A Hybrid Model For Brain MRI Image Segmentation

As one of the leading techniques for medical imaging, MRI provides a useful tool for human brain research, and has been widely used in medicine, neuroscience, cognitive science,psychology, etc. Grey matter, white matter and cerebrospinal fluid (CFS) are three major brain tissues. Accurate segmentation of these three tissues has both clinical and academic significance.Accurate measurement of the volume and morphology of these tissues can help early diagnosis of pathology, as well as understanding the way they change during development,aging, and pathology. Besides, accurate measurement and localization of these tissue, are prerequisite of subsequent analysis, such as visualization, surgical planning, computer integrated surgery, treatment tracing, etc.However, the performance of automatic segmentation of brain MRI images is still hindered by following factors:The complexity nature of the brain anatomy, as well as the inter and intra variations of the these structure bring diffculty to brain modeling; Intensity inhomogeneity caused by inhomogeneity of magnetic field and biological variation within the same tissue;Partial volume effect. Therefore, the existing methods are still far from the clinical requirement for segmentation accuracy, speed, and automation. Therefore, automatic segmentation of brain MRI is still a challenge work.In this thesis, a hybrid model for medical image segmentation, which combines an improved watershed transform and geometric deformable model is proposed. Achievements and innovation points in this thesis are described below:1.We proposed a saliency measurement constrained multilevel immersion watershed transform method. The innovation of our method is that, in contrast to the traditional preprocessing and postprocessing method used to overcome oversegmenation, our method can suppress the oversegmenation during the watershed transform. Besides, additional saliency measurement was added as constraint to avoid important contours from being destroyed. Experiment results demonstrate the superior performance of our method over the traditional watershed method and the improved watershed method provided by ITK, in terms of segmentation accuracy and speed,and thus lays a concrete foundation for the subsequent combination with geometric model.2.Chan-Vese model was used to propagate the initial contour of object of interest which produced by our improved watershed transform, in order to correct its smoothness. This combination can avoid manual initial contour specifying, improve the speed of convergence, and further segmentation accuracy.3.A segmentation system was designed and realized for the proposed hybrid segmentation system, and was tested on real brain MRI image. We provide experiment of brain skull stripping, as well as grey matter, white matter and cerebrospinal fluid (CFS) segmentation to explain the effectiveness of these proposed methods in terms of segmentation accuracy, time complexity and automation.