Accuracy of artificial intelligence applications in periodontics: a thematic narrative review

TL;DR

This review evaluates how well AI works in periodontal diagnostics using various imaging techniques, finding that AI can match clinicians in some tasks but needs improvement in others.

Contribution

The study provides a structured synthesis of AI diagnostic accuracy in periodontics, highlighting performance variability across imaging modalities and model types.

Findings

CNN-based models on periapical radiographs achieved high accuracy (AUCs above 0.88), comparable to clinicians.

CBCT-based deep learning systems outperformed panoramic radiographs with higher accuracy (up to 0.91) and better volumetric assessment.

Intraoral photographic analyses showed inconsistent performance (0.46–1.00) due to imaging and reference standard variability.

Abstract

Artificial intelligence (AI) has been increasingly applied to periodontal diagnostics across periapical, bitewing, panoramic radiographs, cone-beam computed tomography (CBCT), and intraoral photographs. Recent multicenter, external validation, and explainability-focused studies have advanced the field, yet variability in datasets, anatomical sites, reference standards, model architectures, and reporting practices introduces significant heterogeneity. A structured synthesis of current evidence is therefore warranted.

This review synthesizes 35 studies published between 2019 and 2025, evaluating AI applications in four diagnostic domains: detection of periodontal bone loss, measurement of alveolar bone levels, identification of furcation involvement, and detection of periapical lesions. Convolutional neural network (CNN)-based models using periapical radiographs achieved moderate-to-high diagnostic accuracy (0.82–0.85) and AUCs above 0.88, comparable to clinician performance. Panoramic radiographs yielded lower sensitivity and specificity than CBCT, where deep learning systems reached higher accuracy (up to 0.91) and superior volumetric assessment. Intraoral photographic analyses showed variable performance (0.46–1.00), largely due to inconsistent imaging and reference standards. Emerging trends include hybrid segmentation–classification architectures, transformer-based networks, and clinician-in-the-loop approaches. Determinants of performance encompass reference standard quality, dataset diversity, anatomical complexity, and adherence to STARD-AI and TRIPOD-AI reporting frameworks.

AI demonstrates clinically relevant diagnostic accuracy in periodontal imaging, especially for measurement standardization and decision support. Although autonomous diagnosis remains premature, integrating explainable, externally validated AI systems within clinician-guided workflows supported by standardized reporting offers a practical route toward clinical translation.