Research On X-ray Pneumothorax Analysis And Image Segmentation Based On Deep Learning

X-ray film can show the density distribution of various parts of the human body,so it is widely used in the field of medical diagnosis.Whether you go to the physical examination regularly or to the hospital,you need to take an X-ray chest film.A large number of X-rays have brought heavy workload to doctors,which inevitably leads to a long waiting for treatment.Some patients with more serious conditions may not get timely treatment due to long-time waiting,resulting in irreversible conditions.From the doctor’s point of view,diagnosis is a job that requires a lot of energy.If the working time is too long,it is inevitable that the doctor’s mental state will be bad,which will affect the diagnosis of the disease.Therefore,in the field of medical image analysis,the demand for computer-aided diagnosis system which can accurately judge the disease is very urgent.This paper mainly studies the accurate lesion segmentation of X-ray pneumothorax based on deep learning method,hoping to realize the automatic detection and location of pneumothorax.The research contents are as follows:(1)There are two main problems in the existing X-ray pneumothorax detection methods: first,the position of pneumothorax often overlaps with ribs,coracoid bone,clavicle and other parts,resulting in great missed diagnosis in clinical diagnosis;Second,the existing mainstream segmentation algorithms adopt single or double threshold strategy,which is easy to lead to the segmentation effect is not fine.This paper presents a pneumothorax segmentation method based on improved U-Net architecture.Firstly,the contrast limited adaptive histogram equalization is applied to the chest film to remove the noise and restore the image details;The abstract deep features in the image are extracted through the convolution neural network layer with MBConv Block as the encoder module;Then,the extracted feature map is interpolated and reconstructed by the decoder to obtain the binary classification result of each pixel;Finally,the improved triple threshold strategy is adopted to output the results that more meet the actual medical scene.The experimental results show that this method can make the X-ray pneumothorax segmentation have high accuracy,and has the advantages of simple training process and easy to use,which fills the shortcomings in the field of X-ray pneumothorax image segmentation at present.(2)In order to solve the problems of low segmentation accuracy of pneumothorax due to lack of labeled data and weak generalization ability of supervised learning model,a semi supervised learning algorithm PTXSS based on pseudo label is proposed to improve the segmentation accuracy of pneumothorax.For the subsequent mixed cycle training,a more applicable network MNA-Net is proposed;The exponential moving average strategy is used to update the parameters of the teacher model,and noise is added to the mixed cycle training to further improve the generalization ability of the model;In order to improve the data utilization in the training process,the fixed threshold is changed to the dynamic threshold,which can effectively improve the reliability of the pseudo tag and the performance of the model.