Desarrollo de un clasificador de imágenes web de radiografías de antebrazo pediátricas mediante visión por computadora

Abstract

Distal forearm fractures in the pediatric population were among the most prevalent injuries treated in emergency settings, particularly in children aged 10 to 14 years. Accurate diagnosis proved challenging due to anatomical factors inherent to skeletal development, such as the presence of non-ossified cartilage, active growth plates, and morphological variability according to age. These conditions hindered conventional radiographic interpretation and increased the risk of underdetection for fracture patterns such as greenstick, torus, and epiphyseal lesions. In response to this diagnostic challenge, an artificial intelligence-assisted approach was proposed to improve the accuracy and efficiency of fracture identification in pediatric radiological images. Using the GRAZPEDWRI-DX dataset, composed of over 20,000 pediatric forearm radiographs with clinical annotations, three convolutional neural network models—Faster R-CNN, MobileNetV3, and YOLOv8m—were trained and validated. Evaluation was conducted using metrics such as precision, sensitivity, and mAP@0.5, with YOLOv8m selected based on its superior performance (mAP@0.5 = 0.62) and real-time inference speed (≤ 50 ms per image). The model was integrated into a modular web-based platform that enabled secure image upload, automated processing, and interactive visualization of diagnostic results through bounding boxes. The system achieved an average end-to-end latency of 575 ms and demonstrated high reliability in detecting fractures across the most representative clinical classes. Its anchor-free architecture, combined with deep learning techniques, highlighted the potential of artificial intelligence as a complementary tool in pediatric radiology. The developed platform provided a scalable and replicable solution to support orthopedic diagnostics in clinical environments with limited resources and high patient demand.