Multi-Site Assessment of Pediatric Bone Age Using Deep Learning

Abstract

Pediatric bone age assessment is clinically valuable for the evaluation of a variety of pediatric endocrine and orthopedic conditions. Recent studies have explored automated methods for bone age assessment using machine learning techniques, yielding impressive results. However, many state-of-the-art methods rely on manual, fine-grained segmentation of phalanges and have not been validated on an external hospital site. The purpose of this study was to examine the efficacy of a deep learning algorithm for pediatric bone age assessment without the need for time-intensive segmentation. We utilize a novel training regime to achieve results on par with existing approaches, present a systematic analysis of experimental findings via an ablation study, and evaluate generalizability on an external dataset as a function of training data size. The final optimized model achieves mean absolute error of 7.59 months upon internal validation and 11.02 upon validation with data from an external hospital site.