X-ray Bone Age Regression under Decentralized Data Silos

Abstract

Bone age assessment from hand X-ray images is an important tool for evaluating pediatric skeletal maturation and supporting clinical decisions in growth-related disorders. Although deep learning has substantially improved automated bone age regression, further progress increasingly depends on multi-institutional data collaboration. In practice, however, bone age radiographs are often distributed across decentralized institutional silos, where privacy and governance constraints limit centralized data sharing. Existing studies on federated medical learning have paid limited attention to bone age assessment tasks.This paper aims to investigate the challenges of applying federated learning methods to bone age assessment tasks, and to develop an algorithm specifically tailored for this task. Three publicly available bone age datasets, RSNA 2017, RHPE, and DHA, were used to construct a realistic federated regression setting, with each dataset treated as an independent client. A second heterogeneous setting was further simulated on DHA using Dirichlet-based client partitioning to model label-distribution heterogeneity. We use MobileNet as the backbone network for bone age feature extraction. Meanwhile, we also explore the effectiveness of several data augmentation strategies in the federated learning setting.In the cross-dataset federated setting, the proposed method consistently outperformed FedAvg on all three clients and on the overall benchmark. Overall, MAE decreased from 28.295 to 25.971, $R^2$ increased from 0.270 to 0.396, and GM decreased from 18.143 to 16.945, corresponding to relative gains of 8.215\%, 46.648\%, and 6.603\%, respectively. In the Dirichlet-based heterogeneous setting on DHA, MAE was reduced from 34.517 to 31.634, $R^2$ improved from 0.390 to 0.495, and GM decreased from 22.653 to 21.183. Additional analyses showed that the proposed method produced a clearer structure in UMAP visualizations and a smoother convergence trajectory across communication rounds than FedAvg.