A Method Of COVID-19 Lung CT Image Infection Region Segmentation Based On ADID-UNET Model

New Coronavirus disease（COVID-19）broke out in 2019 and has become one of global health crisises because of its highly infectious nature.Currently,it is generally believed that Reverse Transcription Polymerase Chain Reaction（RT-PCR）is used to detect and screen the cases of COVID-19.However,due to the shortage of equipment and the strict requirements of the detection environment,the rapid and accurate screening of suspected cases is limited.Moreover,RT-PCR detection sensitivity is not high enough,resulting in the detection of a large number of false-negative cases,which seriously affects the early detection and treatment of suspected COVID-19 patients.As another method of detecting COVID-19,CT scanning imaging technology can clearly describe the lung features associated with early Ground Glass Opacity（GGO）and late pulmonary consolidation of COVID-19.However,CT scan images also show similar imaging features to other types of pneumonia or normal tissues,so it is difficult to distinguish them.In addition,it is time-consuming and time-consuming to depict the lung area,which is easily affected by personal experience and bias.In recent years,due to the intelligent and efficient feature learning and obtaining ability of deep learning algorithms,it provides a new direction to solve the above problems,and gradually develops many advanced networks,such as Cov-Net,Covid-Net,UNET,residual UNET,Inf-Net,Attention UNET,and RAD-UNET,but rarely involves in the segmentation of COVID-19 lung infection area.This is mainly due to the following challenges:（1）the texture,size,and location of infected regions in CT scan images are different.（2）It is difficult to distinguish GGO from healthy tissue regions in the process of segmentation because of its low boundary contrast and fuzzy appearance.（3）The noise around the infected area is large,which greatly affects the segmentation accuracy.（4）the cost and time to obtain the high-quality real value of pulmonary infection are very high in CT scan images.Therefore,most of the COVID-19 lung infection data without ground truth values are mainly used for diagnosis,and only a few data provide original data and segmented ground truth values.To solve the above problems,the principles and advantages of dense network,improved diffusion convolution module,and attention gate system are used to be for reference and proposes an ADID-UNET network for the segmentation of COVID-19 lung infection region based on the characteristics of COVID-19 CT scan images.There are four innovations of the proposed network are summarized as below:（1）To solve the problem of gradient disappearance in the training process,this paper uses the dense network to replace the traditional convolution and maximum pooling functions.A dense network enhances the feature propagation ability of the network by repeatedly learning and extracting advanced target features.Moreover,the training parameters of the dense network are less,which reduces the complexity and computational overhead of the network.（2）To increase the receiving domain of the output result of the encoder structure and compensate for the problems caused by edge blur,an improved dilation convolution module is introduced between the connecting coding structure and the decoder structure,which increases the receptive fields in different ranges of the prediction results,provides more edge information,and enhances the edge feature learning and extraction ability of the network.（3）Because the edge contrast of GGO is very low and the gray level is uneven,this paper uses an attention gate system to replace the simple clipping and copying operation in the UNET network.By suppressing the background information,the accuracy of network segmentation of infected areas is further improved.（4）Because the number of COVID-19 datasets with ground truth values is limited,which is far from the minimum number of samples required for complex model training,this paper uses data enhancement technology to expand the data based on the collected public datasets.To estimate the segmentation effection of ADID-UNET network,accuracy,Dice coefficient,AUC,precision,sensitivity,specificity,and Structure Measure(₈))are introduced,Enhanced alignment measure(),and F1 score and Mean Absolute Error（MAE）were used as evaluation indexes.Through the analysis of the final qualitative and quantitative results,it can be found that compared with other advanced segmentation models,the ADID-UNET model proposed in this paper can achieve accurate segmentation of the infected areas of COVID-19 CT scan images,and most of the segmentation performance indicators are more than 80%,which has a good clinical diagnosis prospect.