Two-stage deep learning approach for impacted canine automated detection and segmentation on panoramic radiographs

To develop and evaluate an automated two-stage deep learning pipeline for detecting and segmenting impacted canines in panoramic radiographs, addressing the challenges of large image sizes and improving diagnostic efficiency in dental practice. The study was approved by the Ethics Committee (protocol no. S69363). A dataset of 143 panoramic radiographs (2880 × 1504 pixels) containing impacted canines was divided into training/validation and testing sets (114 and 29 images, respectively). A two-stage approach using YOLOv8 model was implemented. An object detection model identified impacted canines with bounding boxes, and a segmentation model performed fine-grained delineation on cropped regions. Both models were trained using fivefold cross-validation with optimized hyperparameters. The system operated alongside general tooth detection networks, with the outputs superimposed on the original images to provide a comprehensive visualisation. The detection model achieved precision of 0.941 and recall of 0.887, whereas the segmentation model demonstrated precision of 0.956 and recall of 0.905. The mean average precision (mAP) at intersection over union threshold 0.50 reached 0.986 for both tasks. Under stricter evaluation conditions (mAP 0.50–0.95), the system maintained robust performance with 0.941 for detection and 0.969 for segmentation across all impacted canine positions. The proposed two-stage deep learning pipeline demonstrates high accuracy in detecting and segmenting impacted canines, offering a reliable automated tool for clinical diagnosis and treatment planning in dental practice.