Research On Convolutional Neural Network-based Automated Bone Age Assessment Method In X-ray Images

Bone age is one of the main criteria for biological age,and it can accurately reflect the growth and development of the children.However,artificial estimation of bone age takes a long time and the results are highly volatile,greatly relying on the proficiency of the radiologist.Therefore,assessing bone age automatedly via computer vision tech-niques is necessary,which is one of the research hotspots in the medical image field.The bone age automated/semi-automated assessment method based on traditional machine learning algorithm mainly assess bone age from X-ray images by segmenting key areas and manually extracting features.However,its assessment accuracy heavily depends on the accuracy of the segmentation and the validity of the extracted features.The bone age assessment system based on deep learning solves the above problems well.It learns to extract features from the data end-to-end without the need to segment and manually extract features,which has attracted researchers,attention.This paper mainly proposes a fully automated bone age assessment method for X-ray images.Firstly,this study proposes a bone age assessment model BoNet+based on dense connection convolutional neural network.Besides,the performance of different cost functions in bone age assessment problem has been studied.It is found that the mean absolute error function is the most suitable cost function for automated bone age assessment problem.Experimental results show that the assessment error of the BoNet+model proposed in this paper is 0.76±0.10,which is much smaller than the bone age assessment method based on traditional machine learning algorithm and slightly better than the VGG-based convolutional neural network model BoNet.Moreover,its per-formance of training speed and parameter number are much better than BonNet,which means that the complexity of BoNet+is much lower than that of BoNet.Then,considering that there may be low-quality X-ray images in real-world scenar-ios,this paper simulates three types of low-quality X-ray images:images with Poisson noise,corrupted images with label,and low-resolution images.In order to reduce the negative effect of image quality on bone age assessment,this paper proposes to use a U-Net-based convolutional neural network to improve the quality of low-quality images.According to whether the type of low-quality image is first classified and then improved,this paper divides the real-world bone age assessment system into two frameworks.Ex-perimental results show that the assessment accuracy of them is similar regardless of whether the classification model is adopted.To explain this phenomenon,this paper proposes that if the fitting ability of the convolutional neural network-based model is strong enough by deepening and widening the layers,multiple tasks can be processed simultaneously only using a single model.