부산대학교 도서관

At present, the bone age assessment scheme based on a convolution neural network generally uses the whole hand, metacarpal finger bone, or multi-scale hand bone image as input, but there is less work to use a carpal bone image to predict bone age. A method of bone age assessment based on the TW3-C Carpal scoring method is proposed to address the less friendly prediction results of bone age for children in the younger age groups by the above scheme. To enhance the detection accuracy of the region of interest, the target detection network is improved by replacing the backbone network in its baseline network to make it more fully extract features and provide a good basis for the subsequent bone age assessment task. After data experiments, the improved detection network mAP@0.5 reaches 99% compared with the original target detection network, and the total loss rate decreases to 0.02, which is significantly lower than the loss value of general target detection algorithms, and the method effectively improves the ability of model detection. Finally, the mapping rules of the skeletal maturity class classification network and maturity curves yielded errors in bone age assessment all within 7 months.