Related Experiment Video
Updated: Jan 13, 2026
Digital Hybrid Model Preparation for Virtual Planning of Reconstructive Dentoalveolar Surgical Procedures
Published on: August 5, 2021
AI-driven gingival segmentation on CBCT: Validation using delineation by intraoral scanning and CBCT-based cotton
Fernanda Bulhões Fagundes1, Dhanaporn Papasratorn2, Bahaaeldeen M Elgarba3
1OMFS-IMPATH Research Group, Department of Imaging and Pathology, Faculty of Medicine, KU Leuven & Department of Oral and Maxillofacial Surgery, University Hospitals Leuven, Belgium, 3000, Leuven, Belgium; Department of Oral Diagnosis, University of Campinas, Piracicaba, Brazil.
Objectives:
To validate a novel AI approach for automated gingival segmentation directly from cone beam computed tomography (CBCT) scans acquired with cotton roll separation.
Methods:
One hundred and eighty arches (101 CBCT) were split into: ground truth (90 arches), internal validation [AI vs manual segmentation (MS); 30 arches], and clinical validation [AI vs expert refinement (R-AI); 60 arches]. Ground truth consisted of manually segmented gingival models from CBCT with cotton roll separation, using intraoral scan outlines as references. AI gingival segmentation was generated from CBCT acquired with cotton roll separation, uploaded to a cloud-based platform (Virtual Patient Creator, Relu, Belgium). R-AI was performed by refining the AI segmentation according to CBCT outlines. The performance was evaluated using voxel-wise and surface-based comparisons. Additionally, time efficiency and consistency were evaluated.
Results:
AI gingiva achieved high overlap with MS, with Dice score (DSC) of 80 ± 4 % and minor surface discrepancies, with median surface distance (MSD) of 0.02 ± 0.09 mm. Similarly, DSC between AI and R-AI was high (91 ± 4 %), and MSD was small (0.07 ± 0.07 mm). Significant differences between arches were found in AI vs R-AI; however, these differences were not clinically relevant (∆DSC∼2 %, ∆MSD∼0.00 mm, p < 0.05). Evaluator consistency was high, whereas AI achieved perfect reproducibility and significantly increased time efficiency (6.7 s) compared to R-AI (1090 s) and MS (4344 s), p < 0.001.
Conclusion:
The presently validated cloud-based AI model achieved fast, accurate, and consistent gingiva segmentation from CBCTs comparable to expert-level performance using CBCT-based cotton roll separation with delineation by intraoral scanning as a clinical reference.
Clinical Significance:
Automated 3D gingival models may assist efficient diagnosis, treatment planning, and outcome visualization in esthetic dentistry, orthodontic treatment planning, and periodontal assessment. This approach has strong potential to transform digital dentistry workflows.
