Artificial Intelligence in Detecting Periodontal Disease From Intraoral Photographs: A Systematic Review

Abstract

This systematic review aims to evaluate the methodological characteristics and clinical performance of artificial intelligence (AI) models in detecting periodontal disease using digital intraoral photographs. This review includes peer-reviewed publications and conference proceedings in English, focusing on clinical studies of human periodontal diseases. Intraoral photographs served as the primary data source, with fluorescent and microscopic dental images excluded. The methodological characteristics and performance metrics of clinical studies reporting on AI models were analysed. Twenty-six studies met the review criteria. Various image acquisition devices were used by the resarchers including professional cameras, intraoral cameras, smartphones, and home-use devices. Ten studies used clinical examinations as reference methods, while 16 used visual examinations. Eight studies involved multiple experts in dataset annotation. Only 9 studies employed multiple intraoral views for their AI models, with the remaining studies focusing solely on the frontal view. Regarding AI tasks, 17 studies used classification, 4 used detection, and 5 used segmentations. Performance metrics varied widely and were assessed at multiple levels. Classification studies showed accuracies ranging from 0.46 to 1.00, detection studies showed accuracies from 0.56 to 0.78, and segmentation studies achieved Intersection over Union (IoU) scores of 0.43 to 0.70. AI models show potential for detecting periodontal disease from intraoral photographs, but their clinical use faces challenges. Future research should focus on improving reporting standards, standardising evaluation metrics, performing external tests, enhancing data quality, and using clinical gold standards as reference methods. Furthermore, efforts should focus on promoting transparency, integrating ethical considerations, minimising misclassification, and advancing the development of explainable and user-friendly AI systems to enhance their clinical applicability and reliability.