The Research On Automatic Bone Age Assessment Based On Ensemble Learning Of Stacking

Human organs has shown phases in growth and development,which is often used to estimate biological age.The most common method to evaluate the age of children and adolescents is to use hand and wrist image to detect the biological age.Bone age assessment(BAA)is widely used in forensic medicine,kinematics and pediatric radiology,which helps clinicians to diagnose and treatment dysgenesis,endocrine abnormalities,genetic disorders,and plan orthopedic surgery in children based on differences in bone age and chronological age.At present,BAA is major traditionally accomplished by manpower with the problem of complex diagnostic process,strong subjectivity,random errors and advanced technical talents requirements.To overcome the limitation of traditional methods,BAA based on computer aided technology has become a research hotspot.However,most of BAA systems have a major shortcoming of using incomplete age coverage of the datasets,and unsteadily estimated results,so that automated robotic BAA systems are not widely applied.Aiming at the problems existing in BAA system,we proposed to construct the BAA model based on left hand-wrist X-ray images datasets outside a children’s hospital in China and the open source data set abroad,combining the ensemble learning of stacking strategy and deep convolutional neural networks(DCNNs).The main contributions of this work are as follows:(1)We proposed a hand-wrist segmentation method using DBSCANU-Net based on feature extraction.The hand area was detected by feature points firstly,and then the whole hand was roughly segmented by DensityBased Spatial Clustering of Applications with Noise(DBSCAN)algorithm,as a result of most background could be removed.Moreover,the coarsesegmented images were input into the U-Net for training and prediction for the purpose of explicit segmentation,a clean hand mask could be extracted by a simple post-processing.Experimental results show that the proposed method is more efficient and accurate than the method of only using U-Net network to segment hand-wrist images.(2)We proposed a framework using stacking-based DCNNs to construct the BAA model.Five DCNNs with better performance among the commonly used BAA models were selected as the basic models,and softmax regression was chosen as the secondary model,stacking strategy of ensemble learning was used to train this multiply framework.Furthermore,the loss function was improved to optimize the prediction error caused by the unbalanced datasets.Experiments conducted on two datasets show that the performance of the BAA model based on the stacked multiple DCNNs was improved by 7%-15% compared with the model constructed by a single DCNN,MAE and RMSE on two datasets were 5.42 and 6.15,7.81 and 8.94,respectively.More importantly,the prediction accuracy of the BAA model constructed by the segmented images was 25% higher than that of the model trained by images without segmentation.(3)We designed and developed the Intelligent Bone Age Assessment System(IBAAS)based on the constructed BAA model,and meanwhile,set up the server and client access platform.The system can automatically predict the bone age through the hand-wrist X-ray image input by users,which increases the clinical application value of the model.The hand bone image segmentation method adopted by IBAAS can improve the accuracy of segmentation and provide reliable pre-processed datasets for the construction of BAA model,effectively improve the prediction accuracy of the model.In addition,the BAA model constructed by IBAAS based on stacking strategy of ensemble learning can accurately predict the bone age of children and adolescents.Compared with the single DCNN based BAA model,although there is a higher time complexity of the multiple DCNNs ensemble,leads to a stronger model generalization ability and a more accurate prediction.IBAAS developed based on this BAA model can be used as an auxiliary tool for clinical application,to provide decision support for radiologists.