Deep learning-based identification, classification and segmentation of infraosseous periodontal defects using a YOLOv8 neural network.
Authors
Affiliations (5)
Affiliations (5)
- Department of Periodontology, Semmelweis University, Szentkirályi Street 47, Budapest, 1088, Hungary.
- Department of Orthopedics, The First Affiliated Hospital of Harbin Medical University, Harbin, 150001, China.
- Department of Pharmacology (State-Province Key Laboratories of Biomedicine-Pharmaceutics of China, Key Laboratory of Cardiovascular Research, Ministry of Education), College of Pharmacy, Harbin Medical University, Harbin, 150081, China.
- Department of Periodontology, University of Bern, Freiburgstrasse 7, Bern, 3010, Switzerland.
- Department of Periodontology, Semmelweis University, Szentkirályi Street 47, Budapest, 1088, Hungary. [email protected].
Abstract
The defect morphology determines single-tooth prognosis, surgical planning, and regenerative potential of intrabony defects. 2D radiographs are often compromised by superimposition of anatomical structures, leading to potential misinterpretation and diagnostic inaccuracies. The aim of this study was to develop YOLOv8 model to identify, classify and segment periodontal defects. A dataset of 329 periapical radiographs from the Department of Periodontology, Semmelweis University, was utilized. All radiographs were preprocessed and manually annotated. The YOLOv8 neural network was trained and mainly assessed by area under the receiver operating characteristic curve (AUC-ROC), macro-average F1-score, Intersection over Union (IoU) and Dice similarity coefficient (DSC). The model achieved an AUC-ROC of 0.8078 (95% CI: 0.6633-0.9257). The macro-average F1-score was found to be 0.6559. IoU and DSC value averaged 0.6881 ± 0.2398 and 0.7789 ± 0.2664. High spatial overlap was observed for one -wall (DSC: 0.8688), three-wall (DSC: 0.8723) and four-wall (DSC: 0.8641) defects. The utilized YOLOv8 model demonstrated the capability to identify, classify, and segment infraosseous periodontal defects, achieving good discriminative power and moderate classification/ segmentation performance. Future work will have to focus on constructing a larger and even more refined training dataset to further enhance model performance.