LAUSR

Skeletal bone age assessment based on hand x-ray is widely used in many fields. There is an urgent need for automated method to alleviate manual labor and address the problem of intra- and inter-observer variability. Most existing methods modeled the task as regression or multiclass classification problems. However, the regression method over-simplifies the relation between image features and bone age as linear while the bone development follows a nonlinear pattern; multiclass classification undesirably ignores the ordinal information of the age labels. In this work, we pioneered the applying of ranking learning to the problem of bone age assessment and proposed a two-stage bone age assessment network. The first stage is a CNN network VGG-U-Net used to segment the hand/wrist from the X-ray image. Then, a conditional GAN network was constructed to assess bone age. The generator was a ranking CNN which consisted of multiple binary classification outputs. We also proposed to use a rank-monotonicity loss to improve the performance. We validated the proposed method using the RSNA2017 Pediatric Bone Age dataset and the usefulness of different components was also investigated. The proposed method achieved an averaged mean absolute error (MAE) of 6.05 month (6.01 month for the male cohort and 6.09 month for the female cohort). Its performance was comparable with other state-of-the-art CNN based methods. Through ablation study, we found all proposed components (including the rank-monotonicity loss, adversary training strategy etc.) took effect and contributed to the final performance. In conclusion, we illustrated that rank-monotonicity enhanced ranking learning is more suitable for the task of bone age estimation. The proposed method is a valuable alternative for automatic bone age estimation.