Artificial intelligence for detecting subtle paediatric tibial fractures in children under three years: An analysis of YOLO architectures with convolutional Block attention

INTRODUCTION: Subtle tibial fractures in children under three are frequently missed, particularly in cases of suspected physical abuse. Early, accurate identification is essential for clinical decision-making and safeguarding. This study assessed whether YOLO-based object detection models can reliably detect subtle tibial fractures on paediatric radiographs. METHODS: A retrospective dataset of 245 tibial fractures and 637 non-fracture anterior-posterior radiographs was collected. Fractures were annotated using radiologist-verified polygonal labels. YOLOv5, YOLOv8, and YOLOv11 were trained with five-fold cross-validation, and the effect of adding Convolutional Block Attention Module (CBAM) was examined. Performance was evaluated using mAP50, mAP50-95, and per-image diagnostic accuracy. Eigen-CAM heatmaps were used to assess model interpretability. RESULTS: YOLOv8 achieved the strongest overall performance. Adding CBAM improved localisation and increased mean detection confidence by 10.6%. Buckle fractures were most reliably identified, while periosteal and spiral fractures remained challenging, likely due to limited representation and subtle radiographic appearance. CONCLUSION: YOLO-based models show potential for detecting subtle tibial fractures in infants and toddlers, and CBAM further enhances performance in difficult cases. Larger multicentre datasets are needed to strengthen generalisability. IMPLICATIONS FOR PRACTICE: AI-assisted detection could support radiologists in suspected non-accidental injury, reduce missed fractures, and improve diagnostic pathways, particularly in settings without specialist paediatric radiology expertise.