Coarse-to-Fine bone age regression by using multi-scale self-attention mechanism

Abstract

Pediatric bone age assessment (BAA) is a widely-used clinical technique employed to investigate various growth, genetic, and endocrine disorders in children. In this article, we propose a novel network architecture called BoGFF-Net that integrates multi-scale hand bone feature maps across different levels, and introduce an adaptive triplet loss (ATL) function that can distinguish sample pairs in regression tasks. Our network incorporates self-attention mechanisms to adaptively learn the most important regions of hand bone images, which deviates from the conventional approach of extracting specific regions in the field of bone age assessment. Additionally, we observe heterogeneous characteristics of hand bone development among different age ranges in adolescents. Therefore, we introduce a two-stage coarse-to-fine framework that can accommodate greater differences in bone modalities across diverse age groups. Quantitative and qualitative results from extensive experiments on two public bone age datasets highlight the performance and effectiveness of our model. Specifically, our model achieves competitive performance with a 3.91 mean absolute error (MAE) on the RSNA test dataset, compared to the latest model proposed by Yang et al. 2023, and a 7.07 MAE on the DHA dataset, setting a new state-of-the-art benchmark. The data and code are available at: BoGFF-Net