Skeletal Maturity Recognition Using a Fully Automated System With Convolutional Neural Networks

Shuqiang Wang, Yanyan Shen, Changhong Shi, Peng Yin, Zuhui Wang, Prudence Wing-Hang Cheung, Jason Pui Yin Cheung, Keith Dip-Kei Luk, Yong Hu

Abstract: In this paper, we present an automated skeletal maturity recognition system that takes a single hand radiograph as an input and finally output the bone age prediction. Unlike the conventional manually diagnostic methods, which are laborious, fallible, and time-consuming, the proposed system takes input images and generates classification results directly. It first accurately detects the distal radius and ulna areas from the hand and wrist X-ray images by a faster region-based convolutional neural network (CNN) model. Then, a well-tuned CNN classification model is applied to estimate the bone ages. In the experiment section, we employed a data set of 1101 hand and wrist radiographs and conducted comprehensive experiments on the proposed system. We discussed the model performance according to various network configurations, multiple optimization algorithms, and different training sample amounts. After parameter optimization, the proposed model is finally achieved 92% and 90% classification accuracies for radius and ulna grades, respectively.