YOLO-HSD: A Multi-Module Collaborative Optimization Network for Caries Detection in Panoramic X-Ray Images

Dental caries, one of the most prevalent oral diseases, significantly impacts quality of life and can lead to severe complications such as pulpitis and periapical periodontitis. Early detection and treatment are critical to slowing its progression , yet traditional diagnostic methods heavily rely on clinician expertise and suffer from high misdiagnosis rates. To address this, we propose YOLO-HSD, a novel caries detection model based on YOLOv11, designed to identify caries in panoramic dental X-rays. Our study utilizes a dataset of 5,290 clinical panoramic X-ray images from diverse sources, encompassing caries at various stages. To enhance image clarity across multi-source X-rays, we introduce a Histogram Transformer Block (HTB). Additionally, we mitigate feature redundancy through a Spatial and Channel Reconstruction Convolution (SCConv) module for optimized feature extraction and replace conventional upsampling with Dynamic 1 Upsampling (DySample) to improve resolution and detection accuracy. These innovations ensure high-quality data input and significantly boost performance, achieving a 28.6% improvement in mAP50-95 over YOLOv11n on the test set.