Research And Implementation Of CT Classification System For COVID-19 Based On Conformer-DRSN

Since 2020,COVID-19 has begun to spread on a large scale around the world,which has brought great challenges to the medical resources of countries around the world.In the treatment process of COVID-19,patients with COVID-19 often need to pay great costs in the detection phase,which also brings great inconvenience to major detection institutions.Based on this situation,people have attempted to combine deep learning methods with medical assisted diagnosis,greatly improving the efficiency of medical diagnosis.Before the prevalence of COVID-19,all kinds of diseases such as cancer and fracture can be classified and identified through the medical auxiliary detection system based on deep learning.However,due to the short outbreak time of COVID-19,the intelligent medical diagnosis system suitable for COVID-19 is relatively scarce compared with the medical diagnosis system for other diseases.According to the above situation,this paper proposes a lung CT image classification algorithm based on deep learning,and designs an intelligent medical image classification system to preliminarily classify and judge whether the lung CT image is COVID-19 CT image,and assist doctors in diagnosis.The work of this article mainly includes the following parts:(1)This experiment uses the dataset published in the virus information database of the National Center for Biological Information and the SARS-CNV-2 dataset as the data sources.Due to the fact that the dataset in this article is collected from two different sources,images that do not meet the classification criteria in both datasets are deleted and preprocessed to maintain consistency in the image data.Fusion of two datasets into a dataset with 6000 CT images.(2)Based on the analysis of the characteristics of COVID-19 CT imaging,a Conforming DRSN algorithm is designed for the classification of different types of lung CT images,which is a specific scene.The backbone network of the algorithm uses the Focus structure and the lightweight Conforming backbone network.The lightweight model structure greatly reduces the number of parameters,and the pixel wise downsampling of the Focus structure can effectively preserve feature information.Introduce the residual shrinkage module into the Transformer to replace the CNN modules in layers 2 to N,in order to solve the noise problem in CT images.The MSRCR algorithm was used to enhance the image dataset and solve the problem of image distortion.Finally,the improved algorithm was compared with other mainstream image classification algorithms on the dataset used in this article.Experimental results have shown that the algorithm proposed in this paper achieves an accuracy of 96.1%and a recall rate of 96.8%,while reducing the model parameter count to 27.31 M.Compared with other algorithms,it has better performance.(3)Build an intelligent CT image detection system for COVID-19.The system has been designed with functions such as image preprocessing module,CT image classification module,and report output module.Image preprocessing module: Adjust the size of the image entering the model,normalize the image,and perform data enhancement.CT image classification module: The core function of this system is to input the CT images uploaded by users to the improved algorithm model for classification.Report output module: Visualize the obtained classification results and test records,and print them out,making it easier for doctors to carry out diagnosis work faster and greatly saving costs for hospitals and patients.