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

Thesis Abstract

The early detection of dental caries remains a critical challenge in preventive dentistry, as traditional diagnostic methods such as visual examination and radiography often lack the sensitivity and specificity required to identify initial or subsurface lesions. This study aims to develop and validate an artificial intelligence (AI)-powered diagnostic tool capable of accurately detecting early-stage carious lesions to facilitate prompt intervention and reduce the progression to cavitation. The specific objectives include designing a robust machine learning model trained on diverse dental imaging datasets, evaluating its diagnostic accuracy compared to conventional methods, and assessing its usability within clinical workflows. The research employs a mixed-methods approach centering on a quantitative experimental design complemented by qualitative usability evaluation. The study population comprises 600 anonymized digital dental images collected from patients at a university dental clinic over a two-year period, capturing a broad spectrum of caries stages in patients aged 12 to 65 years. A stratified random sampling technique was used to select a representative subset of 300 images annotated by expert dentists to serve as labeled training and testing datasets. The AI model development leveraged convolutional neural networks (CNNs) trained using transfer learning from pre-trained models such as ResNet50, implemented in Python with TensorFlow. The model's diagnostic performance was assessed via receiver operating characteristic (ROC) curve analysis, area under the curve (AUC), sensitivity, specificity, and F1-score metrics. Validation included cross-validation procedures and a comparative analysis with assessments made by five experienced dentists on the same image set. Furthermore, qualitative data on user experience was gathered through semi-structured interviews with 15 clinicians, analyzed through thematic content analysis. It is anticipated that the AI model will demonstrate superior sensitivity (>85%) and specificity (>80%) in identifying early carious lesions relative to visual and radiographic assessments alone, with an AUC exceeding 0.90 indicating high diagnostic accuracy. The model is expected to generalize effectively across diverse imaging conditions, thus promising a reliable adjunctive diagnostic aid. Qualitative insights are projected to reveal high perceived usefulness and ease of integration into existing clinical workflows, despite concerns over transparency and interpretability of the AI decisions. This research significantly contributes to the body of knowledge by providing empirical evidence on the application of deep learning in early caries detection, addressing existing gaps in diagnostic sensitivity, and offering a scalable, cost-effective solution for dental clinics. The study also explores the integration of AI within the framework of the Health Belief Model, emphasizing factors influencing clinician adoption of new technologies. The development of this diagnostic tool aligns with the paradigms of precision dentistry and digital health, advocating for enhanced early intervention practices. The main conclusion underscores that AI-driven diagnostic tools can augment clinical decision-making in dentistry, leading to earlier and more accurate detection of caries. Based on the findings, recommendations include adopting AI-assisted diagnosis in routine screenings, further augmenting models with multimodal data (e.g., spectral imaging), and conducting longitudinal studies to evaluate impact on patient outcomes. Future research should explore the integration of AI diagnostic tools with electronic health records and the development of explainable AI models that improve clinician trust and understanding of AI reasoning processes.