Design of bifunctional composite bone age prediction network based on Swin-Transformer

Abstract

Bone age can be predicted by taking hand X-rays. Prediction of bone age is a labor-intensive and time-consuming radiological clinical task. This paper combined a series of prevailing deep learning methods to address this problem. Stages1~3 of Swin Transformer served as backbone, with the multi-scale neck network and detection head of Yolox connected to it. These formed the first period’s hand bone joints detector. After finishing detector training, the pre-trained Stages1~3 were frozen for the second period’s developmental grades classification of corresponding bone joint. Additionally, a linear classification head was attached to Swin Transformer’s stage4, where it functioned as second period’s classifier for different developmental grades. Therefore, a dual-purpose composite network was created like this. It made the bone age prediction model have high integrated level, and the two periods could be applied fusion training. In addition, different attention mechanisms were introduced at different positions, loss functions and optimization methods were also redesigned to ensure improvement of network performance. In the hand bone joints detection period, compared to the original Yolox-X, there was a 5.27% increase in Ap@50 and a 40.12% increase in Ap@50:95. As for the developmental grade classification period, the validation accuracy surpassed that of EfficientNetV2-L by 5.18%, with one-third the training size.