Improved Automatic Radiographic Bone Age Prediction with Deep Transfer Learning

Abstract

Radiograph based pediatric bone age assessment plays a key role in many medical and judiciary tasks. The traditional bone age identification techniques, such as GP method and TW scoring, rely on human experts to compare X-ray image of the hand skeleton with the standard atlas and visually examine the similarities, thus impacted by an individual's experience and judgment. The recent advances of deep neural networks (DNN) have made possible more accurate bone age estimation with the help of computer vision representation and understanding. DNNs usually demands large scale annotated image datasets for training, but this is not guaranteed for bone age estimation because of the incurred cost and the concerns of privacy protection. In this paper, we report our work of developing a convolutional neural network (CNN) based bone age classifier and transferring the low level hand radiograph features learned from the Radiological Society of North America (RSNA) dataset to the small dataset collected from a local hospital. The first few layers of the proposed network are treated as the feature extractors while the remaining parts are further fine-tuned to improve the transfer performance. The experiment results show that the proposed transfer learning framework achieves a mean absolute difference (MAD) of 0.16 years, outperforming other state-of-the-art algorithms.