| Bone Age is an important data indicator to measure the degree of bone development in children and adolescents.The evaluation of hand bone X-ray images of children and adolescents can play an important role in medicine,sports competition and related judicial fields.Aiming at the problems of irregular hand bone X-ray image data collection,slow manual evaluation of bone age,consuming a lot of human resources and large errors in the traditional artificial bone age assessment process,this paper performs relevant preprocessing on the public dataset from Radiological Society of North America with image processing technology,and then builds the corresponding bone age assessment model based on machine learning and deep learning algorithms respectively.After exploring the performance and comparison of bone age assessment under the two models,a new bone age assessment model based on improved Res Net and attention mechanism is finally obtained with stronger generalization and higher accuracy.The main research work and results are as follows:(1)To study bone age assessment methods under different machine learning models and make related improvements.In order to explore the recognition situation and training effect of bone age assessment under the training of traditional machine learning model,this paper firstly uses three machine algorithms of Support Vector Regression,K-Nearest Neighbor and Bayesian to train and test the regression task of bone age assessment.Then,based on the three independent traditional machine learning models as the basic model,different ensemble learning methods Bagging,Boosting and Stacking are used to integrate.Finally,Mean Absolute Error,Mean Squared Error and training time are used to comprehensively measure the bone age assessment effect of the machine learning models.(2)Considering that deep learning can extract and analyze advanced fine-grained features in images,this part combines existing computing resources and analyzes the convolutional neural network used in current bone age assessment.Res Net50 is finally selected as the basic network skeleton from aspects of model complexity and reliability.Then,considering the regression of bone age assessment task,Res Net50 network is improved,adding and deleting some network layers.At the same time,channel attention and spatial attention mechanism are integrated into the improved Res Net50 network in order to obtain higher accuracy and better generalization performance,as well as to consider the features of key areas of bone X-ray images.Through final analysis and comparison of experiments,the bone age assessment model based on the improved Res Net and attention mechanism achieves a mean absolute error of 6.62 months on the test set,and the accuracy rate within 12 months is 85.7%,which shows better assessment performance than several models under machine learning algorithm and the unimproved deep learning models. |
