The Study Of Swarm Intelligence Based Feature Selection Algorithm For Medical Images

With the development of digital information, the image processing technology has been applied to computer-aided diagnostic techniques. Computer-aided diagnosis technology uses computer to make images digital, applies a series of treatments to images using digital image processing technology, thereby providing doctors with objects easy to identify and freeing them from cumbersome and mechanical repetition of classification problem. CAD technology also reduces misdiagnosis, and at last achieves the purpose to improve the diagnostic accuracy. In the machine learning process of CAD technology, for example diagnosis of medical images, feature selection problem is the most important part as a typical global optimization problem.Swarm intelligence methods were introduced in this thesis for the study of feature selection and Levy flight and PSO were chosen. Depending on their characteristics, two feature selection methods are designed and both of them have different advantages.A Levy flight random process based feature selection algorithm (LevyFS) is proposed to improve the speed of feature selection method. A multi-stages heuristic search strategy was defined during optimization process, and Levy flight random process was introduced in local search behavior, and map between Levy flight distances and search operations was defined. During different search stages, maps were introduced to change the probability of search behavior so that the direction and speed of local search behavior was controlled. Local optimum problem was avoided which appears in some feature selection algorithm. Experiment results show that LevyFS algorithm overcomes the limitation of heuristic methods and the average cost is only 1/3 of SFFS algorithm.A BPSO feature selection method which combined global best and personal best is proposed. Probability is introduced to control motion direction, and both social sharing and personal cognitive abilities are used for feature subset optimization.Finally, these methods were applied to medical image datasets, including breast mammogram images and magnetic resonance (MR) image of gliomas. Satisfied results were obtained by the experiment.