AI-driven Intraoral Scanning for Early Caries Detection and Archival Analysis

Thesis Abstract

The rapid advancement of digital dentistry and artificial intelligence presents an opportunity to transform early caries detection through AI-driven intraoral scanning, addressing delays in diagnosis, standardization of imaging, and long-term archival analysis of dental records. Despite improvements in optical scanners, variability in image quality, lack of standardized interpretation criteria, and limited integration with electronic health records hinder consistent early detection and longitudinal tracking of carious lesions. This study aims to develop and validate an AI-enhanced intraoral scanning workflow that (i) enables automated detection of early caries by leveraging high-resolution 3D intraoral scans, (ii) standardizes lesion characterization and grading across diverse patient populations, and (iii) implements a robust archival framework for longitudinal analysis and audit trails. The specific objectives are (1) to design a convolutional neural network-based diagnostic model trained on a curated repository of labeled intraoral scans to identify incipient caries with high sensitivity and specificity; (2) to develop standardized imaging and annotation protocols to harmonize data quality across scanners and clinical settings; (3) to evaluate the diagnostic performance of the AI model against calibrated expert consensus using a multi-center dataset comprising 1,200 patients and 6,500 scan views; (4) to create an archival analytics module capable of longitudinally tracking lesion progression, restorations, and treatment outcomes over a five-year horizon; (5) to assess clinician usability, acceptance, and workflow integration through a mixed-methods study guided by the Technology Acceptance Model (TAM) and the Unified theory of Acceptance and Use of Technology (UTAUT). The theoretical framing incorporates the Technology-Organization-Environment (TOE) framework to examine organizational and contextual factors affecting implementation. A prospective, multi-center study design will be employed, incorporating two phases. Phase I involves data collection from three tertiary dental centers and two community clinics, enrolling 1,200 participants after obtaining informed consent, with each participant contributing standardized intraoral scans using two different high-resolution scanners. Phase II entails longitudinal follow-up at 12, 24, and 60 months to capture caries progression, sealing, fillings, and restorative outcomes. Data collection instruments include calibrated intraoral scanners, a standardized annotation protocol, a validated labeling schema for lesion severity, and a structured survey for clinician feedback. Image data will be preprocessed for noise reduction, color normalization, and geometric alignment, followed by annotation adjudication by a panel of five board-certified pediatric and general dentists to establish ground truth. The AI analytic pipeline will integrate a 3D CNN architecture augmented with attention mechanisms to capture surface texture, subsurface demineralization cues, and morphological changes in lesion pits. Model training will use cross-validated k-fold splits with stratified sampling to preserve caries prevalence, and performance metrics will include area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and F1-score. For archival analysis, a relational database coupled with a time-series schema will store scan metadata, lesion annotations, and treatment events, enabling mixed-effects modeling and survival analyses to quantify progression rates and treatment influence. Statistical analyses will include multilevel logistic regression to adjust for patient- and site-level covariates, survival analysis for time-to-progression, and ANOVA to compare diagnostic performance across scanner types. A qualitative component will analyze clinician interviews using thematic analysis to elucidate usability barriers and perceived value. Expected findings include (i) a diagnostic model achieving AUC ? 0.92, sensitivity ? 0.88, and specificity ? 0.90 across sites; (ii) demonstrable standardization gains with reduced inter-operator variability in lesion labeling; (iii) robust archival analytics revealing progression patterns and predictors of restorative success; and (iv) positive clinician acceptance with TAM- and UTAUT-aligned results indicating high perceived usefulness and ease of use. The study contributes to knowledge by integrating AI-driven 3D imaging with standardized data governance and longitudinal archival capabilities, offering a scalable framework for early caries detection and outcome-oriented dental records. The conclusion anticipates that AI-augmented intraoral scanning will improve diagnostic accuracy, foster consistent documentation, and enable evidence-based preventive strategies. Recommendations include widespread training for clinicians in standardized annotation, continuous model recalibration with new data, investment in interoperable data architectures for dental records, and policy development to govern data privacy and ethical use of AI-enhanced imaging in dentistry.