| In the world,lung cancer has become the most lethal cancer in malignant tumors.As an early manifestation of lung cancer,the earlier the presence of pulmonary nodules is found,it is of great significance to ensure the health and even prolong the life of patients.Compared with natural scene images,pulmonary nodules in medical images have the characteristics of small targets,diverse types,different sizes,and easily confused with the surrounding non nodule lung tissue,which brings great difficulties to the accurate segmentation of pulmonary nodules.In order to solve the problem of poor performance of the existing pulmonary nodules segmentation algorithm,this paper studies the pulmonary nodules image segmentation algorithm based on U-Net network.Feature extraction of input image is a crucial step in segmentation,which is directly related to the quality of segmentation results.Due to the complex imaging characteristics of pulmonary nodule images,it is very difficult to extract the features of pulmonary nodules using neural network.Firstly,this paper proposes a segmentation method of pulmonary nodule images based on dense connected 2D U-Net network.Based on the original 2D U-Net network with encoder and decoder structure,the dense connection structure is introduced,which increases the connection between all levels in the network and enhances the propagation and reuse of features.At the same time,considering the problem of network training speed,the batch normalization layer is added to the network to speed up the network training speed and improve the stability and generalization ability of the network.Experiments show that the proposed 2D U-Net network segmentation based on dense connection has good performance.On the composite data set based on LUNA 16,Dice reached 90.1%.Considering that 2D U-Net only processes a single CT slice and ignores the spatial correlation information in serial CT slices.And in order to further reduce the detection probability of false positive nodules,a segmentation method of pulmonary nodule image based on 3D U-Net network with dense connections is proposed.The dense connection idea is applied to 3D U-Net network,the original structure of encoder and decoder is retained,and the dense connection encoder is designed to extract the input image features.Considering the complexity of the network,a simplified decoder structure is designed and a hybrid loss function is proposed for the training of the network.In addition,3D U-Net network based on dense connection is applied to extract lung parenchyma,which accelerates the speed of traditional image processing method to extract lung parenchyma.The experimental results show that the proposed 3D U-Net network segmentation based on dense connection is excellent,and Dice reaches 92.5% on the composite data set based on LUNA 16.It also performs well in the segmentation of fuzzy boundary contour,adhesion type and small nodules. |
