Research On The Medical Image Segmentation Based On Non-negative Matrix Factorization

Brain diseases and injuries have been serious threats to people’s lives and health in recent years. People have paid more and more attention to brain health problems. It is the current hotspot of medical research to make qualitative and quantitative analyses on brain tissue using medical images.Image processing plays an important role in the medial diagnosis.As the basic image processing, MR image segmentation provides the necessary data used by the high-level processing, and the segmentation results directly influences the conduct of subsequent processing. However, medical image segmentation is still a difficult problem because MRI has the character of strong noise, weak gray-scale contrast and blur organizational boundaries and so on.Clustering method is widely used in the image segmentation. This method can cluster the different data to the distinct clusters. The data have great similarity in the same cluster and dissimilarity in the different clusters. Non-negative Matrix Factorization (NMF) method now is the current research focus in the clustering methods. NMF imposes a strict non-negative constraint in the process compared with the traditional matrix factorization methods, and can guarantee that the original data can be descripted additively. The results of NMF are the basic matrix and the weighted coefficient matrix which represents the clustering tag of pixels of an image in the image segmentation. So we can get the clustering result of pixels by NMF processing and finish the medical image segmentation. Compared with the traditional clustering methods such as k-means method. NMF has the nature of soft clustering which is more suitable for the medical image segmentation.In this thesis, we make use of the clustering nature of Non-negative matrix factorization to segment the medical image, and the contents are as follows:1. This section introduces the theory of magnetic resonance technology, also the structure organization and grayscale characteristics of the MRI. Then we introduces the background and importance of medical image segmentation, and the difficulty faced by medical image segmentation. Finally we briefly present the common methods used in the medical image segmentation.2. We describe the background and the mathematical expression of the NMF, then describes the iterative rules and convergence of the NMF. Finally we analyze the research status and application about the NMF.3. We propose an algorithm for white matter segmentation of brain MR image based on local Walsh transform and non-negative matrix factorization. Local Walsh transform belongs to the digital image processing family, and it has better texture discrimination performance compared with other texture feature extraction methods. The algorithm first operates a local Walsh transform on the pixel of image, and selects the feature with the good texture discrimination to get the feature matrix, and then acquire the results of segmentation by non-negative matrix factorization on the feature matrix.The experiments show that the proposed method is simple with high precision, and can gain reasonable results.4. We combine the NMF method and graph cut theory. The image segmentation methods based on graph theory change the clustering problem into the graph division problem without regarding the shape of the sample space, so they can obtain the global optimal solution. This thesis takes advantage of the NMF and graph theory to the segmentation. Firstly we establish the undirected weighted graph corresponding to the medical image, and then segment the image using the non-negative laplacian embedding to get the segmentation results.The experiments show that the proposed algorithm can acquire better segmentation results comparing to traditional laplacian embedding methods.