Research On COVID-19 Detection Algorithm Based On Convolutional Neural Networ

The China Center for Disease Control and Prevention（CCDC）recently published a statistical report on the prevalence of positive COVID-19 cases across various provinces and cities.According to the report,as of March 2,2023,there were an estimated 11,000 new positive cases reported daily nationwide.COVID-19 testing plays a crucial role in interrupting the transmission chain of the virus and safeguarding public health and well-being.Consequently,the identification of viral infections through CT image recognition among the targeted population holds immense significance.Moreover,such detection methods facilitate the formulation of subsequent diagnostic and therapeutic strategies for patients.The deep learning model Mobile Net V3 exhibits notable attributes such as accuracy,efficiency,and a reduced parameter count,rendering it well-suited for addressing the intricate challenges associated with COVID-19 detection.Consequently,this study endeavors to investigate the application of the Mobile Net V3 model in the context of COVID-19 detection.The primary research objectives are outlined as follows:（1）To address the challenges of high computational complexity and low detection accuracy in deep learning models for CT image detection,a hybrid lightweight detection model called MCDANet（Mixed Detection Lightweight Model based on Deep Learning and Cluster Algorithm）is proposed.MCDANet employs an improved Mobile Net V3 model as the feature extraction scheme and an arithmetic optimization algorithm as the feature selection scheme.In the feature extraction stage,firstly,a cascaded branch residual network is introduced to replace the residual network in the traditional Mobile Net V3 model,reducing the repetition of feature mappings in the intermediate layers of the network.Secondly,a dual-channel feature extraction path is utilized to extract features from bottleneck blocks,capturing more spatial information and enhancing the model’s capability for anomaly detection.Finally,the DW（Depth-wise）convolution kernels in the dual-channel feature extraction path are replaced with3×3 sizes,resulting in a 47.1%reduction in complexity for each bottleneck block.In the feature selection stage,firstly,the search range is determined based on the relationship between the random value’’and the mathematical optimization to accelerate the MOA function.Secondly,in the exploration stage,the solution space is explored globally using multiplication and division operators.Finally,in the development stage,depth searching of the solutions is conducted using addition and subtraction operators to find the optimal solution within the local scope of the obtained solution.（2）In response to the challenge of ineffective detection of early-stage COVID-19 patients due to the limited visibility of lung features in CT images,a fusion algorithm for patient CT image features（NPL）is devised,in conjunction with the MCDANet model,to enhance COVID-19 detection（IFIMNet）.The NPL algorithm employs the Non-subsampled Shearlet Transform（NSST）to decompose the source image into high-frequency and low-frequency subbands.Additionally,an optimization method is employed to determine the time decay factor₀0))of the PAPCNN algorithm,leveraging the image’s highest grayscale value and grayscale distribution.These optimized parameters are then integrated into the PAPCNN algorithm to achieve high-frequency subband fusion.Simultaneously,a latent low-rank representation algorithm is utilized to fuse the coefficients of the low-frequency subbands,effectively preserving significant information within the image’s low-frequency components.Finally,the fused image obtained through inverse NSST serves as the foundation for constructing a novel dataset,enabling the training of the detection model to successfully accomplish COVID-19detection.Experimental results have substantiated the efficacy of MCDANet in achieving remarkable detection accuracies of 97.85%and 98.35%on the COVIDx CXR-3 Dataset and the Chest X-Ray Images dataset,respectively.Furthermore,IFIMNet demonstrates a commendable detection accuracy of 98.73%on the CXR COVID-19 dataset,showcasing a notable enhancement of 1.58%over the MCDANet model.The proposed methodologies in this study effectively mitigate the computational complexity of the network while concurrently attaining superior levels of detection accuracy in the realm of COVID-19 detection.