Hand bone extraction and segmentation based on a convolutional neural network.

• Proposed a key bone extraction network based on YOLOv5. • The GRU and Unet networks are integrated to achieve high-quality segmentation of key bones. • The proposed two-stage key bone segmentation was verified by experiments. • Evaluate the accuracy of partial baseline target detection network, semantic segmentation network and the proposed method. Automatic segmentation of hand bone X-ray images is an important process in skeletal bone age assessment (BAA). The segmentation process is extremely difficult and complicated due to the complex structure and tiny size of hand bone features, which may lead to inaccurate feature localization and incomplete segmentation. A two-stage segmentation method is proposed in this paper. In the first stage, the OSA-YOLOv5 network is used to extract hand bones, and in the second stage, GRU-UNet is used to separate extracted hand bones. The hand bone X-ray dataset consists of a mixed dataset of 1000 left-hand bone X-ray images of adolescents from licenced local hospitals and 1000 left-hand bone X-ray images from the RSNA public dataset. The training efficiency of the OSA-YOLOv5 network is 28 % higher than that of the traditional network, and the proposed network exhibits with a higher extraction accuracy. The segmentation accuracy of the GRU-Unet segmentation model 14.70 % higher than that if Unet. The spatial sequence features with spatial rules are found to have a positive correlation in GRU-Unet. Experimental results show that the proposed segmentation method has good generalization and robustness, which provides a good foundation for subsequent BAA. [ABSTRACT FROM AUTHOR]