Study Of Aided Diagnosis Method For COVID-19 Based On Noisy CT Images

Computed tomography(CT)is one of the most effective diagnostic aids for diseases in today’s clinical practice.However,the high radiation hazard during CT examinations has been a concern,and the blind use of low radiation doses can cause a dramatic increase in the noise of the reconstructed CT images,which greatly affects the diagnostic results.Compared with manual diagnosis,assisted diagnosis using deep learning techniques can more easily focus on the lesion characteristics of noisy CT images.Therefore,this paper proposes a deep learning method for noisy CT images,so that it can be applied to assist in the diagnosis of low-dose CT(LDCT)images,which can assist medical personnel in the diagnosis of LDCT images while minimizing the radiation dose to patients during CT examinations and protecting their health.The main work of this paper is as follows.(1)To address the problem that real LDCT image datasets are difficult to obtain,we constructed a noise-containing LDCT dataset of neocoronary pneumonia(COVID-19-LDCT)independently.In this process,the conventional dose of neocoronary pneumonia CT(COVID-19-CT)dataset was first collected,and this dataset was preprocessed as well as the class imbalance problem in the dataset was solved using the Smote algorithm.Subsequently,the noise components in LDCT images were analyzed and the class was modeled autonomously according to its principles and mathematical relationships.Finally,the COVID-19-LDCT dataset was constructed by adding the above-modeled noise to the COVID-19-CT dataset.(2)In response to the problem that there is a large amount of noise in LDCT images,which is not conducive to accurate diagnosis by neural networks,a module NS(Noise Suppression)that can be used to suppress noise is proposed and embedded in the Residual Block Unit(RBU)of the Res Net model.The Res Net improved in this way has two forms,Res Net-CSNS with the same threshold between channels(ChannelSame,CS)and Res Net-CDNS with different thresholds between channels(ChannelDifferent,CD).The NS module consists of a soft thresholding function and an attention mechanism for the SE channels.The soft threshold function assists the model in filtering the noisy features in the CT images for the purpose of noise suppression;the SE attention mechanism is used to assist in setting the threshold value for each channel of the feature adaptively,solving the threshold setting problem in the soft threshold function.Through experiments on the COVID-19-LDCT dataset,the results show that the module can effectively improve the classification performance of the model for images under noisy CT.(3)Based on the above study,a noise suppression module ENS(Efficient Nosie Suppression)with better performance and lighter weight is proposed.To be able to further improve the diagnostic performance and efficiency of the model on LDCT images,we have further improved the above NS module.Firstly,the Leaky-soft thresholding function is designed to address the problem that the soft thresholding function tends to remove some useful features that are not attended to in time during the early training process by adding an appropriate gradient to the part of its gradient that is 0.The new soft thresholding function enables the model to have noise suppression function and at the same time enhances the protection ability of features to prevent the loss of useful information,so as to improve the classification performance of the model under noisy CT images,and the effectiveness of this improved method is demonstrated through experiments.Secondly,to address the problem of high number of parameters caused by the SE attention mechanism,the classification efficiency of the model is improved by introducing a lighter ECA attention mechanism instead of SE,which greatly reduces the introduction of parameters in the model.(4)In order to avoid the problem of singularity in the applicable scenarios of the noise suppression methods proposed in this study,we applied ENS to Res Net and Mobile Net V2 respectively,and proposed Res Net-ENS and Mobile Net V2-ENS.the former is more accurate,but the number of parameters is relatively large,and it is suitable for realizing the noise suppression in devices such as computers with certain computing power.The former is more accurate,but has a relatively large number of parameters,and is suitable for fast diagnosis in mobile or embedded devices.