Development and Application of Deep Learning-Based Model for Quality Control of Children Pelvic X-Ray Images

Abstract

A deep learning-based method for evaluating the quality of pediatric pelvic X-ray images is proposed to construct a diagnostic model and verify its clinical feasibility. Three thousand two hundred and forty-seven children with anteroposteric pelvic radiographs are retrospectively collected and randomly divided into training datasets, validation datasets and test datasets. Artificial intelligence model is conducted to evaluate the reliability of quality control model. The diagnostic accuracy, area under ROC curve, sensitivity and specificity of the model are 99.4%, 0.993, 98.6% and 100.0%, respectively. The 95% consistency limit of the pelvic tilt index of the model is -0.052-0.072. The 95% consistency threshold of pelvic rotation index is -0.088-0.055. This is the first attempt to apply AI algorithm to the quality assessment of children's pelvic radiographs, and has significantly improved the diagnosis and treatment status of DDH in children.