Research On COVID-19 Medical Image Recognition Method For Lightweight Network Based On Attention Interaction Mechanism

With the development of science and the progress of medical equipment,medical imaging technology has been used as a common means in the diagnosis of living tissues of patients.In recent years,deep learning,as the most representative technology in the field of machine learning,has achieved breakthrough achievements in many fields.The combination of neural network technology and medical image technology,using neural network to extract the characteristics of the task target,and correctly process medical image,has also attracted extensive attention in the field of medical image analysis.All of these can not only assist imaging doctors in rapid diagnosis of patients,but also greatly save time and labor costs,and help multiple patients to diagnose and effectively reduce the burden of doctors.Multi-level neural networks supported by big data constitute the essence of deep learning,and a deep learning model usually contains a large number of network layers and parameters.However,while deep convolutional neural network has high accuracy,it still takes too long to train and test due to the huge amount of data,which extends the diagnosis time and cannot guarantee rapid diagnosis in the early stage of the epidemic.Lightweight neural network generally refers to the neural network model that requires fewer parameters or less cost in the calculation process.Based on these characteristics,lightweight neural network can be applied to some devices with limited computing resources or insufficient hardware conditions,such as computers or smart phones with poor performance,other embedded devices,or some underdeveloped regions with insufficient hardware conditions,to achieve the realization of computing process.Meanwhile,the COVID-19 outbreak caused by Novel Coronavirus SARS-COV-2 has swept the world,impacting medical security and the economy in many countries and regions.The combination of convolutional neural networks and medical imaging applications provides a new feasible direction for COVID-19 diagnosis,and has achieved significant results.How to study and build targeted lightweight neural networks,so as to deal with the possible emergent situation more effectively,is a problem worthy of researchers’ consideration.In view of the characteristics of lightweight neural network,attention interaction mechanism is introduced in this paper,and two optimization models are proposed based on lung medical image as the task target:1.Proposed a medical image classification method for COVID-19 based on MAI-Net model based on multi-channel attention-interaction mechanism.As samples of the same category of X-ray images have the characteristics of high similarity,and the feature variability of multiple different samples is low,In the paper,a multi-channel Attention Interaction Enhancement module is proposed to extract shallow and deep features respectively.Based on this module,MAI-Net,a lightweight neural network based on attention-interaction mechanism,is proposed to classify COVID-19 medical images.The method was validated on a public data set,and experimental results showed that its overall accuracy and COVID-19 category accuracy were 96.42%and 100%,respectively,with a COVID-19 sensitivity of 99.02%.Considering the factors such as accuracy rate,parameter number of network model and calculation amount,MAI-Net has better practicability.Compared to existing work,MAI-Net has a simpler network structure and lower hardware requirements for devices,which can be better used for common devices.2.A medical image classification method for COVID-19 based on LPA-Net model is proposed based on the mechanism of positional pixel attention enhancement.Different from MAI-Net’s multi-channel attention interaction mechanism,LP A-Net uses the GC module to improve the lightweight neural network EfficientNet by combining the location pixel attention with the channel attention module.The method was validated on a public data set,and experimental results showed that its overall accuracy and COVID-19 category accuracy were 93.37%and 98.51%,respectively.Experimental results show that LPA-Net achieves the expected effect,and compared with many classical network models,this method has higher classification accuracy,which proves the effectiveness of the proposed model.