Two-Stage Feature Fusion for Bone Age Assessment

Bone age assessment (BAA) is an important tool for measuring children's growth and development. However, traditional methods are limited by subjective factors and cumbersome operations. Recent advancements in deep learning have provided new approaches for BAA. In this paper, a two-stage feature fusion network (TSFFN) was proposed to address the problems of insufficient feature representation and information redundancy in key regions of hand bone X-ray images. The method utilizes region of interest (RoI) detectors to locate 11 key regions and then dynamically models within and between regions information through the intra-region feature fusion module (IFFM) and inter-region feature fusion module (InterFFM), respectively. It also incorporates a gender cross-attention mechanism to adaptively weight the features. The proposed method achieved an MAE of 4.323 months on the RSNA dataset, outperforming some advanced deep learning methods.