Study On Intelligent Diagnosis Method Of Adolescent Bone Age Based On Convolutional Neural Network

Bone age assessment based on human wrist X-ray images is critical during the diagnosis of growth disorders in children.As a subjective assessment,the accuracy is highly depended on physician clinical experiences.Convolutional neural networks have developed rapidly in recent years and have been widely applied in medical fields,such as medical image analysis,medical image processing.A novel bone age assessment is proposed based convolutional neural networks,which can be realized hand type extraction,hand key bone extraction and segmentation,and key bone score classification.The main work in this thesis is as follows:(1)Dataset construction: The datasets during the bone age assessment datasets is built based on images from RSNA and acquired images from local hospital.Spatial domain transformation and position transformation are applied to X-ray images in the above datasets,which makes the image more clearly with obvious features.(2)The hand segmentation model establishment: A light size hand segmentation networks CSP-Unet is proposed based on CSPNet and Unet.The segmentation results show that the network model size is reduced 70% with similar accuracy,compared with the original Unet.(3)The key bone extraction and segmentation models establishment: A fast-fitting key bone extraction network OSA-YOLOv5 is proposed based on YOLOv5 and OSA.The training speed of OSA-YOLOv5 extraction network is 28% higher than that of YOLOv5.A multimodal segmentation network GRU-Unet containing spatial sequence features are proposed,and efficient segmentation of key bones is realized.The segmentation accuracy of the multi-mode GRU-Unet segmentation model is 13% higher than that of Unet,and validated in experiments.(4)The key bone classification and bone age evaluation models establishment: The multimodal key bone classification network MRL-Efficient Net is proposed.The average classification accuracy of MRL-Efficient Net is 87%,which is high compared with other baseline networks.The classified key bones were scored based on the TW3-C RUS standard percentile curve and the bone age was obtained.The experimental results show that an average difference of 0.34-0.35 years between clinical and automated bone age assessment based on the proposed methods.The automated bone age assessment proposed in this study have improved performance,which provided a solid foundation for further development of intelligent bone age evaluation system.