GENERALIZABILITY OF CLOUD-BASED AI SOFTWARE FOR ANTERIOR TOOTH SEGMENTATION IN MULTICENTER CBCT DATASETS: AN EXTERNAL VALIDATION STUDY.
Authors
Affiliations (7)
Affiliations (7)
- Dental Sciences Graduate Program, Federal University of Espírito Santo (UFES), Vitória, Espírito Santo, Brazil.
- Department of Oral Diagnosis, Piracicaba Dental School, University of Campinas, Piracicaba, SP, Brazil.
- Department of Dentistry, University of Brasília, Brasília, DF, Brazil.
- Department of Propaedeutics and Integrated Clinic, Division of Oral Radiology, School of Dentistry, Federal University of Bahia, Salvador, BA, Brazil.
- Department of Stomatology, Public Oral Health and Forensic Dentistry, Division of Oral Radiology, School of Dentistry of Ribeirão Preto, University of São Paulo, Ribeirão Preto, Brazil.
- OMFS IMPATH Research Group, Department of Imaging and Pathology, Faculty of Medicine, KU Leuven & Department of Oral and Maxillofacial Surgery, University Hospitals, Campus Sint-Rafael, 3000 Leuven, Belgium; Department of Dental Medicine, Karolinska Institute, Stockholm, Sweden.
- Dental Sciences Graduate Program, Federal University of Espírito Santo (UFES), Vitória, Espírito Santo, Brazil; Department of Oral Diagnosis, Piracicaba Dental School, University of Campinas, Piracicaba, SP, Brazil. Electronic address: [email protected].
Abstract
To externally validate the generalizability of a cloud-based artificial intelligence (AI) software for automated anterior tooth segmentation in cone-beam computed tomography (CBCT) scans acquired with five CBCT systems and to identify factors associated with the need for manual refinement. A total of 190 CBCT scans from five systems were analyzed. Automated segmentation was performed using Virtual Patient Creator (Relu, Leuven, Belgium). Two examiners evaluated 879 tooth segmentation maps, with refinements performed when necessary. Automated and refined segmentations were compared using voxel-wise, surface-based, and time-efficiency metrics. Factors associated with the need for refinement were assessed using mixed-effects logistic regression (α=5%). Automated segmentation was adequate in 90.1% of cases. Endodontic treatment (OR=4.43), orthodontic brackets (OR=3.74), and adjacent high-density artifacts (OR=7.88) were significantly associated with a higher need for refinement (p<0.05). Automated segmentations showed high performance across CBCT systems, with Intersection over Union (IoU) ranging from 0.92 to 0.95, Dice Similarity Coefficient (DSC) from 0.96 to 0.97, recall from 0.94 to 0.95, precision and accuracy above 0.97, Median Absolute Distance (MAD) below 0.07 mm, and Root Mean Squared Error (RMSE) below 0.10 mm. Automated segmentation was substantially faster than refined and manual segmentation. The cloud-based AI software showed high performance for anterior tooth segmentation across different CBCT systems, supporting its generalizability under the tested conditions. However, endodontic treatment, orthodontic brackets, and adjacent high-density artifacts increased the likelihood of refinement. The reduced segmentation time and consistent performance across CBCT systems support the potential integration of cloud-based AI segmentation into digital dental workflows, especially for anterior teeth, where accurate morphology is clinically relevant.