Brain Tissue And MS Lesion Segmentation Methods For3D/4D MR Image

Due to the characteristics of non-invention, non-harmful and seldom being effected by object motion, the magnetic resonance imaging has been widely used by radiologists and researchers and provides huge amounts of medical data for clinical research. The MR image segmentation, which is able to provide lots of information about the shape and statistics of the brain tissue and the pathology now, is very improtant in clinical research. These informations provided by segmentation results can be used for the quantification and diagnosis of brain diseases. Besides this, the longitudinal data (4D data) processing is able to dynamically analyse the changes of the brain tissue and the lesions. Therefore, it is very useful to do brain MR image segmentation for predicting the brain disease, studying the lesion change and estimating the effect of the therapy and recovery. The Automatic segmentation and analysis of MR images are very important and necessary, because the manual segmentation for huge amounts of data by the specialists is very time-consuming and very expensive. However, due to the noise, bias field and partial volume in MR images, the traditional segmentation methods are not able to get the accurate segmentation results and are used in clinical application. Besides this, the3D MR image segmentation methods cannot be used directly for4D-image segmentation, because they only segment the3D data in each time point and may not use the statistical dependency over the time change. Therefore, it is necessary to propose a fast and accurate segmentation method by using the statistical information in time dimension.Focusing on above problems and using the MR image analysis of the brain atrophy and MS lesion, we study the characteristic of3D and4D MR image and propose the3D and4D brain tissue and MS lesion segmentation methods. Our work mainly contains the following parts:(1) To overcome the poor brain tissue segmentation due to the intensity inhomogeneity of MR image, we propose a brain tissue segmentation and bias correction methd based on mutual information. In this model, the brain tissue and the corresponding membership is expressed by the mutual information. The bias field is modeled as a linear combination of a set of polynomial basis functions and is incorporated inot the Gaussian probability density function. The Split Bregman algorithm is used for solving the membership during the energy minimization. Compared with other brain tissue segmentation and bias correction methods, this method is able to obtain more accurate segmentation results of white matter, gray matter and CSF effectively. (2) Due to the temporal variability in shape, different imaging equipments and parameters, estimating anatomical changes in longitudinal studies is significantly challenging. In this paper, we propose a algorithm for longitudinal MR brain image segmentation by combining intensity information and anisotropic smoothness term that contains a spatial smoothness constraint and longitudinal consistent constraint into a variational framework. Our model combines the intensity in each time point and spatial-temporal variability, which makes it allowing both segmentation accuracy and longitudinal consistency for analysis of anatomical changes over time. The test results on simulation data and clinical data from OASIS database demonstrate that our method is able to provide the stable4D brain tissue segmentation and is in accordance with the longitudinal characteristic of slowly changeing.(3) Due to the different shapes, regions, brightness and texture, as well as noise, bias field and poor skull stripping results, it is a huge challenge to segment the MS lesion. It is a good choice to use the information fusion from the multi-modality data to improve the accuracy of the tissue or lesion segmentation results. According to different imaging parameters of different modalities, the MS lesion has different brightness in different modalities. In FLAIR image, the intensity of MS lesion is brightest. In T2image, the intensity of MS lesion is similar with the intensity of CSF. While, its intensity is similar with the intensity of gray matter in T1image. In this papre, we propose a novel method based on information fusion and level set method for multiple sclerosis (MS) lesion segmentation in FLAIR and T1images. Firstly, the outlier detection and level set method are used to detect the edges of brightest regions in FLAIR image. Then, we use the space position relations from T1image to reduce the false positive results. The testing results on open source database and clinical database demonstrate that our method is able to effectively detecte MS lesion boundaries and provide extremely stable segmentation accuracy.(4) The longitudinal MS lesion segmentation and analysis is a key point for studying the MS lesion changes in different time, characterizing lesion evloution and detecting active lesions. In this paper, we propose a4D MS lesion segmentation method based on the information from the temporal dimension. Firstly, we use the3D MS lesion segmentation method to segment the3D MS lesion data separately in each time point as the initial segmentation of our4D method. Then the prior knowledge is constructed by the difference of images in two time points and the3D segmentation results. We propose a level set method with this prior knowledge. At last, the active regions are detected and the global volume changes are measured according to the4D MS lesion segmentation. The experimental results demonstrate that our method is robust to segment the longitudinal MS lesions, detect the active regions and measuring the volume.