Research On Feature Selection Algorithm And Its Application In Diagnosis Of Solitary Pulmonary Nodules

In the field of data mining, machine learning and statistical pattern recognition, feature selection, as an important way of data preprocessing, is an essential step of supervised learning algorithm. With the development of computer science and technology, the emergence of high-dimensional data problems such as image processing and bioinformatics propose the stern challenge to the existing feature selection algorithms. This dissertation mainly focuses on feature selection algorithm for high-dimensional data and its application in computer aided diagnosis of solitary pulmonary nodules. The contributions of this dissertation mainly include the following parts.Firstly, this paper makes a specific and in-depth research on current focus and problems of feature selection algorithm, and analyses the definition, process, classification and the routine algorithm models of feature selection. It also puts forward the skills of using the algorithm.Secondly, a new feature selection algorithm based on rough set and genetic algorithm is proposed. It combines the rough set(RS) theory with genetic algorithm(GA) properly, relative attribute dependency of rough set theory is used to design the fitness function and genetic operators. The approach is then applied to image feature selection. Experimentally, the new method illustrates the high algorithm efficiency and excellent results.In addition, on the base of merits and demerits of Filter and Wrapper model, a hybrid feature selection algorithm is proposed based on ant colony optimization. Ant colony algorithm is used to select features while support vector machine is applied to evaluate the performance of feature subsets, then feature pheromone is computed and updated based on the evaluation results. The principle is provided for feature and feature subset selection based on the methods above. Search algorithm can avoid blind searching and convergence quickly. Experiment results on 8 practical datasets show that the algorithm has good comprehensive performance with respect to accuracy, size of feature subsets, and efficiency.Afterward, feature selection algorithms are applied to computer aided diagnosis of solitary pulmonary nodules. In this section, computer aided diagnosis system of solitary pulmonary nodules is introduced briefly and the foundation of knowledge database is described, the hierarchical structure of features is proposed based on the importance of features for SPNs diagnosis. Meanwhile, both of the two feature selection algorithms above are used to calssify the artificial datasets, experiment results show that the feature subset selected by the algorithms is of not only the good classification performance, but also consistency with real facts for medical diagnosis.In the end, this paper concludes by summarizing the research and indicating its future work.