Bone Age Assessment Based on Deep Convolutional Features and Fast Extreme Learning Machine Algorithm

Longjun Guo,Jiaqi Teng,Juan Wang,Yukun Chen

doi:10.3389/fenrg.2021.813650

Abstract

Bone age is an important metric to monitor children’s skeleton development in pediatrics. As the development of deep learning DL-based bone age prediction methods have achieved great success. However, it also faces the issue of huge computation overhead in deep features learning. Aiming at this problem, this paper proposes a new DL-based bone age assessment method based on the Tanner-Whitehouse method. This method extracts limited and useful regions for feature learning, then utilizes deep convolution layers to learn representative features in these interesting regions. Finally, to realize the fast computation speed and feature interaction, this paper proposes to use an extreme learning machine algorithm as the basic architecture in the final bone age assessment study. Experiments based on publicly available data validate the feasibility and effectiveness of the proposed method.

Highlights

In pediatrics, bone age is a significant metric to evaluate the development of child’s skeleton (Manzoor Mughal et al, 2014)
To make use of deep learning (DL)’s ability on feature learning, this paper proposes to use the CNN architectures to extract the important features from each regions of interest (ROI). 3) To further realize the fast learning speed and to improve the efficiency of DL-based bone age assessment (BAA), ELM is considered as the architecture at the last layer
This paper mainly introduces three operations required in the proposed BAA system, such as orientation correction, background removal, and ROIs selection

Summary

INTRODUCTION

Bone age is a significant metric to evaluate the development of child’s skeleton (Manzoor Mughal et al, 2014). In (Lee et al, 2017), a GP-based CNN network called BoNet was proposed to use the X-ray images of the left hand and wrist for BAA and was validated as effective in bone age prediction. Besides the advantages of DL-based models in TW-based BAA, one of the important problems with which we are always concerned is the high computation overhead, especially involving the process of learning deep features from images with back-propagation parameters tuning (Tang et al, 2021) Aiming at these issues in the DL-based BAA study, this paper proposes a new automated BAA system with fast bone age estimation speed.

RELATED WORK ON DATA PROCESSING IN BAA

Background

EXPERIMENTS AND DISCUSSION

Findings

CONCLUSION