Research On Segmentation Method Of Pulmonary Nodules In CT Images

In today’s air pollution environment(such as smoking,oil fume,air pollution,etc.),lung cancer has become one of the malignant tumors with the highest incidence and the fastest growth rate,and it is one of the important factors that threaten people’s lives and health.Since the main symptoms of lung cancer in the early stage are pulmonary nodules,whether lung nodules can be detected in the early stage of lung cancer has a great influence on the later treatment.With the continuous development of Computed Tomography(CT)technology,it has become a common diagnostic method for lung nodule detection.However,for CT images,lung nodules are difficult to find,are small in size,and have variable appearances.They are very similar to some tissues of human lungs.If traditional medical methods are used to detect them,it will rely too much on the experience and experience of the physician.Subjective judgments are prone to omission and over-segmentation.This article summarizes and studies the existing methods,and proposes two lung nodule segmentation algorithms based on deep learning.The main work of this paper is as follows:First,in view of some of the problems that may occur in the U-Net model in the process of lung nodule detection,such as network degradation,over-segmentation or missing segmentation,the author puts forward three opinions on improving the U-Net network,forming a kind of The new network RSEU-Net.This method combines the advantages of the normalization layer,Residual Network(Residual Network)and SE network.First of all,the residual network can effectively avoid the gradient disappearance problem of the model during the experiment,and it can also converge the network model faster and more efficiently.Secondly,the use of the normalization layer can speed up the model training while improving the generalization performance of the model;finally,because the first two networks only extract features in the shallow layer,their feature extraction capabilities are relatively weak,so this paper adds SE networks to extract lungs.Nodules are deeper features.Experimental results show that compared with U-Net network,this method has effectively improved the segmentation accuracy of lung nodules.Second,in the end-to-end lung nodule detection,although the Faster R-CNN algorithm is recognized by researchers as the most accurate,it is still found to be missed and falsely detected in experiments.Especially when examining some small lung nodules,the extracted feature maps will have information loss after passing through multiple pooling layers of the VGG16 network.In response to this problem,this chapter improves the Faster R-CNN algorithm.First,add deconvolution to the back of the VGG16 network to enhance the resolution of the deep feature map;secondly,change the number and size of RPN anchor points to improve the model’s sensitivity to lung nodule detection;finally,do the non-maximum suppression algorithm Improvements are made to improve the model’s detection effect on multiple lung nodule targets.The experimental results show that this model has improved the segmentation accuracy of lung nodules compared with the original Faster R-CNN model.