Automated detection of tooth loss using tooth numbering segmentation in 3D intraoral scans from a population-based sample with artificial intelligence

OBJECTIVES: This study aimed to validate an automated method for detecting tooth loss in intraoral scans (IOS) compared to clinical evaluations in a population-based sample. METHODS: 897 IOSs from 453 participants of the 1982 Pelotas Birth Cohort were included. The participants were clinically evaluated at age 40 by trained, calibrated dentists for several oral health conditions, and lost teeth were identified using the International Caries Detection and Assessment System, later transformed into the DMF-T index. IOSs were acquired using a TRIOS 3® scanner. Ground-truth annotations of present teeth were established using 3DMedX by two experienced researchers, using the same DMF-T index. A deep learning model for tooth segmentation and FDI numbering in IOSs predicted the present teeth. The clinical evaluations and model predictions of missing teeth were descriptively compared to the ground-truth annotations at three levels: any tooth, tooth types (incisors, canines, premolars, molars), and FDI numbers. RESULTS: The model was highly effective for tooth detection (F1-score: 0.994) and tooth typing (macro-F1: 0.998), and less effective for FDI numbering (macro-F1: 0.965). Furthermore, it was more accurate than the clinical evaluations for any tooth (F1-score: 0.946 vs. 0.940) and tooth types (F1-score: 0.917 vs. 0.896), and less effective for FDI numbers (F1-score: 0.768 vs. 0.815), mainly due to extracted premolars and ambiguous molars. CONCLUSIONS: The automated method was accurate for detecting any tooth loss and for detecting missing teeth of a certain type. Automated tooth loss detection in IOS using deep learning appears to be a feasible approach for epidemiological studies. CLINICAL SIGNIFICANCE: The use of IOSs with a deep learning model to automate tooth loss assessment in large epidemiological surveys offers an efficient alternative to traditional clinical evaluation. Although it requires an initial equipment investment and trained operators (dentists or supervised dental auxiliaries), this approach can reduce long-term fieldwork costs by optimizing professional time and enabling large-scale data collection.