Automated Periodontal Bone Loss Detection on Panoramic Radiographs Using You Only Look Once v8 (YOLOv8): A Retrospective AI Approach

Abstract Aim: We aimed to evaluate the diagnostic performance of the You Only Look Once version 8 (YOLOv8) deep learning model in the automated detection of periodontal bone loss (PBL) on panoramic radiographs and assess its potential as a clinical diagnostic aid. Materials and Methods: A total of 500 annotated panoramic radiographs with PBL were retrospectively collected and randomized into training ( n = 400), validation ( n = 50), and testing ( n = 50) datasets. Image annotation was performed using Roboflow and validated by a radiologist and periodontist. YOLOv8 models of five variants (n, s, m, l, and x) were trained using Google Colab. Model performance was evaluated via confusion matrix-derived metrics: accuracy, precision, recall, specificity, F1-score, and average detection time. False cases were also analyzed. Results: After training and optimization, YOLOv8-s showed the highest mAP result (83.7 ± 1.74%) compared to other variants. During testing, YOLOv8-s demonstrated the highest balance between speed and performance, achieving an accuracy of 90.27% (CI 95% [88.49, 92.05]), precision of 91.27% (CI 95% [89.25, 93.29]), recall of 94.64% (CI 95% [92.99,96.29]), specificity of 80.94% (CI 95% [76.76,85.12]), and an F1-score of 92.98% (CI 95% [91.22, 94.78]) at a 10% confidence threshold. The average detection time per image was 22.6 ms. The most frequent false positives were healthy teeth, while missed detections were mainly mild cases of PBL. Conclusion: The YOLOv8 model effectively detects PBL on panoramic radiographs with high accuracy and speed. It holds significant potential as an assistive tool in clinical diagnosis of periodontitis. Future research should focus on severity classification and multi-center validation to enhance clinical integration.