Lung CT Image Segmentation Based On Capsule Network

The mortality of lung cancer has always been the top of all kinds of cancers in China.With the changes in people's living environment,the incidence of lung cancer is gradually increasing.For lung cancer,the most important thing is “early detection,early diagnosis,early treatment”can improve the effectiveness survival rate of patients.CT images are one of the best methods for detecting lung lesions and the early diagnosis of lung cancer is diagnosed by judging the nodules.We apply medical image processing to computer-aided diagnosis systems,which can help doctors identify and analyze lung nodules more quickly.Accurate lung parenchymal segmentation and lung nodule classification algorithms can not only improve the work efficiency of doctors,but also help doctors reduce work pressure and misdiagnosis rates,as well as improve the efficiency of diagnosis.Aiming at the characteristics of strong noise and complex structure of lung CT images,this thesis based on the capsule network,combines traditional image segmentation methods and deep learning,to deeply study the segmentation of lung CT images.The main research results are as follows:(1)In the segmentation algorithm of lung parenchyma,the traditional neural network is easy to lose the spatial information and edge information of the image when performing the maximum pooling operation on the image when segmenting the image.This thesis proposes a capsule network-based The network architecture of lung CT image segmentation.Through the parallel independence between the various capsule layers,the vector neural unit extracts the features of the image,adds a deconvolution layer to achieve the purpose of segmentation,and obtains the segmentation result map,and manual annotation and other compared with segmentation models,the algorithm proposed in this thesis has better segmentation performance.(2)Aiming at problems such as the weak robustness of traditional dynamic routing algorithms and the insignificant clustering effect of hidden categories,a fuzzy clustering algorithm based on positive optimization terms is proposed to optimize the iterative process in dynamic routing.Through the combination of the membership probability distribution problem in fuzzy clustering and the information transmission between the child capsule and the parent capsule,the accuracy and universality of the clustering effect are effectively improved.(3)Aiming at the prediction problem of lung nodule classification,a lung nodule classification algorithm combining dual-path network and capsule network are proposed to predict nodules.Perform feature extraction on two image blocks of different scales through a dual-path network.The Res Net block and the Dense Net block assist each other,and finally perform feature integration on the extracted features.The obtained tensor is used as the input of the capsule layer,and the dynamic routing algorithm is used to feature capsules are divided into two categories.The output vector of the capsule contains the image information of the nodule,and the output modulus length indicates the probability of benign and malignant nodules.Through comparison experiments with manual annotation and other network models,good and stable results are obtained.