Study On Screening Abnormal Chest Radiographs In Children Based On Deep Learning

Chest radiography is one of the main basis for diagnosis of pulmonary diseases in children.However,it is affected by factors such as incomplete development of children’s lungs and low acceptable X-ray radiation dose,the image quality of children’s chest radiography is usually not well guaranteed,which makes it difficult for doctors to interpret children’s chest radiography.Especially in some areas with poor medical resources,due to the scarcity of experienced radiologists,there is a widespread phenomenon of heavy workload and insufficient diagnostic ability of doctors,which leads to low efficiency of chest radiograph interpretation,missed diagnosis and misdiagnosis.With the increasing attention paid to children’s health and the growing awareness of early prevention and targeted screening of diseases,the number of children undergoing regular medical check-up is increasing rapidly,which also makes the above problems becoming more serious.To solve the above problems,this thesis proposes an automatic screening method for children’s abnormal chest radiograph based on deep learning technology,to help doctors prescreen children’s abnormal chest films,so that doctors can focus on the further diagnosis and treatment of children with abnormal chest radiograph.This method can effectively relieve the pressure of reading chest radiograph,improve the efficiency of doctors diagnose.The main work content of this thesis is divided into the following three parts:(1)In the construction of children’s chest radiographs dataset,it is considered that children’s chest radiographs collected in actual medical scenes are prone to the problems of high background noise and insignificant brightness difference,this thesis designs a chest radiograph preprocessing method for children,which is based on YOLOv5 s and Histogram Equalization(HE)algorithm.In this method,the YOLOv5 s model was first used to extract the lung regions in the original children’s chest radiographs,and then the HE algorithm was used to enhance the contrast of the extracted chest radiographs.According to this method,this thesis optimizes the original children’s chest radiography dataset,which effectively solves the problem of large background noise and insignificant brightness difference in children’s chest radiographs.(2)In the screening task of children’s abnormal chest radiographs,considering that the learning of effective features of the lesion area by the model is easily disturbed by other unrelated physiological and anatomical areas,this thesis designed a Dense Net＿ECA model based on ECA attention mechanism,which can learn taskrelated lesion features better and reduce the influence of unrelated physiological anatomical regions in chest radiography on model discrimination,and the screening effect of abnormal chest radiography in children is significantly improved.In order to verify the reliability of the model,a series of ablation experiments and comparative experiments were conducted.The results show that compared with other models,the Dense Net＿ECA model designed in this thesis has a better effect on the screening task of children’s abnormal chest radiographs.At the same time,this thesis also proves the effectiveness of children’s chest radiograph preprocessing method based on YOLOv5 s and HE algorithm through experiments.(3)Finally,this thesis combines the deep learning model with the frontend and backend development technology,design and implement an intelligent auxiliary screening system based on B/S architecture for abnormal chest radiography in children.Doctors can directly use the intelligent screening and other related functions provided by the system through the browser,which provides an effective way for remote diagnosis.