Research On COVID-19 X-ray Image Recognition Method Based On Deep Learning

COVID-19,which began in 2019,is still spreading around the world.Early identification,isolation and treatment of patients is considered a key tool in the fight against the epidemic,and an accurate,rapid and low-cost method to identify infected patients is urgently needed.Based on the practical needs,this thesis studies the COVID-19 recognition and diagnosis from the three aspects of COVID-19 feature characterization,semantic segmentation of lung soft tissue and location of COVID-19 positive patients,proposes the COVID-19 X-ray image recognition algorithm based on deep learning,and completes the design and implementation of the COVID-19 recognition and diagnosis system.The main research work and innovation points of this thesis can be summarized as follows:1)The characterization effect of COVID-19 was studied.Aiming at the recognition problem of COVID-19,this thesis compared and analyzed the features of COVID-19 extracted from different types of image classification networks,and discussed the advantages and disadvantages of different networks from the dimensions of recognition accuracy,recognition speed and learning speed.2)Improved Inception V3 network and proposed COVID-19 X-ray image recognition algorithm based on Mobile-Inception V3 network.The improved network has the advantages of strong generalization ability and fast learning speed.Combined with the image preprocessing algorithm based on multi-algorithm fusion proposed in this thesis,the experimental results on several COVID-19 datasets of Kaggle show that the accuracy of this algorithm reaches 98.77%,and it only takes 16.37 seconds to recognize 7200 images,which is better than the comparison algorithm.3)The semantic segmentation effect of lung soft tissue was studied.Aiming at the problem of lung soft tissue segmentation,this thesis compares and analyzes the lung soft tissue images segmented by different semantic segmentation networks,and discusses the advantages and disadvantages of different semantic segmentation networks from the aspects of segmentation cross ratio,segmentation speed and learning speed.4)The localization effect of COVID-19 positive patients was studied.In this thesis,for the localization of COVID-19 positive patients,the Grad-CAM feature visualization technology was used to carry out the weak-supervised segmentation experiment,and the method was improved according to the medical prior knowledge.The method has a better localization effect of the lesion location.5)Based on the algorithm work,the COVID-19 recognition and diagnosis system is designed and implemented.The system has high recognition accuracy and fast recognition speed for COVID-19 positive images,which effectively helps clinicians in auxiliary diagnosis,improves the recognition accuracy of COVID-19 positive patients,and reduces the probability of clinicians’ secondary infection.Experimental results on multiple COVID-19 datasets of Kaggle show that the system in this thesis can accurately and quickly identify patients infected with the disease even if the quality of training images is variable and the number of training images is small.Through theoretical research,algorithm design,experimental analysis and system implementation,the objective of this thesis has been achieved,which provides a reference for further research on COVID-19 recognition and diagnosis.