X-ray images of wrist bone are often used for clinical bone age assessment to focus on skeletal development in children and adolescents. Among the process, the segmentation of the hand bone region is a key step in the preprocessing, and the accuracy of hand bone segmentation greatly influences the final evaluation results. The traditional threshold segmentation methods have poor robustness in the automatic segmentation process, while using the deep neural network is more accurate but really complex. Aiming at these problems, this paper proposed to predict the optimal threshold by training Softmax regression model to obtain binary image based on threshold segmentation, then use the region growing to extract the whole hand, and normalize the processed images in the end. On the 100 test sets, the proposed method was compared with the traditional threshold segmentation methods-Otsu, maximum entropy threshold and histogram mean segmentation. DSC (dice similarity coefficient), under-segmentation rate and over-segmentation rate were used to quantitatively analyze the results. The experimental results show that the method has achieved the best results with an average DSC value of 0.97, and the under-segmentation rate and the over-segmentation rate are both close to 0. It also shows good performance for complex hand radiographs, which is proved to be more robust than the traditional threshold segmentation methods and can accurately complete automatic hand bone segmentation on skeletal age X-ray images.