Coarse-to-fine segmentation and ensemble convolutional neural networks for automated pediatric bone age assessment

Bone age assessment is commonly used to estimate bone maturity and diagnose disease in children and adolescents. It is significant to develop intelligent bone age estimate methods for clinical applications. In this paper, we propose a novel two-stage automated pediatric bone age assessment method, which cascades a coarse-to-fine hand segmentation method and stacking-based ensemble convolutional neural networks (CNNs). Specifically, we use density-based spatial clustering of applications with noise algorithm to extract coarse feature points of interest from preprocessed hand radiographs. Afterward, we construct a U-Net to fine hand masks, which facilitates feeding clean pediatric hand bone radiographs into the following ensemble CNNs classifier to produce accurate skeletal age estimate. Considering the complementary feature extraction capabilities of different networks, we design an ensemble learning model by stacking five powerful CNNs to improve bone age assessment accuracy from clean hand radiographs. We evaluate the proposed two-stage bone age estimate method using both public and clinical datasets. Experimental results show that our method outperformed existing state-of-the-art skeletal age determination methods and respectively achieved the mean absolute error of 5.42 and 5.98 months on males and females.