Research Of Pulmonary Parenchyma Segmentation And Pulmonary Nodules Classification Algorithm Based On Thin CT Sequence Images

Early lung cancer screening can help patients find the disease and treatment early,and can effectively reduce mortality rates.Thin-section low-dose CT scanning technology has been widely used in early lung cancer screening,because of the small dose of radiation and high accuracy of diagnosis.However,the large volumes of scanned images produced by the technology dramatically increase the intensity of radiologists,which can lead to misdiagnosis or missed diagnosis and can't effectively improve the early detection rate of lung cancer.In the face of such situation,computer aided diagnosis(CAD)technologies based on the images from thin-section low-dose CT become the research hotspots.In computer aided diagnosis(CAD)systems based on images from thin-section low-dose CT,the segmentation of lung parenchyma is the basis of diagnosis of lung nodules(primary morphology of lung cancer)and a necessary procedure to reduce the computation burden of CAD systems.In addition,the performance of pulmonary nodules classification is the key to evaluate a CAD system.Therefore,the two aspects of content above have become hotspot and difficulty in the study of lung CAD.This research focuses on two key aspects based on the sequential images from thin-section low-dose CT: pulmonary parenchyma segmentation and pulmonary nodules classification.First,we propose an improved geodesic active contour(GAC)algorithm to segment lung parenchyma.This algorithm targets two challenges in pulmonary parenchyma segmentation of sequential CT images: 1)significant manual intervention is required;and 2)similarity between adjacent images could be easily overlooked during the segmentation process.This algorithm includes two steps: first,the pulmonary parenchyma in the middle of the CT sequential images is segmented with a spectral clustering algorithm based on Nystrom;the second step is an iterative process,in which the improved GAC model is used to segment lung parenchyma of a target image and the initial condition of the model is the contour of segmented pulmonary parenchyma from the image that is adjacent to the target image and has been segmented.The improved GAC model makes two contributions: 1)applying the spectral clustering algorithm in the segmentation of pulmonary parenchyma;and 2)improving the energy function of the GAC model by adding a gray similarity information term based on the feature of slow change in the gray statistics between adjacent images.Experimental results on 20 group clinical datasets demonstrate that the average values of XOR measure coefficient,Hausdorff distance and Jaccard similarity coefficient corresponding to the segmentation result by our method have been improved effectively compared with the segmentation result by the radiologist.The second aspect concerns pulmonary nodules classification.To improve classification accuracy,we propose a pulmonary nodules classification model,CNN-DPL,by combing the convolutional neural networks(CNN)with dictionary pair learning algorithm(DPL).Using those pulmonary parenchyma images with pulmonary nodules as datasets for the input of CNN-DPL,we extract the features of lung parenchyma images with CNN firstly,and then replace the classification layer of the CNN with a DPL algorithm,using the extracted features as the input of the DPL algorithm to complete lung nodules classification.The proposed CNN-DPL model makes two contributions: 1)this model combines CNN and DPL algorithm in pulmonary nodules classification by replacing the classification layer of the CNN with a DPL algorithm;and 2)we propose a two-step training model for the CNN-DPL that includes pre-training and fine-turning parameters for CNN,as well as training dictionary pair classifier layer.Experimental results show the values of three measurements(classification accuracy,sensitivity and specificity)of CNN-DPL model can meet the expectation of radiologists.