The Research Of Brain Tumor Image Processing Based On Sparse Representation Model

Computer aided diagnosis using brain tumor image processing technology can largely help doctors improve the efficiency and accuracy of checking patients with brain tumors. And it has important clinical significance to improve the cure and survival rate of patients. Medical brain tumor images have low gray scale contrast between organizations. Boundaries between organizations are very blurred, and the same situation exists between normal tissue and tumor region. Moreover, the complicated brain tissue structure brings about partial volume effect in two-dimensional images. All of these make brain image processing difficult.This thesis does research on image processing mainly from two aspects:brain image restoration and brain tumor extraction. Traditional algorithms of image restoration often have the defects such as high computational complexity. In recent years, the restoration algorithm based on sparse representation has become the hot spot of research. And nowadays, in the medicine field, brain tumor extraction mainly depends on the segmentation algorithms. There are already various segmentation methods, but because of the particularity of brain tumor images, a method which is very effective has not been proposed by now. The purpose of this thesis is to seek a more automatic, rapid, accurate and robust processing method of brain tumor images.The linear combination of several basis functions can show original images in the sparse form. The method is called sparse representation. It can demonstrate key features of the information in original images, and reveal the inner structure and the essence attributes of them. In this thesis, brain image restoration and tumor regions detection based on sparse representation model are investigated. The main contents are as follows:The magnetic resonance imaging technology and the characteristics of magnetic resonance brain images are described firstly. And then we introduce the background and significance of brain tumor images processing. The characteristics and difficulties of brain tumor images processing at the present stage are analyzed. At last, we briefly expound the current research status of image restoration and tumor detection around the world.Sparse representation model is the theoretical basis of the research. In this thesis, we introduce the background and basic theory of sparse representation model, and then we analyze two kinds of algorithms to solve the sparse representation model. Finally, we briefly introduce the application of sparse representation model in the field of image processing.Based on the sparse representation model, we propose an algorithm based on L1norm principal component analysis for restoration of brain images and tumor detection. The L1norm principal component analysis can recover feature images of brains and can detect the tumor area at the same time. Compared with traditional principal component analysis algorithm shows the effectiveness of the proposed algorithm in recovering brain images. And the application in extracting tumor areas overcomes the defects of traditional segmentation algorithms. Noise often makes the boundary of tumor areas discontinuous in traditional segmentation algorithms. The results show that the algorithm in this paper is more robust to noise.In addition, an algorithm based on robust principal component analysis is proposed to detect and extract brain tumors. And it takes full advantage of the rotation resistance factor in the RASL algorithm. The algorithm achieves tumor detection and extraction from a set of brain images at one time. The experimental results show that the algorithm is not affected by factors such as image rotation and overcomes the defects of some segmentation algorithms. In them, if the tumor is very small, the algorithm will lose efficacy. Moreover, the algorithm implements a fully automatic detection and extraction compared with traditional segmentation algorithms. The experimental results show that the proposed algorithm has higher accuracy.