Research On Auxiliary Diagnosis Method Of COVID-19 Based On Convolutional Neural Networ

A very contagious infectious illness,the novel coronavirus.Medical imaging studies of the lungs are often used to quickly diagnose new coronary pneumonia.As a result,one of the common diagnostic techniques for new coronary pneumonia is to take a quick X-ray.The COVID-19 auxiliary diagnostic model may significantly lighten the load on medical professionals and lessen the demand on their resources.This thesis improves the accuracy of the new coronary pneumonia from three different perspectives: network optimization,data set generation,and loss optimization.It does this by focusing on the current network conditions of the X-ray auxiliary diagnostic model for COVID-19 and the corresponding data set conditions.The following particular research is done on the X-ray aided diagnostic model’s accuracy:(1)This research builds an FS-TRes Net network,which employs the TRes Net model with high GPU usage efficiency as the core model,optimizes the structure,and provides a new network architecture to address network optimization and improvement.The TRes Net model’s characteristics may be enhanced by a compound attention mechanism that combines local and non-local attention,increasing the model’s training precision.By comparing the proposed FS-TRes Net model to the existing basic network,this thesis demonstrates its effectiveness.Ablation experiments were conducted to investigate the improvement effects of compound attention and TRes Net model optimization and to validate the efficacy of each component.(2)The X-ray image data set of new coronary pneumonia has an imbalance in the number of categories,therefore a generation diffusion network with a superior generation impact of tiny data sets is employed to generate the data.Additionally,this research employs an updated generative diffusion model to create COVID-19 category data due to the generative diffusion model’s poor pace.The COVID-19 Chest X-ray Database was used to train the dataset creation model in this research,and the FS-TRes Net network that was suggested in this article served as both the training and testing network for the enhanced dataset.According to the test findings,the F1 value of the model under test has risen by roughly 1.2%when utilizing the produced model and an adequate generation ratio compared to the FS-TRes Net network without generation.(3)This study suggests GFLoss,which combines the benefits of multivariate cross-entropy,Focal Loss,and GHM Loss to increase the training accuracy of unbalanced data sets,to address the issue of an imbalanced number of distinct categories.The FS-TRes Net network is used as the foundational network for training,and the incremental new coronary pneumonia X-ray image data set produced by the model from the previous generation is used as the data set.The GFLoss created in this work is used by The Loss.The efficacy of GFLoss has been shown by the achievement of a 1% improvement.In this study,the most accurate X-ray diagnostic model for new coronary pneumonia had an accuracy rate of94.36%.The best model is used for visual research,and the heat map produced by gradient inversion may perceptually identify the most crucial portions of the original picture,enhancing the reliability of the computeraided diagnostic system.A new crown pneumonia classification demonstration system is created for evaluating human-computer interactions based on the best model.The system is capable of picture input,image presentation,and display of the results of the new crown pneumonia classification.