Ahmed Ammar ; Manaf Abdul

Abstract

Wrist fractures are highly prevalent among children and can significantlyimpact their daily activities, such as attending school, participating insports, and performing basic self-care tasks. If not treated properly, thesefractures can result in chronic pain, reduced wrist functionality, and otherlong-term complications. Recently, advancements in object detection have shownpromise in enhancing fracture detection, with systems achieving accuracycomparable to, or even surpassing, that of human radiologists. The YOLO series,in particular, has demonstrated notable success in this domain. This study isthe first to provide a thorough evaluation of various YOLOv10 variants toassess their performance in detecting pediatric wrist fractures using theGRAZPEDWRI-DX dataset. It investigates how changes in model complexity, scalingthe architecture, and implementing a dual-label assignment strategy can enhancedetection performance. Experimental results indicate that our trained modelachieved mean average precision (mAP@50-95) of 51.9\% surpassing the currentYOLOv9 benchmark of 43.3\% on this dataset. This represents an improvement of8.6\%. The implementation code is publicly available athttps://github.com/ammarlodhi255/YOLOv10-Fracture-Detection