Develop an esthetic-zonal non-invasive periodontal assessment tool based on YOLOv8 and intraoral images

Objective Non-periodontal specialists (e.g., orthodontists and prosthodontists) often struggle to efficiently monitor periodontal health in the anterior esthetic zone. This study aimed to develop the Esthetic-zonal Non-invasive Periodontal Assessment Tool (ENPAT), a you look only once (YOLO)v8-based tool that screens oral health conditions and grades periodontal abnormalities from intraoral images. Methods From 3,608 frontal intraoral images, a lightweight YOLOv8 classification model produced unlabeled oral health grading (OHG, Fair/Acceptable/Poor) to pre-screen image labelability. For unit-level assessment, triangular gingival regions were annotated in 2,029 images for modified gingival index (MGI) and 1,847 images for papillae filling index (PFI). We trained YOLOv8s-seg and a GhostNet-modified variant (YOLOv8-Ghost) with five-fold cross-validation, using ResNet-50 as a baseline classifier. A retrospective real-world test set (MGI n = 156 and PFI n = 121) was used to assess generalization, while two junior dentists then graded the same set twice (with and without artificial intelligence (AI)) to analyze agreement (Cohen’s κ and weighted κ) with experts’ diagnosis and efficiency (ΔTime%). Results OHG achieved mean accuracy of 0.872. In cross-validation, YOLOv8 outperformed Ghost-YOLOv8 on PFI (macro-accuracy 0.927 vs. 0.920, F1-score 0.762 vs. 0.743, mAP@50 0.726 vs. 0.702; P ≤ 0.001), with smaller grade-specific gains on MGI (e.g., accuracy of MGI0 0.867 vs. 0.853, P = 0.017). Both YOLOv8 variants exceeded ResNet-50 on real-world set, while YOLOv8 showed the strongest overall PFI performance (accuracy 0.940, mAP@50 0.854, F1-score 0.838). YOLOv8-Ghost reduced complexity substantially (−44.2% params and −29.0% GFLOPs). With AI assistance, junior dentists’ diagnosis rose to satisfactory levels (weighted κ up to 0.799), and per-image evaluation time decreased by 18.10% for MGI and 22.79% for PFI (P < 0.05). Conclusion ENPAT delivers real-time, multi-class grading of periodontal conditions from routine intraoral photos. The standard YOLOv8s-seg model offers the best overall accuracy and sensitivity, while YOLOv8-Ghost provides a compelling lightweight alternative. These results support ENPAT’s potential to enhance periodontal screening and education for non-periodontal practitioners.