Deep Learning-Based Detection of Periodontal Infrabony and Furcation Defects on Periapical Radiographs: A Feasibility Study.
Authors
Affiliations (4)
Affiliations (4)
- Division of Periodontics, School of Dental Medicine, Department of Surgical Sciences, Faculty of Medicine, University of Cagliari, Cagliari, Italy; College of Dentistry, American University of Iraq Baghdad (AUIB), Baghdad, Iraq. Electronic address: [email protected].
- Division of Periodontics, School of Dental Medicine, Department of Surgical Sciences, Faculty of Medicine, University of Cagliari, Cagliari, Italy.
- Department of Industrial and Information Engineering and Economics, University of L'Aquila, L'Aquila, Italy.
- Division of Periodontics, School of Dental Medicine, Department of Surgical Sciences, Faculty of Medicine, University of Cagliari, Cagliari, Italy; Department of Industrial and Information Engineering and Economics, University of L'Aquila, L'Aquila, Italy.
Abstract
Accurate radiographic detection and classification of periodontal osseous defects are essential for prognosis and surgical planning in regenerative periodontology. Traditional diagnostic methods offer limited morphological information, and interpretation can be operator-dependent. Recent advances in artificial intelligence (AI) have shown potential in medical image analysis, but their application to detailed classification of periodontal defects on periapical radiographs remains underexplored. A total of 7464 periapical radiographs were retrospectively collected from the clinical archive of the University Hospital of Cagliari. After expert annotation, 581 images containing at least 1 periodontal osseous defect were included. Defects were categorised into 4 types: 1-wall defects, 2-or-more-wall defects, crater-like defects, and furcation involvements. A YOLOv8 large (YOLOv8l) object detection model was trained using a patient-independent split (406 trainings, 58 validations, 117 testings). Performance was assessed using mean Average Precision (mAP) at IoU thresholds of 0.5 and 0.5:0.95, as well as class-wise precision and recall. The model achieved an overall precision of 0.592, recall of 0.435, and [email protected] of 0.504. Furcation involvements showed the highest precision (0.669) and [email protected] (0.577), followed by crater-like defects and multi-wall defects. One-wall defects were the most difficult to detect. Qualitative analysis revealed that smaller or radiographically ambiguous defects were more frequently missed. AI-assisted object detection demonstrated feasibility in classifying periodontal defects, but current performance remains limited. Although current performance is limited by dataset imbalance and the inherent constraints of 2D imaging, these models may enhance diagnostic consistency and support treatment planning. This study demonstrates the feasibility of using an AI-based object detection model to classify such defects on standard periapical radiographs. By supporting clinicians in the radiographic interpretation of defect morphology, the proposed system may contribute to more consistent diagnoses, improved case selection, and enhanced predictability of surgical outcomes in periodontal therapy.