The Research Of Tumor Interactive Segmentation Algorithms Based On MR Brain Images

Magnetic resonance imaging technology can draw high-quality 3D images of internal body structure, and it is widely applied to clinical medicine now, particularly in the diagnosis of brain tumors and other brain lesions. It is conducive to better treatment plan by extracting region of interest from MRI brain image segmentation. Therefore, the brain tumor segmentation has a very important theoretical significance and practical value in the medical field and the computer vision field. Medical image segmentation algorithms can usually be divided into manual segmentation methods, automatic segmentation methods, semi-automatic segmentation methods. Manual segmentation methods are time-consuming, currently automatic segmentation methods are unable to achieve the required accuracy in clinical application. So interactive segmentation methods become the main research direction in the area of medical image segmentation.This thesis studies GrabCut which based on interactive segmentation method, analyzes the advantages of the algorithm, and demonstrates on 2D slice images of MRI brain tumor. At the same time, this thesis presents an improved algorithm, The GrabCut is applied to 3D images of MRI brain tumor through the constraint of Structural Trajectories, which reduces user interaction and improve the segmentation efficiency. Details are as follows:(1)This thesis analyzes different segmentation algorithms which is based on threshold, edge and region according to different image features, applies these segmentation algorithm to glioblastoma multiforme datatset, and compares their performance.(2)This thesis describes the basics of graph theory, analyzes a variety of segmentation algorithms based on graph theory. Then gives a detailed description of the Graph Cuts and its improved algorithm GrabCut. Finally, the algorithm is applied to 50 slice images which belong to GBM dataset, and we calculates Jaccard score, verifies the advantages of CrabCut.(3)Considerring the above-mentioned segmentation algorithms are only applied to 2D slice images of MRI brain tumor, we propose an improved algorithm for 3D image. We need select a slice and label it according to asymmetric, then the left slices will be labeled by the algorithms GrabCut with hard constraints. Because of greatly reducing the user interaction, so there will be less decrease in the accuracy of segmentation, but segmentation has been greatly improved efficiency.