Research On Lung CT Image Segmentation Algorithm Based On Deep Learning

Lung cancer is a high-incidence disease worldwide,and the mortality rate is still high.Early detection and early treatment are important means to improve the cure rate of lung cancer and prolong the life cycle of patients.Pulmonary nodules are the main manifestation of early lung cancer.Therefore,early diagnosis and analysis of lung nodules is the key to improving the survival rate of lung cancer patients.The use of computerized tomography(Computed Tomography,CT)to screen for pulmonary nodules is currently a commonly used diagnostic method.With the increasing number of patients,the lung CT data is also growing exponentially,which undoubtedly brings huge challenges and burdens to the manual screening work of physicians.Therefore,computer aided diagnosis(Computer Aided Diagnosis,CAD)technology is used Pulmonary nodule detection and segmentation is very necessary,which can greatly improve the diagnosis efficiency of physicians and further improve the accuracy of lung cancer diagnosis.Due to the variability of lung nodules in size,shape and similarity with lung blood vessels and other tissues.When using traditional segmentation methods to segment lung nodules,it relies too much on the doctor's prior knowledge and subjective judgment,which leads to the situation of missed segmentation and over-segmentation.The segmentation process using the deep learning algorithm no longer requires artificial selection of features,and can extract more specific and more identifiable information.The use of deep learning algorithms for medical image segmentation has now become an important research direction.U-Net network has been widely used in the field of medical image processing due to its simple structure and strong generalization ability.In this paper,based on the U-Net model,the lung nodule segmentation algorithm for lung CT images is completed.The main research content of this article includes the following three parts:First,in view of the possible network degradation and poor segmentation effect of the U-Net model in the segmentation of lung nodules,this paper designs a new network model structure RIU-Net based on the U-Net model,which combines The advantages of Residual Network and Inception network are used.The residual and parallel convolution are used to replace ordinary neural units as basic blocks,and the two-dimensional convolution operation in the original U-Net is changed to a three-dimensional convolution operation.The introduction of the residual network avoids the problem of gradient disappearance during the network model training to a certain extent,and accelerates the convergence of the network model;the Inception module can parallel the convolutional layers in the network model,making the convolution after parallel The layers have convolution kernels of different sizes,which makes the network model adaptable to multiple scales.The three-dimensional convolution operation improves the model's ability to extract spatial information from three-dimensional medical images.In addition,the batch normalization layer(BN)introduced by the model makes the neural network loss function space smoother and increases the model's robustness.Experimental results show that the RIU-Net model has effectively improved the accuracy of lung nodule segmentation compared to the U-Net model.Second,the RIU-Net model has a poor effect on feature utilization and lacks pertinence.This paper combines the Squeeze-and-Excitation(SE)and pyramid pooling modules to design on the basis of RIU-Net The network model structure RISEU-Net is introduced.Compared with the RIU-Net model,this model has a larger receptive field,and can suppress the useless features in the image by weighting the feature channels.In addition,a new loss function is designed to reduce over-segmentation and undersegmentation in the network model.Experimental results show that this model has further improved the accuracy of lung nodule segmentation compared to the RIU-Net model.Thirdly,there is a large number of false positive lung nodules in the candidate region obtained by the above segmentation model.In this paper,a false positive screening model for lung nodules based on 3D CNN is designed to further screen the segmentation results and determine whether the candidate lung nodule regions obtained by segmentation are true nodes.Experimental results show that classifying the candidate lung nodules generated by the segmentation model can effectively reduce the number of false positives of lung nodules.