Research Of X-Ray Bone Age Assessment Based On Deep Neural Networks

Bone Age Assessment(BAA)is a pediatric examination by comparing the differences between children's bone age and their chronological age.Comparing to the chronological age,either the bone age is too older or too younger often reflects hormone problems and other abnormalities in children's growth process in some extend,bone age is also used to predict children's future height and so on.The main work of the BAA is predicting the bone age of Children,for predicting bone age accurately and efficiently,a mature BAA method is very necessary.The method of BAA has gone through a long period of development,previous works had upgraded the BAA from traditional techniques to the more accurate expert system.But the problem is also obvious,the existence of human assessment process will make the final results filled with subjectivity and sariability,and wasting time.The coming of the deep learning age makes the more accurate and more efficient BAA method possible.Deep Neural Network Models have a high requirement of samples,but original X-ray data of hand bones exist the trouble of too much noise information and the number of samples are unbalanced between different classes,these troubles will influence the final assessment result of the models.To solve the above problem,this paper explores the method of BAA based on deep learning.The main research contents are as follows:1.To solve the trouble of too much noise information,this paper proposes a X-Ray hand bone region extraction method based on deep neural network.This method mainly includes fixing the location of hand bone,sliding window sampling for X-Ray of hand bone,sample classification and segmentation which is accurate to pixels.Not only removes the noise information of original X-Ray of hand bone,this method also ensures the unification of hand bone position in all image data.2.To eliminate the human intervention,this paper proposes a Bone Age Assessment method based on deep neural network.This method mainly includes different sample classification style,different deep neural network model,classification for X-Ray of hand bone and evaluation of experimental results.As for original X-Ray of hand bone data,this paper finds the appropriate sample classification style which solves the trouble of unbalanced numbers of samples between different classes.Analysis of experimental results use three indicators: absolutely right,one-year error and two-year error.The best model assessment results are 0.507,0.851 and 0.964.3.Images of human hand bone show gradual changes,the smaller the bone age gap is,the smaller the difference between images of hand bone.Aiming at this characteristic,this paper proposes a BAA method based on distance measure neural network.This new method uses Triplet loss function to train deep neural network firstly.Based on this network,three means are used to achieve BAA(Feature Base Matching,SVM and Feature Vector Reclassification).The average value of three means assessment results are 0.542,0.843 and 0.969.This paper mainly explores the method of BAA in many ways,it is of great significance to the maturity of bone age assessment method.