Development of an AI-powered Diagnostic Tool for Early Caries Detection

Thesis Abstract

Early detection of dental caries is critical for preventing the progression of decay and subsequent enamel destruction, yet existing diagnostic methods often lack sensitivity and specificity, leading to delayed intervention and increased treatment costs. This study aims to develop an advanced artificial intelligence (AI)-driven diagnostic tool capable of identifying early-stage carious lesions with high accuracy using digital imaging data. The specific objectives include (1) designing a deep learning model utilizing convolutional neural networks (CNNs) trained on a curated dataset of intraoral images; (2) evaluating the diagnostic performance of the AI model in comparison with traditional visual-tactile examination and radiographic assessment; and (3) exploring the usability and clinical integration potential of the proposed tool within dental practice settings. The research adopts a quantitative, diagnostic accuracy study design grounded in a positivist paradigm, reflecting the objective measurement of AI model performance. The study population comprises 500 patients presenting for routine dental check-ups at private dental clinics over a six-month period, from which a stratified random sample of 250 patients will be selected to ensure demographic and clinical diversity. Intraoral photographs and bitewing radiographs for each participant will serve as the primary data sources. Data collection involves capturing standardized digital images with a calibrated intraoral camera and obtaining corresponding radiographs, all stored securely in a HIPAA-compliant database. The development of the AI model involves label annotation by expert dentists to establish ground truth, followed by training, validation, and testing phases utilizing a 70/15/15 split. Analytical techniques will include the application of convolutional neural network architectures such as ResNet-50 and InceptionV3, optimized through cross-validation to prevent overfitting. Model performance will be assessed via receiver operating characteristic (ROC) curve analysis, with sensitivity, specificity, positive predictive value, and negative predictive value calculations. Statistical comparison of the AI model's diagnostic accuracy against standard methods will be conducted using McNemar’s test for paired proportions and kappa statistics for inter-method agreement. Further analysis incorporates calibration plots and decision curve analysis to evaluate the clinical utility of the tool. Expected findings suggest that the AI-powered diagnostic tool will demonstrate superior sensitivity and specificity in detecting early carious lesions compared to visual-tactile and radiographic methods. The model is anticipated to achieve an area under the ROC curve (AUC) exceeding 0.90, indicating high discriminative ability. The study also expects to reveal that the integration of AI-assisted diagnostics can significantly enhance early detection accuracy, reduce diagnostic variability among practitioners, and streamline clinical workflows. This research contributes novel insights into the application of deep learning in contemporary dental diagnostics, addressing a critical gap in early caries detection capabilities. By validating the effectiveness of AI in a real-world clinical context, it advances the understanding of technology-driven diagnostics and supports the development of standardized, objective assessment tools that can be incorporated into routine practice. The findings are anticipated to inform policy guidelines on digital diagnostics and to foster further research into AI-enabled applications in dentistry. The study concludes with recommendations for integrating AI diagnostic tools into clinical protocols, emphasizing the importance of clinician training and user-centered system design to maximize utility and acceptance. Future research directions include longitudinal studies to assess the impact of AI-assisted early detection on treatment outcomes and oral health status over time, as well as exploring the application of the model across diverse clinical populations and different imaging modalities. This study ultimately aims to enhance early diagnosis, improve patient outcomes, and contribute to the digital transformation of dental healthcare delivery.