Research On Lung Nodule Detection Based On Deep Convolutional Neural Network In Computed Tomography

Lung cancer is the most commonly diagnosed cancer in the world,which is the leading cause of cancer death in men and the second leading cause of cancer death in women.Patients with lung cancer can be treated early and then their life quality can be improved with early screeningLung cancers in early-stage often present as lung nodules.Therefore,lung nodule detection accurately is of great significance for the early diagnosis of lung cancer.The Computer-Aided Detection(CADe)system of lung nodules can assist radiologists to detect lung nodules in medical images,improve the accuracy of lung nodule detection,and relieve the workload of radiologistsScreening with low-dose spiral CT scanning is the preferred method for lung cancer.Based on the research of the existing CADe algorithms for pulmonary nodules in CT images,the experiments in this thesis had achieved satisfactory results in three tasks:lung parenchyma segmentation,lung nodule detection and false positive lung nodule reduction.The main research contents of this thesis are listed as follows(1)There are some common problems during the process of the lung parenchyma segmentation.Firstly,juxta-pleural nodules would be easily missed Secondly,adhesion would occur between the main trachea or bronchial tree and lung parenchyma.Last,adhesion would occur between the two lung regions.In order to solve the problems mentioned above,a novel method of lung parenchyma segmentation was proposed in this thesis based on the methods of the three-dimensional region growing and the gray-level integrated projection.Firstly,the lung parenchyma is roughly segmented by the threshold method and other methods;secondly,the main trachea and bronchial tree were extracted and removed by the three-dimensional region growing method;then,the gray-level integrated projection method combined with the line scan method was used to judge and separate the adhesion between left and right lung.Finally,the boundary of lung parenchyma was repaired by the rolling ball method and then the segmentation of the lung parenchyma was completed(2)Juxta-vascular lung nodules are often hard to be segmented;moreover,these nodules after segmentation are not easy to be distinguished from vessel bifurcations In order to solve the above problems,a novel method of lung nodule detection was proposed in this thesis based on multi-scale dot selective enhancement filter and weighted support vector machine.In the beginning,the nodule candidates were divided into two groups using three-dimensional labeling technology:the vessel tree group and the non-vessel tree group.For the vessel tree group,multi-scale dot selective enhancement filter was first constructed based on the Hessian matrix to segment lung nodule candidates.Then,a 3D constrained region growing algorithm was used to constrain the growth of nodule candidates.Moreover,according to the different grayscale distributions between juxta-vascular nodules and bifurcations,the features based on gradient,shape,grayscale and others were extracted.Finally,the weighted support vector machine was adopted to solve the class imbalance problem in lung nodule detection task.The proposed method could detect juxta-vascular nodules very well.For the non-vessel tree group,the obvious false positive lung nodules were removed by using the rule-based method.Then,the false positive lung nodules were further removed based on the dot selective enhancement filter.Nodule candidates are classified as true nodules if the candidates survived the rule-based classifier and were not screened out by the dot filter.The sensitivity of lung nodule detection is 87.81%,and the number of false positives is 1.057 per scan(3)In order to solve the problem that micronodules can be missed,a novel method of lung nodule detection was proposed in this thesis based on a 3D convolutional neural network combined with multi-scale prediction strategy.Firstly,the 3D convolutional neural network was constructed to detect small cubes containing lung nodules.Secondly,as small cubes of fixed size were easy to miss small lung nodules,multi-scale prediction strategy was utilized to detect small lung nodules with small-scale cubes.Finally,the Density-based spatial clustering of applications with noise(DBSCAN)algorithm was used to cluster the small cubes containing lung nodules at different scales to detect lung nodules.The proposed method can detect small lung nodules very well.The sensitivity of lung nodule detection is 92.93%,and the number of false positives is 4 per scan(4)The radiologists will have to consume much more time to review the results of CADe when there are excessive false positive lung nodules.To improve the specificity of lung nodule detection,a novel method of false positive lung nodule reduction was proposed based on multi-level dense convolutional neural network Firstly,a 3D dense convolutional neural network was constructed,and the shallow and deep features were simultaneously reused for lung nodule identification.Then,five different levels of networks were constructed to extract different contextual information.Finally,the integrated learning method was adopted to fuse the prediction results to improve the accuracy for false positive nodule reduction.The sensitivity of false positive lung nodule reduction is 92.70%,and the number of false positives is 4 per scanIn summary,the methods proposed in this thesis have advantages and novelties in lung parenchyma segmentation,lung nodule detection and false positive nodule reduction.These methods and they can assist radiologists to detect lung nodules and improve detection accuracy.