Lung Nodule Computer-aided Detection Based On Selection Ensemble Algorithm

Currently lung cancer has become one of the high mortality rate malignant diseases. A large number of clinical data show that early detection and early treatment can effectively improve the survival rate of patients. Now the main screening method for detecting lung cancer is based on the sequence of CT images. Pulmonary nodules are important signal of lung lesions on CT images, especially small and solitary pulmonary nodules. Due to diversity and heterogeneous of pulmonary nodules detection performance of traditional classifiers are very poor. In recent years, ensemble learning has drawn widespread attention of the scholars in the machine learning field. Compared with using only a single classifier, ensemble learning makes the decision by multiple classifiers has stronger generalization ability.This thesis focuses on the application of integration algorithm in pulmonary nodules detection. Bagging algorithm, Adaboost algorithm and the integration algorithm based on feature selection are applied on the data of lung nodules. A novel approach based on Dynamic Multiple Classifiers Selection (DMCS) integration algorithm is proposed. Main contents and achievements of this thesis are as follows:1. We test main integration algorithms on lung nodules detection data. The experiment results demonstrate the advantage of integration algorithm in lung nodule aid detection.2. We propose a dynamic multiple classifiers selection integration algorithm (DMCS).The algorithm randomly divides the feature space into a number of feature subsets using the random feature subset selection algorithm. Because different feature subsets have different distributions, every feature subset should choose an appropriate base classifier.The experimental results show that the ensemble learning has obvious advantage in lung nodule aid detection. The DMCS algorithm proposed in this thesis has better stability and generalization ability than the existing representative lung nodule detection algorithms, and has more extensive applications.