The classic Tanner and Whitehouse 3-Chinese RUS (TW3-C RUS) method strictly divides the key epiphyseal regions of hand bone development into 9 grades,and does not fully consider the continuity of epiphyseal development, leading in certain errors in bone age assessment.To solve this problem,a modified method for assessing bone age based on TW3-C RUS is proposed.The idea of threshold method is adopted to dynamically select the first N level probability values output by the network model,and use the first N probability values as weights to calculate the hand bone weighted score,so as to reduce the error caused by the single level judgment of the hand bone.Aiming at the problem of network model redundancy,the cross stage partial network (CSP-Net) is used to lighten the deep residual network 50 (Resnext50).Experiments show that the mean absolute error (MAE) of the improved method for evaluating bone age is 0.421 4 years old for men and 0.412 8 years old for women.Compared with the classic TW3-C RUS method, the bone age evaluation efficiency and accuracy are significantly higher.The amount of network parameters after the lightweight is 46.28 MB,which is significantly lower than that of the Resnext50 network model.