| CT image,as a medium to show the morphology of human internal organs,is an important research object in medical diagnosis and treatment.At the beginning of 2020,Corona Virus Disease 2019 swept through China and spread to many countries around the world.The health of the global population is never before challenged.With the infringement of industrial waste gas emissions and smoking on human lungs,more and more people suffer from lung diseases,especially lung cancer.Usually,doctors can examine and diagnose pulmonary nodules by examining thin slice CT scans of lungs through their eyes and judging them based on their own experience.Therefore,there are too many human intervention factors.In addition,for a set of tens or even hundreds of CT images of the lungs,because the patients do not have professional knowledge,they cannot have an intuitive understanding of them,which is not conducive to the communication between doctors and patients.In view of the above pain points,many scholars have conducted in-depth research on the automatic diagnosis of pulmonary and threedimensional reconstruction of the lungs,but there is still a long way to go before it becomes more practical.The following three parts are mainly explored in this article:segmentation of pulmonary parenchyma,segmentation and detection of pulmonary nodules,and threedimensional reconstruction of pulmonary parenchyma area.Firstly,a pulmonary parenchyma segmentation method based on fully convolutional networks is introduced.A network architecture based on fully convolutional networks for pulmonary parenchyma segmentation is proposed and trained.The method does not need to manually select seed points or select thresholds,just input CT images of the lungs,then the network can output the segmentation results,which improves the degree of automation of pulmonary parenchyma segmentation.In order to solve the problem of pulmonary parenchymal cavities that may exist in the predicted results of the network,the lung is repaired with the morphological method,which improved the accuracy of the whole segmentation and achieved a better segmentation of the pulmonary parenchymal region.Secondly,the input of sample form of U-Net network and the number of network layers are analyzed and improved,and a multi-level aggregated U-Net network are proposed.The network can synthesize the effects of the deep and shallow U-Net network to segment lung nodules better and more accurately.It also has the characteristics of pruning,which can be further satisfied when there is a need to reduce the amount of parameters.For possible false positive problems,a convolutional neural network is used to remove false positives to a greater extent,which can reduce the false positive rate and improve the practicality of the entire system.Finally,in order to solve the problems of the marching cubes algorithm which is commonly used in 3D reconstruction of medical images,such as redundant traversal and poor model surface effects,the algorithm are improved.Firstly,the sequential traversal of the original algorithm is improved.It is proposed to judge the intersection of neighboring voxels and isosurfaces based on the intersection of isosurfaces and cube voxels.It only traversed the cubes with intersect conditions to improve the traversal efficiency.At the same time,for the problem that the noise may affect the reconstruction effect during the rendering of 3D reconstruction surface,a method which can smooth the normal vector of the triangles to improve the visual perception of the reconstruction model is proposed. |
