The Research Of Bone Age Assessment Method Based On Convolutional Neural Networks And Multi-scale Feature Fusion

Bone age assessment infers the actual growth of children through bone development morphology,which has various clinical applications such as diagnosis of endocrine disorders and prediction of final adult height for adolescents.According to the development process,bone age assessment methods can be divided into three parts: the methods based on artificial estimation,the methods based on traditional machine learning and the methods based on deep learning.The first two methods are very dependent on the experience of clinicians and the design of manual features,which has a great constraint on the popularity of methods application.With the development of deep learning technology,its powerful ability of data expression has attracted people's extensive attention.Bone age assessment methods based on deep learning have become an active direction in the field of medical image analysis.Because the bone age assessment is a recognition task about small data set,it can't provide rich training samples for neural network.Therefore,how to obtain the qualified network model in the case of small samples becomes a serious challenge in the current development.In view of the problems in bone age assessment task,this dissertation proposes a multiscale data fusion framework of X-ray image based on the combination of non-subsampled contourlet transform and convolutional neural networks.The existing bone age assessment methods based on convolutional neural network directly use the original spatial image as the network input data,which require a large number of spatial images as training samples.In this dissertation,the input image is preprocessed by performing the non-subsampled contourlet transform,and the rich feature sets are extracted to obtain its multi-scale and multidirectional representation.Because the number of annotated samples in bone age assessment task is typically quite limited,taking these features as training data of network can alleviate the problem of insufficient training data to a certain extent.The obtained non-subsampled contourlet transform coefficient maps at each scale are fed into a convolutional neural network model individually and the information from different scales are then merged to achieve the final prediction.For the design of the network model,this dissertation uses VGGNet as the basic structure of the model,and sends information of different scales into VGGNet respectively,forming a complex network model composed of multiple VGGNet branches.According to the different fusion strategies of model branches,this dissertation designs two convolution neural network models: a regression model with feature-level fusion and a classification model with decision-level fusion.At the same time,in order to determine the optimal model structure,the factors such as the number of non-subsampled layers and decision-level fusion rules are analyzed in detail.Experiments on the public bone age assessment dataset Digital Hand Atlas demonstrate that the proposed method can obtain promising results and outperform many state-of-the-art bone age assessment methods.In particular,the proposed approach exhibits obvious advantages over the corresponding spatial domain approaches(generally with an improvement of more than 0.1 years on the mean absolute error),showing great potential in the future study of this field.