Research And Implementation On Convolutioanal Neural Network Based Automatic Pediatric Bone Age Estimation Algorithm

Bone age is an indicator to measure the physiological growth and development of a living body.Moreover,bone growth and development in children under 18 is more intense than in any other age group.Clinically,children patients are required to take an X-ray film from the fingertips to the wrist area of the left hand.Radiologists then determine the bone age based on the development of the ossification centers from the metacarpophalangeal bone,carpal bone,the radius,and the ulna.Bone age assessment is widely used to diagnose endocrine disorders,genetic and metabolic diseases in children.The GP-Atlas method,the TW method,and the CHN-05 method are commonly used to evaluate bone age clinically.However,these manual evaluation processes have two aspects of drawbacks: the evaluation process relies heavily on the subjective judgment of doctors,and the process is time-consuming.For a long time,semi-automatic or automatic bone age assessment has been a hot topic in medical image processing.Recently,a method of training convolutional neural networks with large amounts of data has been successful in the field of computer vision,which also provides a way forward for the bone age assessment.Therefore,this paper collects a large number of pediatric hand X-ray data from multiple sources,designs a reliable algorithm based on convolutional neural network,and conducts a comprehensive study on ossification center localization and bone age assessment issues.The main contributions of this paper are as follows:1.A simple and effective pre-processing algorithm is designed to standardize the X-ray images of hand with different sources and great quality differences,so as to improve the applicability and robustness of the algorithm.2.For the ossification center localization task,to solve some of ossification centers do not always appear in X-ray images of specific ages,this paper proposes the OC-Net which is supervised by a dual loss function.The image-level landmark classification loss function,and the pixel-level landmark localization loss function optimize the network parameters at the same time.The experimental results proved that OCNet greatly reduces the localization error of ossification center.3.For the bone age assessment task,this paper proposes a neural network BA-Net based on multi-task learning technology to simultaneously estimate bone age and locate the ossification centers of phalanges,metacarpal bones and carpal bones.The experimental results showed that the performance of bone age assessment task was improved after adding ossification center localization task.This is because the local appearance of the ossification center area is more indicative of bone development.Multi-task training is similar to the attention mechanism,which enables the network to learn more relevant features.This phenomenon proves the effectiveness of using ossification center location to improve the accuracy of bone age assessment.