RCTE: A multi-class object detection framework for dental panoramic radiographs.
Authors
Affiliations (4)
Affiliations (4)
- School of Stomatology, Xinjiang Medical University, Urumqi, Xinjiang, China.
- College of Information Science and Technology, Shihezi University, Shihezi, Xinjiang, China.
- The First Affiliated Hospital of Shihezi University, Shihezi, Xinjiang, China.
- Key Laboratory of Intelligent Decision-Making for Frontier Security, Xinjiang Production and Construction Corps, Shihezi, Xinjiang, China.
Abstract
Dental panoramic radiographs play an important role in oral disease screening, computer-aided diagnosis, and clinical treatment planning. However, accurate multi-class detection remains challenging due to small lesion sizes, ambiguous anatomical boundaries, insufficient multi-scale feature representation, and missed detections of complex dental structures. This study aims to develop an accurate and efficient multi-class object detection framework for dental panoramic radiographs to improve the detection performance of small lesions, weak-boundary structures, and complex anatomical targets. A novel re-parameterized cross-scale attention-enhanced framework, named RCTE, was proposed based on the YOLOv8n detector. The proposed framework integrates three complementary components: a Cross-Scale Channel Transformer (CSCT) module for cross-scale contextual interaction among P3, P4, and P5 features, a RepNCSPELAN4-based Re-parameterized Feature Pyramid Fusion (RPF) structure for enhancing multi-scale feature aggregation, and a Multi-Scale EMA (MS-EMA) mechanism for feature recalibration before the detection head. Experiments were conducted on a publicly available dental panoramic radiograph dataset containing 11 categories of dental structures and lesions. Model performance was evaluated using Precision, Recall, F1-score, mAP50, mAP75, and mAP50-95. Compared with the original YOLOv8n baseline, RCTE improved mAP50, mAP75, mAP50-95, and Recall by 2.55, 3.88, 2.40, and 5.10 percentage points, respectively. The proposed framework achieved better detection completeness and localization accuracy compared with other YOLO-based detectors. Furthermore, RCTE maintained real-time inference capability, demonstrating a favorable balance between detection accuracy and computational efficiency. The proposed RCTE framework effectively improves multi-class object detection performance in dental panoramic radiographs by enhancing cross-scale feature interaction, multi-scale feature fusion, and detection feature recalibration. This method provides a potential solution for computer-aided dental image analysis.