Research On Bone Age Assessment Method From X-Ray Images Of Pediatric Hand Bone Based On Deep Learning

Bone age is an important basis for reflecting the maturity of individual development,and is of great significance for evaluating pediatric development,disease diagnosis and treatment.Bone age assessment is usually performed using X-ray images of the non-dominant hand bone in medicine.However,the traditional methods of bone age assessment have problems such as the assessment process is too complicated,the assessment results are influenced by subjective factors,and the professional level of doctors is also required.Thus,this thesis proposed a bone age assessment method for X-ray images of pediatric hand bone to achieve high-precision pediatric bone age assessment automatically.The proposed method employed deep neural networks consisting of hand bone segmentation network and bone age assessment network.The X-ray images of the hand bone were firstly input to the improved Mask R-CNN for hand bone region segmentation,and based on the segmented hand bone,the bone age assessment was completed by the improved Xception.The main research work are as follows:(1)An automatic hand bone segmentation method based on deep learning was proposed.The hand bone segmentation network was improved based on the Mask R-CNN model,and the feature extraction network ResNet-50 was replaced with DenseNet-201.Then the hand bone X-ray images were segmented and fused with the original images to obtain the hand bone Xray images with the hand bone region and background separated.(2)An automatic bone age assessment method based on deep learning was proposed.The bone age assessment network was improved in three aspects based on the Xception model.Firstly,a pooling method SoftPool that can replace any pooling operation in the network was introduced,so that it retained more effective information in the reduced feature map activation map.Secondly,a residual attention module that can be independently embedded in the network was proposed,which enabled the network to enhance feature perception from two different dimensions,channel and space,and extracted more representative features from hand bones.Thirdly,gender feature was added to the network and worked together with image features to reduce the impact of gender differences on bone age assessment results.(3)An intelligent auxiliary assessment system for pediatric bone age that integrated the hand bone segmentation and bone age assessment algorithm was developed.Using PyQt5 and PyInstaller to encapsulate the model into a human-computer interaction software on PC and using Android Studio to encapsulate the model into a mobile human-computer interaction software.Model performance was evaluated on the RSNA public dataset.The mean absolute error of bone age assessment was 4.68 months,which was a 42.86% reduction compared to only using Xception to assess bone age.The accuracy of bone age assessment error within 1 year was 94.4%,which was 11.2% higher than that of only using Xception to assess bone age.By segmenting the hand bone region,replacing the pooling operation as SoftPool,adding the residual attention module and synthesizing gender feature,which can effectively reduce the error of pediatric bone age assessment.The method proposed in this thesis can provide an efficient and accurate automatic assessment of pediatric bone age for clinical practice.