A Lightweight U-Net Architecture Multi-Scale Convolutional Network for Pediatric Hand Bone Segmentation in X-Ray Image

Abstract

Bone age assessment (BAA) is a common radiological examination used in pediatrics based on an analysis of ossification centers and epiphyses of hand bones. Segmentation of hand bones could help give specific descriptions of hand bone features in medical records and assess bone age automatically. This study proposes a lightweight U-Net architecture multi-scale convolutional network for pediatric hand bone segmentation in the X-ray image. The compact structure is based on U-Net architecture with two down-sampling and up-sampling operations and multiple filters with different kernel size are adopted for countering hand bone scale variations during growth in children. This is the first-hand bone segmentation study with deep learning and the experiment results indicate promising performance in hand bones segmenting, especially for small bones of the hand.