Application Of Deep Learning Method In Pulmonary Lobe Segmentation And Pulmonary Nodule Detection In CT Images

Pulmonary disease,as one of the diseases with the highest incidence,has been affecting the health and well-being of many patients.Because most pulmonary diseases are difficult to detect at an early stage,and it is generally manifested in the form of pulmonary nodules at first.Also,Computed tomography(CT)is the preferred method for non-invasive detection of pulmonary nodules in clinical practice.If the lesions can be detected early in the routine CT imaging physical screening,and a certain auxiliary diagnosis can be given,early and effective treatment can greatly reduce the fatality rate of pulmonary diseases.Therefore,designing and developing an effective method for detecting pulmonary nodules,reducing false positives and give some auxiliary diagnosis will help doctors to formulate a reasonable diagnosis and treatment plan.This will be a work with theoretical research value and clinical application value.Consequently,the application of deep learning method based on pulmonary lobe segmentation and pulmonary nodule detection in CT images,as well as the auxiliary diagnosis of lesion structure and quantitative analysis of pulmonary function,is carried out in this thesis.The main content of this thesis can be divided into three parts.The first part is the study of pulmonary tissue segmentation in CT images,which includes pulmonary parenchyma segmentation,pulmonary lobe segmentation,pulmonary blood vessel segmentation and pulmonary tissue superpixel segmentation based on contour-based convex polygon.The second part is the research of deep learning method in the detection of pulmonary nodules in CT images,including the detection of pulmonary nodules and the strategy of reducing false positives.The third part is the research of auxiliary diagnosis of lesion structure and quantitative analysis of pulmonary function.In pulmonary parenchyma segmentation and pulmonary lobe segmentation,this thesis proposed a deep mutual learning segmentation method based on heterogeneous models,trains 3D U-Net and V-Net models as two heterogeneous models at the same time,restricts each other through the KL divergence value of the output results of the two models,realizes the mutual supervision between the models,and continuously iterates,so as to improve the stability of network training and the efficiency of the model.Under the test,the DSC index of pulmonary parenchyma segmentation of normal CT is 98.08%,and the DSC value of pulmonary lobe segmentation can higher than 90%,which is significantly improved compared with the traditional methods.Meanwhile,CT images including COPD,pulmonary fibrosis,new crown pneumonia and other pathological images were trained.The test results showed that the model can adapt to multi pathological data,and has strong robustness and model segmentation effect.In the study of pulmonary tissue superpixel segmentation,a contour-based convex polygon superpixel segmentation method is proposed in thesis.The contour characteristic information of the image is obtained by the canny algorithm and the image expansion method,the anchor points are selected around the contour,and the Voronoi diagram is constructed by the anchor point homogenization.Compared with the traditional convex polygon super-pixel segmentation method based on LSD line detection,the contour based super-pixel segmentation method can obtain more image information,and the method solve the problems of unclear super-pixel construction at the contour in the traditional SLIC and other clustering based super-pixel segmentation methods.Tested on the BSDS300 image dataset,the US index of this thesis is better than the above-mentioned traditional methods.Finally,this thesis applies this algorithm to lung images,and also compares and analyzes the traditional methods,which also shows its advantages.In the research of pulmonary nodule detection,this thesis is mainly based on the framework of deeplung model.Aiming at the problems of insufficient data,poor data diversity and unbalanced data under the specific task of pulmonary nodules,the optimization strategies such as multivariate data fusion,expansion of the number of input channels,selection and design of loss function and post-processing are used to improve the experimental effect.Also,Due to the problem of high false positives after detection of pulmonary nodules,three different strategies were proposed to solve the problem:confidence threshold screening,morphological threshold screening and constructing a false positive network model.Finally,the confidence threshold method is used after using deeplung model and we test the results in the mixed dataset of the public dataset and the data provided by a hospital in Shandong.The accuracy rate is 93.45%and the sensitivity is 95.9%.In the study of auxiliary diagnosis of lesion structure and quantitative analysis of pulmonary function,it consists of two parts.The first part is a simple judgment of the structure of nodule lesions,which mainly includes four parts:the determination of the size of pulmonary nodules,the location of pulmonary nodules,the analysis of morphological types of pulmonary nodules,and the three-dimensional visual analysis of pulmonary blood vessels.The second part is to carry out the quantitative analysis of pulmonary function,and to use some quantitative indicators of the pulmonarys to achieve the purpose of auxiliary diagnosis.In this part,indicators such as emphysema index,low attenuation area ratio,and functional pulmonary volume ratio are mainly used.Through the statistics and comparisons of these indicators in different pathologies,calculating the statistical value range of these indicators.The actual test shows that the proposed method for auxiliary diagnosis of lesion structure and quantitative analysis of pulmonary function can effectively assist doctors in pathological diagnosis.