Automatic Bone Age Assessment Based On Attention Region Localization Algorithm

Bone age assessment plays an important role in the selection of young athletes,clinical medicine,forensic medicine and other fields.The doctor evaluates the bone age by viewing the X-ray image of children’s hand bone,and then compares the bone age with the actual age,the growth and development of children can be judged by the difference between the two.Clinical bone age assessment is usually carried out manually,but this process takes more time and requires professional clinical experience.Many assessment methods based on deep learning directly input the hand bone X-ray image into the neural network for training,but ignore the local fine-grained features of the hand bone image and the edge noise of the X-ray image.Some researchers also propose to train segmentation or detection network to cut important region of hand bone before evaluation,but such methods need additional manual annotation and can not be trained end-to-end.Based on the foreign published bone age data set and the data set provided by a domestic children’s Hospital,this thesis proposes a bone age evaluation method to locate the key parts of hand bone based on the high response value area(Attention Region)in the process of attention training,which effectively improves the accuracy of bone age evaluation and reduces the complexity of training.The main work of this thesis is as follows:1.Aiming at the problem that additional manual annotation is needed to locate the hand area,a multi-stage bone age evaluation method based on CAM(Class Activation Map)is proposed.Firstly,the classification network is trained to generate the thermal map of hand bone X-ray image,and the hand area most concerned by the network model is cut out on the thermal map,so as to realize the hand area location without manual annotation.Input the cropped image into Xception＿CBAM with attention mechanism completed bone age assessment.This method effectively improves the evaluation accuracy,and proves that paying attention to the important area of hand bone image has an important impact on the evaluation results of bone age.2.In order to make the evaluation process pay more attention to the local fine-grained characteristics of hand bone,an end-to-end bone age evaluation model AHPR-Net(Attention Hand and Parts Region Networks)based on attention region localization is proposed.AHPR-Net first uses the selective convolution descriptor aggregation method to locate the hand region,then locates the attention local region with high discrimination characteristics based on the principle of region generation network,and then diversifies the positioning part and bone age evaluation part into an overall network framework by sharing the backbone network and full connection layer to realize end-to-end training.After the model automatically locates the hand area and the local area containing fine-grained features,the feature information of these areas is fused to improve the evaluation effect.The absolute errors on public data sets and private data sets are less than 6.02 months and less than 4.33 months respectively,which meets the application requirements of clinical testing.Finally,after training the bone age evaluation model based on the domestic data set,an automatic bone age evaluation system is developed.After inputting the X-ray image of children’s hand bone,the user can get the evaluation result of bone age,realize the function of computer-aided medical diagnosis,and have important clinical application value.This study completes the bone age assessment based on the method of attention region positioning,which effectively reduces the training cost and improves the assessment accuracy.The developed bone age assessment system can be used as an auxiliary tool for clinical detection and reduce the workload of doctors.