Smartphone-Based Deep Learning for Dental Plaque Detection Among Preschool Children

Introduction and aims Dental plaque causes common paediatric oral diseases, and kindergarten children have limited access to dental services. This study aims to evaluate the technical feasibility of four deep learning models in detecting dental plaque from intraoral smartphone images of primary and early mixed dentition taken in kindergarten outreach settings. Methods A total of 3888 intraoral images were collected from 648 preschool children (352 males and 296 females; mean age 4.80 ± 0.97 years, range: 2.75-7.17 years) without disclosing agents. A set of six intraoral images per child was captured using the smartphone (iPhone 7, Apple Inc) with flashlights, including the buccal surface of incisors, canines, and molars from the upper and lower jaw. The dataset was randomly divided into a training set comprising 3402 images and a test set of 486 images at child level. Mask R-CNN, MaskDINO, Mask2Former and PointSup were trained and evaluated. The accuracy, sensitivity, specificity, F1-score, mIoU (Mean Intersection over Union) were applied to assess the detecting performance. Results All four deep learning models demonstrated moderate capabilities in detecting dental plaque, with the accuracy ranging from 0.68-0.78. Among them, Mask R-CNN exhibited the most balanced overall performance, achieving the highest accuracy of 0.78 (95% CI: 0.76-0.80), sensitivity of 0.70 (95% CI: 0.66-0.74), F1-score of 0.77 (95% CI: 0.73-0.80) and mAP50 of 0.46. MaskDINO exhibited the highest specificity of 0.94 (95% CI: 0.93-0.96) and PPV of 0.91 (95% CI: 0.88-0.94). PointSup achieved the highest mIoU of 0.47. Conclusion This study demonstrates the potential of deep learning models in detecting dental plaque using smartphone-captured, unstained intraoral images among preschool children. Further external validation across diverse populations and multiple smartphone devices are required to confirm its generalizability. Clinical relevance Deep learning-assisted plaque screening shows potential to be an effective tool for non-invasive oral health assessment in the paediatric populations in kindergarten outreach settings.