Performance comparison of deep learning methods on hand bone segmentation and bone age assessment

Abstract

Bone age is the biological age that reflects the growth and development of human body. Bone age assessment has been applied and plays an important role in clinical medicine, sports science and justice. Reasonable convolution neural network (CNN) models can greatly improve the accuracy and efficiency of bone age assessment. By comparing various hand bone segmentation models trained by classical convolutional neural networks, we found that with intersection over inion (IoU) and dice similarity coefficient (Dice) as evaluation indexes, the segmentation model trained by U-Net had the best performance. Its IoU reached 0.9746, and its Dice reached 0.9871. This is contrary to our inherent recognition that the U-Net++ model is superior to the U-Net model. Based on the images segmented by U-Net, we applied five kinds of common convolutional neural networks to bone age prediction, with mean absolute error (MAE) and error accuracy within two years as evaluation indexes. The results showed that the MAE of Xception was 7.635 and the accuracy of errors within two years reached 97.59%. In this paper, we provide an optimal scheme for bone age image segmentation and bone age assessment, and provide a theoretical basis for the design of bone age assessment system.