Dental caries is one of the most common diseases globally. It affects children and adults living in poverty, who have the most limited access to dental care. Left unexamined and untreated in the early stages, treatments for late-stage and severe caries are costly and unaffordable for socioeconomically disadvantaged families. If detected early, caries can be reversed to avoid more severe outcomes and a tremendous financial burden on the dental care system. Building upon a dataset of 50,179 intraoral tooth photos taken by various modalities, including smartphones and intraoral cameras, the present study developed a multi-stage deep learning-based pipeline of artificial intelligence algorithms that localize individual teeth and classify each tooth into several classes of caries. This study initially assigned International Caries Detection and Assessment System (ICDAS) scores to each tooth and subsequently grouped caries into two levels: level 1 for white spots (ICDAS 1 and 2) and level 2 for cavitated lesions (ICDAS 3 to 6). The system’s performance was assessed across a broad spectrum of anterior and posterior teeth photographs. For anterior teeth, 89.78% sensitivity and 91.67% specificity for level 1 (white spots) and 97.06% sensitivity and 99.79% specificity for level 2 (cavitated lesions) were achieved, respectively. For the more challenging posterior teeth due to the higher variability in the location of white spots, 90.25% sensitivity and 86.96% specificity for level 1 and 95.8% sensitivity and 94.12% specificity for level 2 were achieved, respectively. The performance of the developed AI algorithms shows potential as a cost-effective tool for early caries detection in nonclinical settings.
Keywords: artificial intelligence, caries, convolutional neural network, deep learning, dental public health, machine learning
