From Photos to Indices: Deep Learning-Based Photographic DMFT (pDMFT)

Introduction Timely, accurate Decayed-Missing-Filled Tooth (DMFT) data are essential for caries surveillance, but conventional examinations are resource-intensive. Objective This study aims to develop and evaluate pDMFT-AI, an automated system for photographic DMFT (pDMFT) scoring from intraoral images, as a scalable proxy for remote caries monitoring. Methods We retrospectively analyzed 5000 intraoral photographs (89,825 permanent teeth) acquired in oral health surveys in Vietnam, evenly distributed across five standardized views. Two dentists independently assigned pDMFT codes to each tooth, with disagreements adjudicated by an expert (κ = 0.82-0.83). Teeth were automatically segmented and enumerated using SegmentAnyTooth to crop tooth-level images and detect missing teeth. A multi-label classifier with a ConvNeXt-Tiny backbone, trained with self-supervised and transfer learning, predicted four conditions: decayed, filled, prosthetic, and not recorded. Data were split 70/10/20 into training, validation, and test sets; performance was evaluated on 18,247 teeth. Results On the test set, pDMFT-AI achieved mean sensitivity 0.917 (95% CI: 0.902-0.932), specificity 0.989 (95% CI: 0.988-0.990), and average precision 0.948 (95% CI: 0.933-0.959). Per-class sensitivities were 0.881 to 0.940 and specificities 0.982-0.998. Subject-level agreement between pDMFT-AI and expert annotations was excellent (intraclass correlation coefficient 0.90, 95% CI: 0.87-0.93). Grad-CAM heatmaps indicated the model focus on clinically relevant regions. Conclusion pDMFT-AI showed high accuracy for automated pDMFT assessment from intraoral photographs and may support large-scale photographic caries surveillance. Clinical relevance By automating tooth- and arch-level pDMFT scoring from intraoral photographs, pDMFT-AI could streamline high-volume screenings and facilitate standardized monitoring when chairside examinations are limited.