Research On COVID-19 Lung Infection Region Segmentation Algorithm Based On Context Modeling

The sudden outbreak of Corona Virus Disease 2019(COVID-19)brought great disaster to the whole world.Up to now,450 million people have been infected and 6million people have died.In the clinical diagnosis of COVID-19,the imageology judgment based on computed tomography(CT)of the lung region is one of the necessary means of diagnosis.Therefore,COVID-19 automatic segmentation technology for lung CT images can help doctors realize rapid diagnosis.However,the scattered distribution of infection areas and complex background interference make it very challenging to accurately and completely segment COVID-19 lung infection areas.Thus,this paper focuses on COVID-19 infection region segmentation technology for lung CT images,and designs two models based on the convolutional neural network(CNN)and Transformer.Firstly,with CNN as the main architecture,considering the boundary constraints,context relationship,and semantic guidance,a COVID-19 lung CT infection region segmentation network is proposed,named boundary-context-semantic network(BCSNet).The BCS-Net includes three progressive Boundary-Context-Semantic Reconstruction(BCSR)blocks in the decoder stage.In each BCSR block,the attentionguided global context module aims at highlighting important spatial and boundary positions,modeling the global context dependence,thereby learning the most discriminative features.The semantic guidance unit is designed to refine the features of the decoder by aggregating multi-scale features at the intermediate resolution and generate the semantic guidance map.A large number of experiments show that the proposed network is superior to the existing comparison algorithms both qualitative and quantitative experiments.Secondly,with Transformer as the main framework,a Cross-Scale Transformer Network(CST-Net)is proposed to locate the COVID-19 lung CT infection regions,including a multi-level parallel gated axial-attention module and a cross-scale Transformer module.Specifically,the multi-level parallel gated axial-attention module is used to extract multi-level self-attention in parallel,thereby obtaining efficient feature representations and semantic relations containing global contextual information.The cross-scale Transformer module makes full use of the features of low and high dimensions through the effective combination of high-level semantic information and low-level features to make up for the defects of simple semantics and unstable structure of medical image.The comprehensive experiments and comparisons show that the proposed Transformer-based CST-Net can achieve better performance in the COVID-19 lung infection segmentation.