Bone Mature: A Deep Learning Framework for Bone Age Detection

Abstract

Bone age estimation is one of the most important processes in paediatric radiology which is used to determine the level of skeletal development and identify various developmental problems. Hand radiograph interpretation is manually used in the estimation of bone age, but this method is rather subjective and inconsistent. In this paper, the author presents the Bone Mature, a two-step deep learning system to estimate bone age without the need of manually annotated region of interest. During the first step, an annotation-free cascaded critical bone region localization model is trained using InceptionV3 with an extra convolutional block attention module (CBAM) and gradient-weighted class activation mapping (Grad-CAM). In particular, carpal and metacarpal-phalanx bones, which are anatomically important, are automatically detected in hand radiographs by the proposed cascaded critical bone region localization model. The second step involves the localized areas being inputted into ResNet50 and Xception models with mid-level feature fusion and gender-assisted estimation.