A primary and secondary feature interactive learning network for bone age assessment

Abstract

Bone Age Assessment (BAA) is a method used to assess the degree of development of adolescents. Existing Bone Age Assessment deep learning models require additional manual annotation to locate several important regions. However, manual annotations are costly and subjective, and these regions are independent and lack mutual information exchange. In this paper, a primary and secondary feature interactive learning network (PSILNet) is proposed to locate important regions and learn their features from each other. PSILNet is specifically trained to place more attention to those bone age related regions in the X-ray image. Specifically, a backbone network with an attention erasure module is introduced to obtain primary and secondary features of hand bone images. Then, a feature self-learning module (FSLM) is designed to enhance both primary and secondary features. FSLM allows the network attention to focus on more than just the most noticeable regions. Furthermore, we designed a feature-mutual learning module (FMLM) to simulate the interaction of feature information, allowing the network to focus on the regions that are different from both. PSILNet adopts an end-to-end training method without any additional manual annotation. Experiments show that the mean absolute error of PSILNet is 4.09 months on the Radiological Society of North America (RSNA) dataset. It has competitive performance with existing deep learning models for BAA. In order to verify the generalizability of the proposed network. Experiments are conducted on The Asian Face Age Dataset (AFAD) and achieve the best result with a mean absolute error of 3.14 years.