Efficient Lightweight YOLO for Adenoid Detection and Segmentation in Cephalometric Radiographs

Adenoid hypertrophy is a common pediatric and adolescent condition that can cause mouth breathing, snoring, adenoidal facies, and craniofacial developmental issues. In clinical ENT diagnosis, medical imaging is often required to evaluate adenoid status. In this paper, we introduce a lightweight YOLO-based model for automatic detection and segmentation of adenoids and upper airway regions in lateral cephalometric radiographs. The model incorporates lightweight depthwise separable convolutions and an improved attention fusion module (C3K2CSA), significantly reducing parameter size and computation cost. Experimental results show a Box mAP₅₀ of 0.922 and a Mask mAP₅₀ of 0.94, with reductions of 17.5% in model size and 13.7% in computation. These results demonstrate the model's effectiveness and suitability for clinical deployment.