Date on Master's Thesis/Doctoral Dissertation


Document Type

Master's Thesis

Degree Name



Oral Biology

Degree Program

Oral Biology, MS

Committee Chair

Scarfe, William

Committee Co-Chair (if applicable)

Santaella, Gustavo

Committee Member

Metz, Michael

Author's Keywords

Dentistry; artificial intelligence; dental caries; bitewing


Background: Interpretation of bitewing radiographs is influenced by factors such as acquisition parameters (e.g. exposure, type of sensor), clinical technique, visualization (e.g. monitor type and calibration) and the observer (e.g. experience and fatigue bias). We hypothesized that the use of artificial intelligence (AI) will reduce visualization and observer factor in bitewing interpretation and improve diagnostic accuracy. Objective: The purpose of the present study was to evaluate the use of AI in the form of a machine-learning algorithm to detect and quantify proximal lesions compared with human trained observers. Methods: 16,000 anonymized, digital bitewings of patients were hand searched and non-bitewing, poor quality images with personal health identifiers were excluded from the study. The images were randomly assigned into four sets: a) Training dataset for training AI, b) Calibration dataset for training 3 experts and 3 evaluators with AI software interface use, c) Ground truth set displayed to 3 experts used to provide a consensus truth, and d) Testing Subset displayed to three general dental practitioners (GDP) and used to evaluate the performance of the AI and GDPs compared to the experts. Sensitivity and specificity were calculated and receiver operator characteristic analysis was used to determine accuracy and compared using ANOVA (p≤0.025). Results: Overall sensitivity for AI (0.62) was greater compared to observers (mean, 0.52; range, 0.33-0.74) whereas specificity for AI (0.71) was reduced compared to observers (mean, 0.94; range, 0.87-0.98). Overall ROC for AI (0.7; CI: 0.66-0.74) was similar to the observers (0.74; CI: 0.69-0.78). Sensitivity increased for observers overall with increasing lesion (0.22 to 0.75) size but remained steady for AI (0.40 to 0.58). Conclusion: Using a limited learning dataset, AI provided a higher sensitivity for proximal lesion detection and greater accuracy for incipient sized lesions than observers. Further AI training is necessary to increase the specificity of dental proximal lesion detection.