (4.9078 ± 4.9058). All models showed no significant volume bias (p > 0.05) and high volumetric correlations (ρ ≥ 0.997). UNETR was fastest (0.48 ± 0.38 min), followed by U-Mamba Bot (0.60 ± 0.40 min).

nnU-Net ResEncM showed the strongest overall performance, while U-Mamba Bot and U-Mamba Enc excelled in pulp and primary-tooth segmentation. All three models performed well across age groups, with primary tooth segmentation more accurate in younger children. External testing confirmed adequate performance, though lower than on the internal dataset, supporting the feasibility of automated CBCT segmentation.

Automated segmentation of pulp and hard dental structures can streamline CBCT assessment and enhance workflow efficiency in managing developing teeth and complex orthodontic or surgical cases. This study provides a framework for incorporating AI-assisted analysis into pediatric dental imaging by including both primary and developing permanent dentition, as well as anomalous cases.