Research On The Assisted Diagnosis Method Of Pneumothorax Disease Based On Deep Learning

Pneumothorax is a common acute lung disease,and rapid and accurate diagnosis and treatment of pneumothorax disease can help reduce the harm of the disease.At present,the mainstream pneumothorax auxiliary detection mainly adopts a single-stage image segmentation algorithm,but the pneumothorax image has the characteristics of uneven distribution,large changes in lesion shape and size,and in obvious features,and the single-stage image segmentation algorithm is easy to cause the problem of missing detection of pneumothorax disease,resulting in a decrease in accuracy.Therefore,it is of practical significance to solve the problem of missed detection of pneumothorax disease,improve the accuracy of pneumothorax disease judgment in deep learning model,and develop an auxiliary diagnosis system for pneumothorax disease to assist doctors in diagnosis.The main work and innovation of this article are as follows:(1)In response to the current single stage pneumothorax disease segmentation auxiliary diagnosis network model which has the problem of missing lesions in pneumothorax images,this study proposes a two-stage pneumothorax disease auxiliary diagnosis algorithm.The first stage is the image classification algorithm,which first ensures the accuracy of positive case judgment and sends positive cases into the segmentation algorithm.At the same time,negative cases are screened as much as possible,and the screened negative cases directly output the results.The second stage is the pneumothorax lesion segmentation algorithm,which extracts the positive cases screened in the first stage for lesion segmentation.Compared to the single stage segmentation algorithm,the two-stage algorithm can reduce the difficulty of image segmentation tasks,making the segmentation network model focus on lesion extraction,in order to solve the problem of low accuracy in assisting diagnosis of pneumothorax diseases in the single stage algorithm.The experiment shows that the two-stage pneumothorax disease auxiliary diagnosis algorithm achieved an accuracy of 88.13%,which is 1.98% higher than the single stage method.(2)A pneumothorax disease classification algorithm based on EfficientNet is proposed.Firstly,in this study,the SIIM-ACR Pneumothorax segmentation dataset was reconstructed into a categorical dataset.At the same time,Focal loss is used for training to increase the weight of positive cases and ensure the accuracy of positive case judgment.In terms of network model,on the basis of EfficientNet-B2 network,the Atrous Spatial Pyramid Pooling module is added to enable it to obtain a wider Receptive Field,and the SE module in the original MBConv structure is replaced by CBAM module,which can make the network pay more attention to the characteristics of small targets,so as to solve the problem of inaccurate judgment of positive cases by convolutional neural network.The experiment showed that adding only the Atrous Spatial Pyramid Pooling module improved the accuracy by 0.9%,and using only the CBAM module improved the accuracy by 0.7%.By applying two improved methods simultaneously to the model,the accuracy of the improved EfficientNet-B2 network was improved by 2.1%,reaching 87.3%.Meanwhile,the detection rate of positive cases in the test set was 90.1%,an increase of 2.9%.The improved EfficientNet network is more accurate in determining positive cases and screening negative cases more accurately.(3)A new segmentation algorithm for pneumothorax disease is proposed based on U-Net3+ network.In this study,the pneumothorax disease data set was reconstructed first,and the negative cases in the original data set were eliminated to make the training data set conform to the actual scenario.Secondly,in terms of network structure,the Residual Module and the improved CA attention mechanism are added to the U-Net3+ full-scale feature fusion network structure,which increases the encoder’s ability to extract richer texture information,so that the network can fully mine the detailed feature information of the image and ensure the accuracy of lesion segmentation.Mix Loss,a combination of Dice,Focel and BCE,is used in training to accelerate model convergence.Experiments show that the improved U-Net3+ network has a Dice coefficient of 89.97% and an IOU coefficient of 82.77%,which effectively solves the problems of unclear segmentation boundary and inaccurate segmentation area in the current pneumothorax disease segmentation task.Finally,by deploying the two-stage neural network model,the development of the auxiliary diagnosis system for pneumothorax disease was completed.This system has the function of logging in,which can perform auxiliary diagnosis of pneumothorax disease on local chest X-ray,and perform lesion segmentation and extraction according to the auxiliary diagnosis results,and store the results in the local designated path,the basic function of auxiliary diagnosis of pneumothorax disease is realized.