Research On Segmentation And Classification Of Benign And Malignant Pulmonary Nodules Based On Deep Learning

With the continuous improvement of computer imaging and artificial intelligence technology,the automatic detection rate of lung cancer has been significantly improved,and the technical barriers of traditional methods have been gradually broken.The early symptoms of lung cancer are difficult to detect,and are often clinically diagnosed as middle to late stage(II to IV),even with standard treatment,the five-year survival rate is still less than 5%.However,if it is detected and treated in time,the survival rate of lung cancer patients can reach 80%.Lung cancer often appears in the form of lung nodules on imaging,so early detection,diagnosis and treatment are very important.Due to differences in understanding and experience of pathological images,subjectivity and instability often exist in the empirical analysis and evaluation of lesion boundaries.At the same time,as the volume of lung CT data continues to grow,a whole lung scan can produce 10500 images in a routine case,and more images will need to be processed if extensive lung cancer screening is performed.Computer aided tuberculosis diagnosis system based on deep learning can reduce people is workload and improve diagnosis accuracy.This paper aims to study the key technologies of computer-aided diagnosis of lung cancer through deep learning,and propose solutions,including lung nodule data preprocessing,lung nodule segmentation,and classification of benign and malignant lung nodules.Current methods for diagnosing lung nodules are mostly based on taking a CT scan of the entire lung,then isolating the nodules and identifying lesions by identifying and classifying positive nodules.However,this approach lacks an understanding of the region of interest(ROI),limiting the accuracy of segmentation and classification methods.The texture and shape characteristics of lung nodules are similar to those of other non-nodular tissues,which results in inconsistent performance and low classification accuracy in the segmentation task of different types of nodules.Therefore,this paper studies and discusses two new methods to solve these problems.The first method is lung nodule segmentation based on SKV-Net network.The method improves segmentation accuracy by identifying ROI,thereby avoiding classification errors due to similar texture and shape features.The SKV-Net network enables efficient segmentation of lung nodules by combining multi-resolution analysis and spatial attention mechanisms.The second method is based on the improved 3D-CNN classification of benign and malignant pulmonary nodules.Through a more detailed analysis of ROI,it is possible to effectively distinguish between benign and malignant pulmonary nodules.The main innovation of these two methods lies in the more accurate understanding and analysis of ROI,which improves the accuracy and efficiency of lung nodule segmentation and classification.The main work contents are as follows.(1)Study on segmentation of lung region.In the computer-aided diagnosis of pulmonary nodules,the primary task is to select appropriate data and preprocess it to meet the input requirements of the algorithm.The objective of this study was to obtain experimental data suitable for semi-automatic segmentation and benign and malignant classification by cutting and preprocessing the basic data set without changing the Lung Nodule diagnosis database,Lung Nodule Analysis 16(LUNA16).In this chapter,V-Net is used as the baseline for improvement,making the model lightweight and adaptive to different size target segmentation tasks,and finally a lightweight segmentation network SKV-Net is obtained.The adaptive target size based on multi-scale makes the network segmentation of large target lung areas and small target lung nodules achieve good results.(2)Segmentation of pulmonary nodules.Pulmonary nodules differ in morphology and texture,but they are visually very similar to lung tissue,making the distinction between nodules and tissue difficult and resulting in less accurate segmentation.In this chapter,SKV-Net is used as the backbone network and soft attention fusion multi-scale information design is adopted to make the edge segmentation of small target segmentation tasks such as pulmonary nodules clearer in the model.The effectiveness of this method is proved by the comparison of comprehensive indexes.(3)Classification of benign and malignant pulmonary nodules.A new classification method for benign and malignant pulmonary nodules based on improved 3D-CNN was proposed to solve the classification difficulties caused by different sizes and shapes of pulmonary nodules in medical images.In order to better extract the information features of lung nodules,the 3D-CNN model is improved by using the segmentation results of lung nodules obtained above.Through comprehensive experiments,the model shows good performance in the classification task of benign and malignant pulmonary nodules.