Lung Nodule Segmentation And Benign Or Malignant Classification Based On Deep Learning

The rise of computer imaging technology and artificial intelligence has dramatically increased the rate of automatic diagnosis of lung cancer.Lung cancer has no unique symptoms in the early stages of the disease,or its symptoms are extremely difficult to detect,and the clinical diagnosis of lung cancer is mostly intermediate to advanced(stage ?-?).Despite standardized treatment,the 5-year survival rate of patients is still less than 5%.If early patients can be treated in time,the survival rate can reach 80%.The initial manifestations of lung cancer in imaging are mostly lung nodules,so early detection,early diagnosis,and first treatment are essential for patients.At the same time,because different doctors have different understandings of pathological images and there are differences in diagnosis and treatment experience,radiologists often analyze the lesion boundaries and make judgments through experience,which makes the reading results subjective and unstable.With the sharp increase in CT image data of the lungs,a full lung scan of standard cases will produce 100-500 images.If large-scale lung cancer screening is performed,it is even more necessary to screen and judge from massive images.Reading each image not only costs A lot of time and energy,and long-term observation is straightforward to understand the film fatigue,resulting in misjudgment.Computer-aided diagnosis system for lung nodules based on deep learning can effectively reduce the workload of radiologists and improve the accuracy of diagnosis.This thesis focuses on the fundamental techniques of computer-aided diagnosis of lung cancer based on deep learning.It proposes solutions to the critical problems in the construction of lung nodule database preprocessing,lung nodule segmentation,and benign and malignant classification of lung nodules.Most of the existing diagnosis methods of lung nodules are based on whole lung scanning,segmentation of nodules in the whole CT image,and on this basis to determine the classification of positive nodules.However,this method lacks the understanding of the specific location of lesions,that is,region of interest(ROI),which limits the segmentation method and classification accuracy.Therefore,in view of the visual similarity between pulmonary nodules and other non-nodule tissues in texture and shape features,which results in the different performance of the algorithm in various nodule segmentation tasks and the low classification accuracy,this paper proposes a lung nodule segmentation method based on multi view residual block network and a lung nodule benign and malignant classification method based on multi input convolution neural network,the main work and innovation points as follows:(1)Selecting data and cropping it to meet the input requirements of the algorithm is the key to the computer-aided diagnosis of pulmonary nodules.Without changing the underlying database Lung Image Database Consortium / Image Database Resource Initiative(LIDC-IDR),to reduce the ambiguity of the data,the necessary data set is trimmed and preprocessed,the filtering rules are designed,and the slice sequence extraction method is obtained.Training,verification,and test data for semi-automatic segmentation methods and benign and malignant classification algorithms.(2)This article addresses the problem that different types of nodules in the lungs have different shapes and contrasting textures,and some types of nodules are visually very similar to lung tissues.The lung nodule segmentation method based on multi-view residual module convolutional neural network applies to all kinds of nodules proposed in this thesis,and comprehensive experiments verify the performance of the method.(3)A benign and malignant classification method of lung nodules based the multi-input convolutional neural network was proposed.This research focuses on the classification of benign and malignant lung nodules in medical images.Due to the different sizes and shapes of the lung nodules,classification is difficult.Based on the lung nodule segmentation results obtained above,the multi-input module information of the pulmonary nodules is designed to facilitate the network to extract feature such as the edges and texture of the pulmonary nodules and apply them.Sub-branch structure for three different inputs.Extract and fuse features to obtain classification results,and finally verify the model performance through comprehensive experiments.