Research On COVID-19 CT Image Segmentation Algorithm Based On Boundary Perception And Attention Mechanism

Since the end of 2019,the COVID-19 virus was first discovered in China and has since spread worldwide.Due to its high infectivity,it has seriously affected human social activities.Detection,diagnosis,and treatment of the disease have become important means of prevention and control.Due to the significant changes in lesion areas during the early,middle and late stages of COVID-19,even professional physicians can easily lead to misdiagnosis and missed diagnosis in the diagnosis of COVID-19.Due to the development of deep learning,computeraided analysis of COVID-19 CT images can improve the diagnostic accuracy of lesions and help patients receive prompt treatment.Compared with traditional medical image segmentation methods,deep learning has advantages such as higher accuracy and robustness of models.For the segmentation of COVID-19 lesions,this paper proposed and designed two neural network models: a boundary-perception convolutional neural network model and a cross-reverse attention-based convolutional neural network model.The research content of this paper is as follows:1.This paper proposed and designed a boundary perception axial attention network(BP-Net)for segmentation of COVID-19 lesions in lung CT images,which are characterized by fuzzy lesion boundaries,irregular lesion areas,and strong noise in lung regions.In the encoding stage,the model used Resnest101 as the encoder to extract features and provided improved feature representation through multi-path representation and attention mechanisms.The fusion feature map of high-level features enhanced by receptive fields is used as input to multiple decoding layers.In the decoding stage,parallel axial attention mechanisms are inserted into high-level features to further highlight foreground targets and improve the segmentation accuracy of the network model.The outputs of the decoding stage and the boundary distribution map are supervised separately,and decoders with different morphologies of high and low-level features are combined to enable the network to learn rich semantic information.The accuracy of the network model is validated on the COVID-Semi Seg dataset in this study.2.The lesion area of COVID-19 varies greatly in the early,middle,and late stages of the disease,and is subject to significant noise interference.To address these challenges in COVID-19 lesion segmentation,this paper proposed and designed a convolutional neural network model based on cross-reverse attention(DCRA-Net),which achieved accurate segmentation of COVID-19 lesions.The network aggregated multi-scale contextual information through enhanced receptive fields and multi-scale features to achieve effective segmentation of foreground lesion targets.Using Res2Net101 as the feature extractor for encoding,the network obtained finergrained multi-scale features,enhancing the receptive fields of each layer.The deformable feature pyramid is used to extract rich contextual information from deep feature maps.The cross-reverse attention mechanism is embedded in the model to superimpose the reverse segmentation target on the output image of the parallel decoder,refining the boundary information of the segmentation target.This article verified the model’s high efficiency and accuracy on the COVID-Semi Seg dataset.In summary,to address the challenges in segmenting the COVID-19 lesion regions,this paper proposed and designed two neural network models based on convolutional neural networks to improve the issues of blurred boundaries and varied shapes of lesions.The excellent performance on the COVID-19 lung CT dataset demonstrated that the two network models proposed in this paper have accurate and efficient segmentation performance.