Research On Automatic Detection Of Pulmonary Nodule In High Resolution Computed Tomography Images

Lung cancer is not only a kind of malignant tumor which has the highest mortality rate worldwide,but also the highest cancer mortality rate and number of new additional disease of malignant tumor in China,brought great threat to human health.Lung cancer CAD plays an important role in early detection,but it doesn’t get great clinical results due to complex tissue structure and numerous lung diseases.How to improve the detection accuracy rate is the focus of the present research work.This paper mainly study the lung parenchyma detection,candidate nodule detection,feature extraction and false positive removal of lung cancer CAD,in order to improve the accuracy of detection.As the first step in the automatic detection of pulmonary nodules,Lung parenchyma segmentation can effectively reduce the search scope and reduce the amount of the subsequent processing tasks.The adherent vascular and pleural nodules are easy to be split out using threshold segmentation method because of the adherent vascular and pleural nodules on CT value is closed to the chest CT value,which are easy to be missed.Therefore,it is necessary to carry out lung repairing to include the type of nodules.This paper proposed an improved vector method to repair the internal lung nodule sag mainly,based on the two dimensional convex hull algorithm.This method got the initial lung parenchyma by turning the lung CT image to binary image and removing the trachea firstly.In view of the poor effect of two dimensional convex hull in repairing internal nodules sag,this paper repaired internal contour sag by using the improved vector method.It was proved that this method can effectively improve the accuracy of detection by comparing with existing methods.Detection of candidate nodules is based on the segmentation of lung parenchyma.It is difficult to extract the true nodules and retain as few false nodules as possible.This paper detected the candidate nodules based on the threshold method firstly.According to the excessive number of false positive after the detection of the candidate nodules,this paper put forward a hierarchical method to remove partial false positive.The partial small false positive was removed firstly according to the area threshold and then the partial tubular blood vessels were removed according to the degree of linearity.The experimental results showed that this method can effectively remove some false positive and reduce the computation of subsequent processing.The purpose of the removal of false positive is to retain the true nodule and remove false nodule as possible,which is the key to the lung cancer CAD.Feature extraction is the basis of the removal of false positive.In this paper,a total of 20 features,such as gray level,geometry,morphology and texture,were extracted.Because of the negative samples in the candidate nodule data were much more than the positive samples,it was a typical unbalanced data.The Gentle Adaboost method based on cross validation was used for classification in this paper.This method sought the best iterative parameters of the Adaboost algorithm by cross validation.By comparing with SVM classification algorithm,the classification accuracy of this method was proved.