Deep Learning-based Segmentation Of COVID-19 Lung CT Images

At the beginning of the 21 st century,medical imaging technology has gradually become an important way to assist diagnosis in clinical medicine.In recent years,with the development of artificial intelligence,it is a very promising direction to use artificial intelligence method to help doctors make clinical diagnosis.For example,at the early stage of the novel coronavirus pneumonia,a large number of patients including suspected cases,confirmed cases and follow-up cases need to do chest CT examination.The chest CT will set a small range of Ground-Glass Opacity(GGO)to observe the changes and severity of the pneumonia.At that time,the number of patients surged in a short period of time and medical personnel were insufficient,so the disease put a huge burden on medical professionals.In the early stages of COVID-19 pneumonia,lung CT may show small,subpleural and circumferential GGO,which will take more time to diagnose than diffuse GGO or consolidation.Therefore,it presents a huge challenge for the radiologist,requiring a great deal of work and high diagnostic accuracy.Moreover,the radiologist’s eye fatigue increases the potential risk of missing some small lesions.So,developing an AI approach for computer-aided diagnosis of COVID-19 could be very helpful to radiologists.This paper proposes a medical image segmentation network model based on boundary information to segment COVID-19 lesion areas in lung CT.The main work and innovation points of this paper are summarized as follows:1.Lung CT image segmentation method based on deep learning: a new deep learning network framework is proposed for the segmentation of infected areas in medical images.By using the aggregation module to aggregate high-level features,the global location map is obtained,which serves as the initial boot area for subsequent steps.We designed a boundary guidance module integrating features of each layer for feature extraction of each layer to further extract boundary information.While using reverse attention module step by step from high to low level,we combined the boundary guidance module with it to further excavate hidden details of each layer and obtain lateral output of each layer.Finally,we fused features of each layer.So that the network can fully extract the details of the past model is difficult to notice.We carried out experiments on different data sets and proved that this design could enable the model to further extract the hidden lesion region boundary information,thus greatly improving the segmentation effect.2.COVID-19 image segmentation method based on improved U-Net network:On the basis of the above,we use the framework containing boundary guidance module to improve the classical medical segmentation network U-Net.We modified the concat and skipconnection operations in U-Net.Starting from the deepest upsampling,we gradually used the reverse attention module to extract features from high to low.Combined with the boundary guidance module,the improved network was more sensitive to the boundary information of the target segmentation region.In addition,we designed the model to be convenient for pruning.The advantage of this is that all sub-networks can learn the required knowledge only by training the main network,and the network size can be trimmed by pruning,so as to achieve a balance between the size and accuracy of the model.Experiments show that the segmentation performance of this network is better than that of U-Net,and we can greatly reduce the number of model parameters by proper pruning while keeping the accuracy of prediction results with little change.