Comparative Analysis of Bone Age Assessment Techniques Using Hand X-Ray Images and Gender Feature

Abstract

Estimated bone age is an essential measure of the growth and development of a person. As an indicator, the chronological and bone ages of the person predict the height and maturity time of the body. Statistic shows that men have greater average height than women, indicating that gender plays a role in human growth. This research aimed to utilize machine learning methods to predict bone age using hand X-ray images and analyze the effect of the presence of gender features on bone age prediction ability. Aiming at this problem, this paper proposes the use of convolutional neural network (CNN) with transfer learning methods in predicting bone age. We used VGG19, ResNet50V2, and DenseNet201 architectures as the pretrained models for comparison. The models were trained using hand X-ray images with and without gender variables. In this paper, the models are trained on a dataset consisting of 5000 hand X-ray images from RSNA2017. The images came with the corresponding gender and actual bone age ranging from 0 to 228 months old. In the experimental design, our models are divided into two scenarios. In the first scenario, we used our proposed transfer learning methods with three architectures, but gender information was not included as an input. In the second scenario, we used the same proposed models but with gender provided as an additional feature input. In this research, the models were evaluated by Mean Absolute Error (MAE) in months with a smaller MAE indicating better performance. The experimental results indicate that models added with the gender feature have lower MAE than the same models without gender feature. Models with VGG-19 architecture also have lower MAE than other architecture models. The model with VGG-19 that takes only hand X-ray image achieves an MAE of 14.544 months, whereas the same model that takes both hand X-ray image and gender feature achieved an MAE of 11.507 months.