Multi-class anatomical landmark detection in periapical radiographs with deep learning

Abstract

This study aims to develop a deep learning model for the detection and segmentation of multiple anatomical landmarks on periapical radiographs from the maxilla and mandible. A total of 1930 paralleling-technique periapical radiographs with 21 annotated anatomical landmarks were divided into training (80%), validation (10%), and test (10%) sets. Geom etry-preserving preprocessing was applied before dataset splitting, while appearance-based augmentation was performed exclusively on the training subset after the split. A YOLOv8x-seg architecture was trained for multi-class detection and instance segmentation. Performance was evaluated using precision, recall, F1-score, Dice coefficient, Intersection-over Union, mean average precision, and receiver operating characteristic analysis. The model demonstrated stable training and consistent performance. Overall precision, recall, and F1-score were 0.820, 0.725, and 0.769, respectively, with an overall Dice coefficient of 0.621. High detection accuracy was achieved for well-defined structures such as the maxillary sinus, nasal fossa, nasal fossa floor, and nasal septum, whereas low-contrast landmarks showed reduced performance. Confidence-dependent analysis indicated optimal performance at low confidence thresholds (approximately 0.05–0.10). In conclusion, the proposed model effectively detected major anatomical landmarks on periapical radiographs while demon strating expected limitations for small or low-contrast structures. Despite substantial anatomical variability across maxil lary and mandibular regions, anterior–posterior sites, and projection-dependent appearances of similar structures, these findings demonstrate that deep learning can reliably identify key anatomical landmarks, supporting safer, more consistent, and clinically meaningful radiographic interpretation in routine dental practice.