Fine Segmentation Of Lung Nodules By Combining Image Processing And Deep Learning

Lung cancer has been the leading cause of death in all different cancers,ranking first in cancer mortality in China for a long time.The use of computed tomography(CT)images for detection is an important strategy for early diagnosis of lung cancer and improvement of survival time.Segmentation of lung nodules is extremely important in the diagnosis and treatment of lung cancer.However,the recognition and diagnosis of CT images is time-consuming and laborious,which requires a lot of human-computer interaction.Therefore,the development of an accurate and consistent automatic segmentation model has great significance to saving the lives of patients.Lung nodules are varied,with large differences in the shape and texture.The ambiguity of nodule morphology and the influence of background noise such as lung wall and blood vessels make the accurate segmentation of nodules remain challenging for both humans and computers.In this paper,we investigate the method of lung nodule segmentation by combining image processing and deep learning.Firstly,we use image processing-based methods to preprocess CT images to roughly localize nodular lesion regions,followed by using neural networks to accurately segment lung nodules.Secondly,we propose a U-Net-based dual branch network(DB U-Net)combined with graph neural network,which can effectively explore the feature information in slices and the relationships between adjacent slices.Further,we use feature fusion and fully connected conditional random field(CRF)to optimize the segmentation algorithm.The main work and innovations of this paper are summarized as follows.1)Lung nodule segmentation based on the combination of image processing and neural networks.In this paper,using the feature information of different types of lung nodules,in view of the blurred morphology of ground-glass nodules and the problems of juxta-pleural nodules and juxta-vascular nodules which are greatly affected by background noise such as lung walls and blood vessels,we use the image processing-based segmentation method to coarse divide the lung nodule.The method effectively removes the interference of nodule morphology and background noise on the nodule segmentation,allowing the neural network to focus on learning the features around the boundaries.After that,the result of the coarse divide of the nodule is used as the input of the neural network.We use multiple different segmentation models to achieve the segmentation of the nodule.2)Lung nodule segmentation based on the combination of dual branch neural network and graph neural network.Since both 2D and 3D network segmentation models have their limitations,the 2D network model is difficult to learn the spatial information of CT images,and the 3D network model can effectively learn the spatial information of CT images,but the complexity is higher.Therefore,we propose a two-branch neural network by combining the advantages of 2D and 3D networks.In the 3D branch of the model,multilayer slices adjacent to the CT image are used to help the network learn the spatial information of the CT image.In addition,we incorporate graph convolution into the neural network,which can be used to learn the similarity in the spatial structure of CT images and enhance the sensitivity of the model to feature channels.3)Optimization of lung nodule segmentation algorithm based on feature fusion and fully connected conditional random field.In the process of training lung nodules using neural networks,the strategies of feature fusion and CRF are used respectively.Feature fusion can fuse the information of deep and shallow features of the data,which can provide the model with high-resolution information of CT images,while CRF is used to capture the detailed information of nodule boundaries,which can effectively improve the segmentation of lung nodules.This paper evaluates the performance of the proposed algorithm through the public dataset of the Lung Image Database Association(LIDC)and the dataset of the cooperative hospital.The accuracy of the two datasets is 85.06%and 82.64%,which is better than the mainstream segmentation algorithm.