Optimized Yolov8 feature fusion algorithm for dental disease detection

Abstract

In oral panoramic film dental disease detection, image magnification distortion and low contrast often result in unclear details and features of target regions, increasing the difficulty of accurate detection. Although mainstream object detection algorithms have shown excellent performance in various fields, their direct application to dental disease detection has been suboptimal. To address these challenges, this study proposes an improved YEM-SAFN model to enhance the recognition of dental conditions. The proposed model incorporates a novel small-target network structure to address the multi-scale and significant size differences of targets in dental disease detection. Additional detection heads for different scales were introduced to improve the model's ability to recognize various dental diseases effectively. To mitigate the issue of tissue offset or overlap in panoramic dental films, the HCSA attention mechanism was integrated, enabling the model to focus on feature extraction in disease-specific regions. Additionally, a redesigned weighted fusion module enhances the utilization of features across scales, improving the model's feature representation capability. The improved YEM-SAFN algorithm achieves a 3.2 % increase in mAP compared to the original YOLOv8s algorithm, attaining an mAP of 86.7 % and outperforming other mainstream algorithms. This model provides an efficient and accurate method for dental condition identification and diagnosis.