Performance assessment of five artificial intelligence‐based algorithms for automated tooth segmentation and labeling on intraoral scans

PURPOSE: This study aimed to evaluate and compare the performance of five AI algorithms for tooth segmentation and labeling on intraoral scans, as their performance remains unclear. MATERIALS AND METHODS: A total of 100 intraoral scans in the STL file format were classified into two main groups: complete dentition (C) and partial dentition (P, fewer than 12 teeth). Each group was further divided by arch into four subgroups (n = 25 each): complete maxillary (Mx-C), complete mandibular (Md-C), partial maxillary (Mx-P), and partial mandibular (Md-P). The algorithms tested were the Tooth Group Network (Team CGIP), Dentbird Studio (Dentbird), Medit Ortho Simulation (Medit), NemoSmile 3D (Nemotec), and MovumStudio (MovumTech). Manual segmentations by an expert operator served as the ground truth. Performance was assessed using Python and five metrics, with Intersection over Union (IOU) as the primary indicator. Statistical analysis included permutation tests with the Bonferroni-Holm correction (α = 0.05). RESULTS: Significant differences were observed between groups and algorithms (P < 0.05). IOU scores ranged from 0.72 to 0.92 in complete dentition and showed greater variability in partial dentition (0-0.928). The Tooth Group Network and MovumStudio consistently outperformed the others, with MovumStudio achieving the highest performance across all metrics and groups. Its performance matched that of a human expert when compared against a subset of the data. CONCLUSIONS: Tooth segmentation and labeling performance vary depending on dentition completeness and algorithm choice. MovumStudio demonstrated the most robust and consistent results, comparable to expert human annotation.