Research On Segmentation Technology Of COVID-19 Lung CT Images

Since the outbreak of the novel coronavirus(COVID-19),it has caused huge loss of life and property around the world.Lung segmentation and infection region segmentation for COVID-19 CT have great significance for large-scale screening and severity assessment of the disease.However,lung CT of COVID-19 patients presents a high level of complexity,which poses severe challenges to lung segmentation and pneumonia lesion segmentation techniques.First,it is difficult to realize a robust lung segmentation method based on deep learning because of the serious shortage of samples in the early stage of the disease.Although the traditional lung segmentation algorithm does not rely on a large number of samples,it sitll cannot solve the problem of losing large areas of high-density areas in the lung segmentation results;Second,different from pulmonary lesions such as pulmonary nodules,lung tumors,and interstitial pneumonia lesions,the new coronary pneumonia lesions have various types and huge morphological differences,the existing lung lesion segmentation models have poor performance on the COVID-19 dataset.In view of the above challenges,this paper aim to dive into the methods of segmentation and lesion segmentation in COVID-19 CT images.The main work of this paper is as follows:Firstly,we propose a two-stage COVID-19 lung segmentation method based on rib contour recognition.The rib shape information is introduced into the lung segmentation process through the rib inner edge line recognition technology,which solves the problem that the existing lung segmentation methods cannot handle the problem of missing large areas of high density.In addition,a contour correction algorithm is proposed,which adjusts the rough segmentation contour to the lung boundary with the greatest probability by finding the local optimal balance point of boundary density,gradient and shape.The method can achieve accurate lung segmentation on the COVID-19 dataset without requiring a large number of labeled samples.Finally,the effectiveness and robustness of the method are verified by experiments.Secondly,we designed a segmentation model of pneumonia lesions based on a hierarchical self-attention mechanism.Through the hierarchical self-attention mechanism and multi-scale feature fusion mechanism,the model has both good localization ability and detail capture ability.The proposed optimized loss function solves the problem of unbalanced positive and negative samples in lesion segmentation and makes the model pay more attention to difficult pixels.Experiments show that our model structure and loss function achieve superior pneumonia lesion segmentation performance.