Pre-diagnosis Research Of Lung Cancer Based On CT Image

Lung cancer has long been difficult for the medical community to surrender.It seriously threatens the lives of human beings and is the most deadly cancer in the world.Scientific research shows that early detection and diagnosis of lung nodules can significantly increase the survival rate of lung cancer patients.In order to help experts diagnose pulmonary nodules easily,this paper proposes a new method of diagnosis of pulmonary nodules,and uses the LIDC-IDRI public database to evaluate system performance as a future research work's benchmarks.The thesis describes the relevant common techniques of pulmonary medical images and the diagnosis of lung cancer,and proposes a computer-assisted diagnosis of lung cancer consisting of three parts: image preprocessing,nodule detection,and diagnosis of malignant nodules.In the preprocess stage,the median denoising and Gaussian filtering are selected by comparing various denoising algorithms.This can remove the image noise and preserve the required details well.At the same time,the boundary contour of the image can be highlighted.At the detection stage,the threshold segmentation and morphological operation were mainly used to segment pulmonary nodules,and then the difference between the nodules and other lung tissues(such as blood vessels and trachea)in image gray values was considered.The false positive rate was reduced by introducing the Tsallis entropy value.Then the candidate nodule features extracted were fused,and the improved SVM-RFE algorithm was combined with the simulated annealing strategy to optimize the selection process.Then the sequential method was adopted to select the feature set with the smallest redundancy and correlation.Later,cross-validation was used to train and test in the SVM classifier.At the diagnosis of malignant nodules,72 cancerous pulmonary nodules and 52 non-cancerous lung nodules were selected and classified in the previous stage,and four radiological experts rated the main pathological features.The two-dimensional and three-dimensional texture features of the nodule are extracted using the gray level co-occurrence matrix and the gray-scale spatial dependency matrix respectively.After the feature normalization and nodule similarity measure,a new nodule feature subset was generated,and a SVM classification model based on decision tree is constructed(DDAG SVM)for training and testing of the final features.Through the experimental comparison on the public data set,the accuracy of nodule segmentation was 96.25%,the sensitivity of the CADe system was 85.6%,and the average FP of each scan/case was 4.0.At last the CADx had an Az value of 0.88.The experimental results show that the classification model proposed is feasible and has achieved better classification effect compared with the traditional classification model.The effectiveness of the experimental method has been tested.