Research On Lung CT Image Segmentation Algorithm Based On Deep Learning

The incidence rate and mortality rate of cancer are extremely high.Each year,more than 2.5 million people die from lung cancer,accounting for 22%of the total number of cancer deaths.Therefore,lung cancer is also known as the number one cancer.A benign or a malignant lung cancer can be judged by the types of lung nodules.Therefore,to clarify the types of lung nodules through segmentation has an important significance on research.Traditional methods in manually segmenting lung nodules require professional knowledge of doctors.Insufficient experience often leads to missing segmentation.In early lung cancer images,lung nodules are generally too small and too shallow for observation,which is very easy to cause misdiagnosis.Providing doctors with segmented lung images through image segmentation can greatly improve the efficiency of physicians and reduce the probability of misdiagnosis.In this thesis,lung CT images from the LIDC-IDRI database is taken as the research object,the main research contents are as follows:First,in view of the poor segmentation effect of the vascular adhesion part in segmenting lung nodules,this thesis designs a network model RAU-Net based on the U-Net model.To begin with,the ordinary convolution structure of the up and down sampling layers in the U-Net model is changed to a residual structure with the help of the residual network.This can effectively avoid the problem of gradient disappearance while obtaining more image features,and also can accelerate the convergence speed of the network model.Subsequently,a dual-channel attention module is added to the model,and the low-resolution spatial attention channel is used to increase the model's interest in the lung parenchymal region,and to strengthen the network's ability of feature extraction from region.Finally,the added normalization layer effectively reduces the training time of the model and improves the segmentation accuracy of the model.The experiment shows that the RAU-Net model has an improvement in the accuracy of lung nodule segmentation.Second,this thesis designs a DSU-Net network model based on the U-Net model so as to improve the left and right lung adhesions that are easily generated from segmenting the lung parenchymal region,and also to optimize the segmentation accuracy of the lung parenchymal boundary.It combines the U-Net model with the atrous spacial pyramid pooling layer and the densely connected network structure.After the last downsampling,the output result is input to the pyramid pooling layer.With different pixel values extractd,the model performs better in grasping details,and the calculation of useless features is also suppressed during the segmentation process.To enable the model with more feature information,this thesis replaces the U-Net convolutional layer with a densely connected structure.By conducting so,the pixel information can be better remained;meanwhile,it can enhance the robustness of the model and reduce the phenomenon of over segmentation.The binarization preprocessing of the lung image can help to achieve a better segmentation effect by distinguishing between the foreground and the background.The improved loss function increases penalty weights for over segmentation and under segmentation,thus effectively reducing the over-segmentation rate.The experiment shows that the DSU-Net model has an improvement in the accuracy of lung parenchyma segmentation.Third,this thesis designs a network model RACGAN based on the GAN model so as to improve the segmentation accuracy of small nodules,use RAU-Net as the splitter,the original lung image is preprocessed through simple linear iterative clustering enhancing regional features of the shallower and smaller lung nodes.This reduce missing segmentation of small nodules and the interference of useless pixels such as blood vessels in segmentation.Then the preprocessed image is input to the conditional segmented adversarial network with the RAU-Net as the segmenter.It uses discriminative loss to modify the network parameters and improve the segmentation performance of the model.The experiment shows that the conditional segmented adversarial network model has effectively improved the segmentation accuracy of lung nodules compared with the RAU-Net.Through the above improvements,the network model DSU-Net used in this thesis has its segmentation results and the dice overlap rate improved,and can distinguish the left and right lung parenchyma well,when segmenting the lung nodule,the network model RACGAN can identify small nodules well,while the false positive rate has been significantly reduced.Thus,the experimental methods used in this thesis are proven to be able to help doctors with their diagnostic accuracy.