Research On COVID-19 Segmentation Algorithm Based On Dual-stream And Context

Novel coronavirus pneumonia,abbreviated as COVID-19,is an acute respiratory infectious disease.In clinical diagnosis,CT images can accurately display lesion information,and the automatic segmentation of CT images can be realized through computer-aided diagnosis and treatment technology,which can provide doctors with efficient and accurate lesion analysis methods and significantly improve the clinical diagnosis efficacy rate.In this thesis,the infected region of the COVID-19 CT image is segmented by constructing the ContextSE-Unet algorithm model and the dual-stream ContextSE-Unet algorithm model,and the main works of the whole thesis are as follows:(1)In order to solve the problem of boundary blurring and multi-scale lesions in the COVID-19 segmentation task,a ContextSE-Unet segmentation network model is designed and constructed.Firstly,the local context module is used to introduce multi-scale contextual information in the same layer,and the underlying dependencies between the semantic information of the same layer are represented and integrated through the two-way gated context aggregation method to obtain richer semantic information and capture remote context dependencies.Then,the global context module is used to selectively combine different levels of attention features to achieve cross-layer information complementarity,so as to provide richer detailed features for the later segmentation of infection regions.In addition,in order to improve the segmentation accuracy of the network,the SCR module is used to constrain the region-level feature structure relationship,and the EA module is used to optimize the boundary loss and improve the expression of the boundary features of the target region.Experimental results show that the ContextSE-Unet network model has a significant improvement in the segmentation effect of the COVID-19 region compared with other segmentation models.(2)In order to solve the problem of small size of lesions in the COVID-19 segmentation task,a dual-stream context ContextSE-Unet segmentation model is proposed.Firstly,the dual-stream mode is introduced,which takes the shape information as an independent processing flow,and uses a new gating mechanism to connect the information interaction between the two streams,so that the shape flow can focus on processing the shape information and eliminate the interference of other information.Then,a new fusion method is used to fuse the prediction information of the conventional flow with the shape information of the shape flow in the fusion module.In addition,in order to further improve the accuracy of boundary prediction,a new loss function is used to make the final prediction result plot closer to the real label.The experimental results show that the network model further improves the segmentation accuracy of the COVID-19 infected region,and the segmentation effect on small-sized lesions is more excellent.