Research On X-Ray Image Classification Algorithm Of Pneumonia Based On Deep Learning

Objective:Since the outbreak of COVID-19 epidemic,there has been a devastating impact on the health and well-being of people worldwide.Early identification of clinical manifestations is essential to be able to initiate right preventive measures and supportive treatment in a short time to avoid possible complication.Reverse transcriptase-polymerase chain reaction(RT-PCR)detection is a major tool for COVID-19 detection and combine with imaging to better screen for COVID-19.This work proposed a two-stage network model based on deep learning to solve the problem that the X-ray images of COVID-19 and viral pneumonia are similar in imaging characteristics.In the first stage,the U-Net network with atrous spatial pyramid pooling(ASPP)was used to segment the lungs.In the second stage,X-ray images of COVID-19,other viral pneumonia and normal chest were classified by deep learning networks(Dense Net121,Res Net18)for reducing missed diagnosis and improving accuracy,which help doctors make a better diagnosis.Methods:In this work,a two-stage network is proposed,which firstly uses segmentation network to extract the lungs,and then uses classification network to classify the segmented images.The twostage network includes two processes: segmentation and classification.Firstly,the U-Net network with ASPP is adopted to achieve efficient dense feature extraction to segment the lung tissue.The structure introduces ASPP module,which effectively realizes mapping by using multiple parallel atrous convolution layers with different dilation rate.The bottom layer of the UNet network is replaced by ASPP to extract multi-scale information and thus improve the segmentation accuracy.The structure is used to realize a prediction mask for X-ray images.The lung region is extracted and interference information outside the lung region is removed to eliminate its influence on the model classification.Next,deep learning networks of Dense Net121 and Res Net18 were used to compare the training effects of the datasets with or without extracted lung regions.Results:The U-Net with ASPP model show the test set accuracy,dice similarity coefficient,and intersection over union of 99.37%,98.66%,and 97.36%,respectively.Dense Net121 network was trained with the original dataset,and the result shows that macro-P,macro-R,macro F1,and accuracy was 92.42%,92.32%,92.36%,92.32%,respectively.Dense Net121 network was also trained with the segmented dataset,gives the macro-P,macro-R,macro F1,and accuracy of94.07%,94.05%,94.06%,94.05%,respectively.Dense Net121 model using the segmented dataset compared to the original dataset showed an improvement in macro-P,macro-R,macro F1,and accuracy of the test of 1.65%,1.73%,1.70%,and 1.73%,respectively.Res Net18 network was trained with the original dataset,and the result shows that macro-P,macro-R,macro F1,and accuracy was 92.54%,92.56%,92.55%,92.57%,respectively.Res Net18 network was also trained with the segmented dataset,gives the macro-P,macro-R,macro F1,and accuracy of93.96%,93.93%,93.94%,93.93%,respectively.Res Net18 model using the segmented dataset compared to the original dataset showed an improvement in macro-P,macro-R,macro F1,and accuracy of the test of 1.42%,1.37%,1.39%,1.36%,respectively.Conclusion:Lung field segmentation is an important prerequisite for chest radiographs in computeraided diagnosis system,which precisely defines the observation regions.The results show that segmenting the lungs from chest X-ray images has an improvement on model classification.Lung segmentation can help convolutional neural network identify the main region of interest.Furthermore,it can help deep learning models make diagnostic decisions to improve classification performance and increase the reliability of computer-aided diagnosis.