Automated Bone Age Assessment Based On X-ray Images And Convolutional Neural Network

Bone age assessment is of great significance for the assessment of adolescents’ physical development and can be used to guide adolescents in exercising and eating properly.The most commonly used clinical method of bone age assessment is to visually compare the bones of the left hand with the standard hand diagram to evaluate the bone age.This kind of methods are very subjective and their accuracy depends largely on the experience of the practitioner.Therefore,it is very important to develop an efficient,stable and objective method for assessing bone age.Researchers at home and abroad have long carried out related research on the automation and computerization of bone age assessment tasks,have proposed various constructive ideas,among which convolutional neural networks have achieved the best results.However,there are still some problems in the current research.Firstly,in order to achieve higher accuracy,many researchers have introduced image segmentation networks and target detection networks into bone age assessment tasks.These methods do improve the accuracy,but also make the entire network very complicated,which is not conducive to the automation of bone age assessment tasks.Secondly,many researchers use convolutional neural networks to extract high dimension features of hand X-ray images,and then use traditional machine learning algorithms to predict bone age.There are two shortcomings in those ideas.One is that there is no convolutional neural network designed for bone age assessment tasks,and the other is that there is no feature selection for the extracted high dimension features.Thirdly,no researchers tried to compress these convolutional neural networks used to assess bone age.In view of the above problems,this thesis proposed two ideas.The first scheme improves the feature expression ability of the convolutional neural network,without resorting to the target detection network and the image segmentation network.The second scheme,on the premise of ensuring the accuracy of bone age assessment results,compresses the network.Therefore,this thesis introduces attention mechanisms and knowledge distillation technology into the task of bone age assessment.The main work and contributions of this thesis are as follows:(1)This thesis designs bone age assessment network based on the attention mechanisms.First,a basic convolutional neural network is built for assessing bone age.Then,these attention mechanisms are encapsulated into independent modules.Finally,the basic convolutional neural network embedded in these modules is used to evaluate bone age.The purpose of attention mechanisms is to replace the segmentation network and the detection network,and make the bone age assessment network automatically focus on important parts in the image,and weaken the background interference.The experimental results show that compared with the segmentation network,the attention mechanism has the similar performance,but has fewer parameters.(2)This thesis designs bone age assessment network based on attention mechanisms and support vector regression.Here,attention-based bone age assessment network is used as a feature extractor to extract high-level features.Principal component analysis and support vector regression algorithms are introduced to process these high-level features.Experimental results show that the network greatly reduces the error,compared with the attention-based network.(3)This thesis designs bone age assessment network based on knowledge distillation technology.First,a small convolutional neural network is built for assessing bone age.Then,the bone age assessment network based on the attention mechanism assists the mall network training to improve the performance of the small network.Experimental results show that the network is not only small-scale,but also has good performance.In summary,the contribution of this thesis lies in introducing attention mechanism and knowledge distillation technology to the bone age assessment task for the first time,which provides different ideas for subsequent researchers.