Research Of Lung Segmentation In CT Image Based On Deep Learning

Statistically,lung cancer continues to be the most vicious of all cancers for several years.With development of related theory and simulation method,medical imaging has become a very important auxiliary method to help identify various cancers.Segmentation of organ provide valuable information for the analysis of disease and is important for planning of treatment strategies,monitoring of disease progression.The widely used medical image segmentation methods include traditional segmentation methods and deep learning methods.Deep learning theory has been extended to many fields in recent years,including computer vision and speech recognition.It extracts features from massive data hierarchically and can achieve high accuracy.Therefore,the thesis studies the lung CT image segmentation model based on deep learning theory.Based on the classical U-Net model and Dilated Convolutions(DC),the thesis presents DC-U-Net model that handles the tasks of lung CT image segmentation.It adds a 1 × 1convolution layer before the model output layer to increase the nonlinearity of the model.In the Keras framework,the DC-U-Net model is implemented using Python language.The tasks include data preprocessing,model building and network configuration.Finally,there are numerical results and conclusions.The research content of the thesis mainly focuses on the following two aspects:First,dilated convolution adds a gap in the convolution,which can expand the receptive field and obtain the more characteristic information while the parameter quantity is constant.In the thesis,dilated convolution is introduced into the U-Net model,in which the dilation rate is designed according to the HDC principle;on the other hand,1×1 convolution is used before the output layer of the model.It can increase the nonlinear expression of the model.Second,in the engineering realization stage,the original image is preprocessed,mainly including denoising processing and data enhancement.This is followed by the model building and network configuration process.In the network configuration process,the truncated normal distribution is used to initialize the weight and the dynamic learning rate is used to updated parameters.Through the numerical experiments on the lungs data in Kaggle,the lung CT segmentation map is given and compared with the other models.The numerical results show that the model has a good effect on lung segmentation and has certain application value.The structure of this paper is as follows: Chapter 1 briefly introduces the background and research status of medical image segmentation;Chapter 2 introduces the traditional imagesegmentation algorithm and segmentation algorithm based on deep learning;Chapter 3 gives the DC-U-Net model;Chapter 4 gives the engineering implementation process based on the model to deal with the image segmentation problem;the preliminary numerical experiments were carried out on the lungs data,the numerical results were listed and the conclusions were drawn.Last part is the summary.