Research On Deep Learning-based Automated Bone Age Assessment Method In X-ray Images Of Hand Bone

Bone age assessment is a common clinical method to study endocrine,genetic and growth disorders in children.The traditional manual detection method of doctors is timeconsuming and laborious,and is affected by the doctor’s subjective factors,and there are operator errors.Existing automatic bone age detection methods based on automatic extraction of clinical features also have the problem of low accuracy and difficulty in generalization due to inaccurate feature extraction.The bone age evaluation method based on deep learning automatically learns the characteristics of the data through the multilayer network structure,showing a good evaluation situation,but most of the existing methods are non-end-to-end and focus too much on the feature area of the clinical method and ignore the convolution Feature extraction by neural network.Aiming at the limitations of existing bone age assessment methods,this thesis proposes an automatic bone age assessment method based on deep convolutional neural networks.The main research contents and innovations are as follows:（1）Fine-tuned deep convolutional neural networks Inception v3,Inception v4,Inception Resnet v1,Inception Resnet v2 models are trained from scratch on the public data set The Digital Hand Atlas（DHA）and applied to bone age assessment;mean absolute error（MAE）is not higher than 0.7 years old,which proves the superiority of the Inception module in the field of bone age assessment;（2）An Inception bilinear convolutional neural network based on the attention mechanism is proposed.The attention mechanism module and bilinear structure are introduced on the basis of the Inception module.At the same time,attention is paid to the bone age-related areas and visual appearance of the hand bone X-ray image characteristics.Using the improved stratified 5-fold cross validation method on the DHA data set,the optimal result MAE is 0.27 years old;（3）Visualize the feature regions extracted by the deep convolutional neural network,and compare the results with the regions of interest of traditional artificial methods;emphasize the advantages of the convolutional neural network itself for feature extraction in the automatic bone age assessment.The experiments in this thesis on the public bone age assessment data set DHA show that the proposed method has obvious advantages compared with the deep learning methods evaluated on the same data set,and the MAE has generally decreased by more than 0.3 years.