Study Of Detection Methods Of Pulmonary Nodules Based On CT Images

Lung as an important organ of human gas exchange,often accompanied by many diseases,which seriously affect human health and safety of life.With the continuous improvement of CT imaging resolution,more and more small lung tissue information is revealed from CT images,which poses serious challenges for doctors to detect suspected pulmonary nodules based on CT images.Therefore,based on the suspected Pulmonary nodule extraction decision came into being.At present,computer-aided determination of pulmonary nodules are mostly based on two-dimensional features of pulmonary nodules such as texture,shape,size,sphericity,and average grayscale,resulting in a large number of false-positive pulmonary nodules and reducing the accuracy of the discrimination.However,most physicians determine that pulmonary nodules are mostly based on the three-dimensional characteristics of pulmonary nodules.Therefore,it is a beneficial attempt to extract the three-dimensional feature information of suspected pulmonary nodules based on the CT images and to make computer-aided determination accordingly.Lung nodule detection program is divided into three steps: 1.based on CT images of the lung parenchyma segmentation;2.suspected three-dimensional feature extraction of pulmonary nodules;3.pulmonary nodules based on three-dimensional characteristics of suspected pulmonary nodules.This article has carried on the thorough research to the relevant algorithm of the above three steps,the main work includes the following three aspects:1.In order to realize the effective segmentation of pulmonary parenchyma in CT images,an improved algorithm of Otsu method is proposed by comparing the Otsu method,the maximum entropy method and the k-means clustering algorithm,and its application in lung parenchyma segmentation Extraction of the rough outline of the lungs,on the basis of the physical division of the left and right lungs.2.In order to extract the three-dimensional feature information of suspected pulmonary nodules,a three-dimensional DBSCAN algorithm is proposed to improve the density neighborhood and to apply it to extract the three-dimensional features of suspected pulmonary nodules.However,due to the large initial set of core points,the proportion of invalid information finally extracted(non-pulmonary nodules)is too large,which seriously affects the efficiency of adjudication of suspected pulmonary nodules.To solve thisproblem,a fusion algorithm of Fast-RCNN and 3D DBSCAN algorithm is designed.Fast-RCNN was used to extract the two-dimensional feature of suspected pulmonary nodules in CT images and the extracted results were used as the initial core set of three-dimensional DBSCAN algorithm to extract the three-dimensional features of suspected pulmonary nodules.3.Taking the three-dimensional feature extracted from the fusion algorithm of convolutional neural network and three-dimensional DBSCAN algorithm as input,the three-dimensional convolution neural network model was used to determine the positive pulmonary nodules.From the judgment results,we can see that compared with the two-dimensional characteristic lung nodule detection algorithm,the false positive rate of the suspected pulmonary nodule extracted by this paper is lower than the one based on the two-dimensional characteristic pulmonary nodule detection algorithm,thus improving the accuracy of pulmonary nodule detection.