Research And Application Of Bone Age Assessment Technology Based On Deep Convolutional Neural Network

Bone age is an explanation of bone maturity,which can determine the biological age of human beings and judge if a child is precocious or retarded.It has been decades since the development of bone age application.Bone age assessment not only plays an important role in the clinical practice of Pediatrics,but also is extended to various fields.At present,many professional doctors still assess bone age with traditional methods,such as scoring method,atlas method and so on.These methods are time-consuming and laborious,easily interfered by subjective factors and difficult to be popularized.Before the rise of deep learning,some traditional image processing methods were applied to assist doctors in order to overcome the problems of artificial bone age assessment.But they did not fundamentally solve the problem.With the development of deep learning theory and related technology,bone age assessment technology based on deep convolutional neural network has become the mainstream research direction.The paper studies the bone age assessment technology based on convolutional neural network.The quality of hand images is uneven.Some noise,background information and hand bone posture seriously affect the accuracy of bone age assessment.The preprocessing of hand image based on deep convolutional neural network is studied in this paper.A hand segmentation method based on DensUNet and a hand posture correction method based on ResNet152 are proposed.The hand image with main background separation and consistent hand posture is obtained.On this foundation,the feature extraction of bone age based on deep convolutional neural network is studied.By integrating the Squeeze-and-Exception module,some attention modules and gender features into the backbone,a bone age assessment model based on improved InceptionV3 is proposed.Multiple experiments based on open data sets are designed in this paper,which verify the effectiveness of image preprocessing technology and bone age assessment model.Then,hand heat-maps are generated with Grad-CAM method.Finally,an automatic bone age assessment system is designed and implemented to verify the practicability of the research work.Specifically,the main research contents of this paper include the following aspects.(1)Based on the review of traditional bone age assessment methods and image-based bone age assessment methods,this paper summarizes the research status of bone age assessment methods based on deep learning,analyzes the problems to be solved in bone age assessment based on deep convolutional neural network,and obtains the research content of this paper.(2)To eliminate interference information in the hand images,such as noise,image scale and so on,this paper studies the image preprocessing technology.A hand segmentation method based on DenseUNet is proposed.The traditional cross entropy loss function is discarded in this model,and instead,we optimize the dice function,which is used as the loss function of the model for gradient update,so that it can segment hand more precisely.A hand posture correction method based on ResNet152 is proposed to solve the problem of uneven hand posture due to some factors such as shooting angle.Using the above image segmentation and posture correction technology,the hand bone is separated and corrected to a unified posture,which provides the basis for bone age evaluation.Preprocessing lays a foundation for improving the accuracy of automatic bone age assessment model.When training the above models,we do not have training labels,so iterative training method is applied to solve the problem of time-consuming manual annotation.(3)The Squeeze-and-Exception(SE)module is integrated into InceptionV3.The network is combined with Position Attention Module(PAM)and Channel Attention Module(CAM)to extract bone age features.An automatic bone age assessment model based on the improved InceptionV3 is proposed,into which the gender features of patients are effectively integrated.(4)The data set used in this paper is hand radiography provided by the Radiological Society of North America(RSNA).We evaluate the accuracy of bone age from three aspects,which include the data set before and after preprocessing,the model before and after integrating gender features,and the model before and after adding multiple attention modules.The experimental results verify the effectiveness of the proposed method.According to the hand heat-maps,we can summarize the changes of the areas that network focus on when the model assesses bone age,which provide help for understanding the relationship between bone age and hand image.(5)This paper integrates the DenseUNet model,the ResNet152 model and improved InceptionV3 bone age assessment model.An automatic bone age assessment system based on deep convolutional neural network is designed and implemented,and the feasibility of using related technologies in the actual system is explored.