Research On Brain Tissue And Lesion Segmentation Based On Magnetic Resonance Imaging

Magnetic resonance imaging(MRI)has several advantages over other medical imaging modalities,including high contrast among different soft tissues,relatively high spatial resolution across the entire field of view and multi-spectral characteristics.Therefore,it has been widely used in quantitative brain imaging studies.Quantitative volumetric measurement and dimensional visualization of brain tissues are helpful for pathological evolution analyses,where image segmentation plays an important role.With the development of medical imaging techniques and deep study of the brain,the segmentation of brain tissues has been divided into three levels:global brain tissue segmentation,brain structure segmentation and pathological tissue segmentation.However,MR images suffer from several major artifacts,including intensity inhomogeneity,noise,partial volume(PV)effect and low contrast,which make MR segmentation remains a challenging topic.Therefore,in this thesis,we focus on brain MR image segmentation based on the non-local characteristics of brain images from three aspects,including the global brain tissue segmentation,brain structure segmentation and pathological tissue segmentation.The main work of this paper can be concluded as follows:(1)A modified Gaussian mixture model is presented to overcome the impact of noise by combining the non-local information for global brain MR image segmentation.Based on the discussion of the strategies for the segmentation of brain MR images with noise,we analysis the major drawbacks of conventional Gaussian mixture model(GMM)and the ideas of local information construction for current improved methods.Then we introduce non-local information to compute the weights of the neighborhood which makes the similarity between the pixels is replaced by the similarity between the image patches,and a search window is provided for each pixel,only the pixels in the searched window are considered.Experimental results show that our method can further improve the segmentation accuracy of Gaussian mixture model and suppress the noise effectively.(2)A novel brain MR image segmentation algorithm is presented based on Markov random field information.We discuss the strategies for the segmentation of brain MR image with intensity inhomogeneity.By using the local spatial neighborhood information,the segmentation can overcome the noise and preserve more details of the image.Meanwhile,by introducing the KL distance into the prior probability and posterior probability as an entropy penalty,the proposed algorithm could get better segmentation results through smoothing this penalty term.Experimental results show that the method can further improve the segmentation accuracy of brain MR images,and can better retain the details of the image information.(3)A progressive label fusion framework for multi-atlas segmentation by dictionary evolution is proposed for hippocampus segmentation.Different from the conventional label fusion methods,in our proposed method,we construct a sequence of intermediate dictionaries in a multi-layer manner to progressively optimize the weights for label fusion.In this way,our proposed method seeks for the representation weights that can be progressively improved for final label fusion,instead of employing only the original intensity patches based representation weights,which are used in the conventional methods.A large number of experimental results show that the algorithm can achieve higher segmentation accuracy for hippocampal segmentation.The framework can also be considered as an extension of the traditional single-layer method,thereby allowing for further applications.(4)A method for brain tumor segmentation based on multimodal MRI is proposed.Brain tumors are divided into two parts:edema and tumor nucleus by the combination of FLAIR and T1ce modalities.First,the OTSU method is used to initialize the initial contour of level set on FLAIR modality.Then we introduce the non-local patch information into a region-based active contour model to detect’ the abnormal regions on FLAIR modality,and a variation level set formulation is applied locally to approximate the contour.Consequently,k-means is applied to distinguish the edema and tumor tissues in the abnormal regions based on the T1ce modality.Compared with traditional one-modality methods,our model can better represent the specific tissue of tumor in a simple way.The validation experiments on synthetic and clinical brain magnetic resonance images demonstrate the effectiveness and simplicity of the proposed method for brain tumor segmentation.