Brain Image Segmentation Based On Multi-weighted Probabilistic Atlas

One of the prime processes of medical image analysis is image segmentation. Image segmentation is an important pre-processing step. Image segmentation is the process to partition the image into non-overlapping, constituent regions which are homogeneous with respect to characteristics such as intensity or texture. This preliminary low-level vision task has been widely studied in the last decades since it is of critical importance in many image understanding algorithms, computer vision and graphics applications. Medical image segmentation is an important part of medical image analysis and study, and has a great value for computer-aided diagnosis, cancer radiation therapy technology, medical operation,3D visualization and other areas of medicine. In the segmentation of medical images, the objective is to identify different regions, organs and anatomical structures from data acquired via MRI or other medical imaging technique. Segmentation is a technique in medical image processing and it has been useful in many applications including lesion quantification, surgery simulations, surgical planning, multiple sclerosis, functional mapping, computer assisted diagnosis, image registration and matching, etc. However, there is no one standard segmentation technique that can produce satisfactory results for all types of imaging applications.In order to solve the problem of medical image segmentation, researchers have done a lot of research work to set up various segmentation models respectively based on the basis of gray, texture, area and other various constraints. The basis image segmentation algorithms can be divided into eight categories:(1) Thresholding approaches, (2) Region growing approaches, (3) Boundary checking approaches, (4) Classifiers, (5) Artificial neural networks, (6) Wavelet transform models, (7) Active contour models, (8) Markov random field models. Thresholding is a simple yet often effective means for obtaining segmentation in images where different structures have contrasting intensities or other quantifiable features. However, Thresholding typically does not take into account the spatial characteristics of an image. This causes it to be sensitive to noise and intensity inhomogeneity. Thresholding is often used as an initial step in a sequence of image processing operations. Region growing is a robust, simple, fast segmentation method. Region growing’s primary disadvantage is that it requires manual interaction to obtain the seed point. Thus, for each region that needs to be extracted, a seed must be planted, However, Region growing suffer from sensitivity to the selection of initial seed points. Region growing also be sensitive to noise, causing extracted regions to have holes or even become disconnected. Conversely, partial volume effects can cause separate regions to become connected. Classifier methods seek to partition a feature space derived from the image using data with known labels. Clustering algorithms require an initial segmentation (or equivalently, initial parameters). Clustering algorithms do not directly incorporate spatial modeling and can therefore be sensitive to noise and intensity inhomogeneity. Artificial neural networks (ANNs) use training data determining the weights and then the ANNs is used to segment new data. A difficulty associated with Markov random field (MRF) models is proper selection of the parameters controlling the strength of spatial interactions. Too high a setting can result in an excessively smooth segmentation and a loss of important structural details. In addition, MRF methods usually require computationally intensive algorithms. Recently, level set methods have become increasingly popular for use in medical image segmentation. Level set methods have several desirable characteristics such as sub-pixel accuracy, ease of formulation and flexible incorporation of prior knowledge. In the traditional level set method, a signed distance function is commonly used as the initial level set function. An obvious shortcoming of the signed distance function is periodical re-initialization, which is necessary to maintain a stable evolution of the zero level set curves. The re-initialization of the level set function has an intensive computation cost and a side effect of numerical inaccuracy. In addition, it is difficult to determine when and how to re-initialize the level set function.Human brain imaging was preferred in the study of the various tissues through MR imaging. The majority of research in medical image segmentation pertains to its use for MR images, especially in brain imaging. This is because of MR’s ability to provide a combination of high resolution, excellent soft tissue contrast, and a high signal-to-noise ratio. One of the prime processes of medical image analysis is image segmentation. Although classic segmentation methods work well on many applications, the models are often insufficient to reflect the characteristics of brain images, which include deformable object structures, inter-subject variability and ambiguous boundaries between organs. A major difficulty in the segmentation of brain MR images is existence of noise and the intensity inhomogeneity artifact, which cause a shading effect to appear over the image. This artifact can significantly degrade the performance of methods that assume that the intensity value of a tissue class is constant over the image. Many segmentation algorithms suffer from limited robustness to outliers, over-smoothness for segmentation and limited segmentation accuracy for image details. Initially segmentation of brain MR images has been done based manually by human experts. But manual segmentation is a difficult and time consuming task. The accuracy of the segmentation depends upon the qualified experts. The errors occur due to low tissue contrast, unclear boundary, poor hand to eye coordination and interpretation of the operator in manual segmentation. Hence, a priori anatomical information is essential for simplifying the segmentation task.Recently, a popular method to incorporate additional information in the segmentation process is atlas-based segmentation (sometimes also called label propagation). In this framework, prior knowledge is introduced through an atlas image. Here, an atlas is defined as the combination of an intensity image (template) and its segmented image (the atlas labels). After registering the atlas template and the target image, the atlas labels are propagated to the target image. At this point, the segmentation turns into a registration problem. Atlas-based segmentation has been successfully applied to a number of applications. For example, Rohlflng et al. compared several atlas-based segmentation techniques for the segmentation of structures in bee brains. Heckemann et al. segmented 67 brain structures using a non-rigid registration approach and label fusion of 29 atlases and Klein et al. applied a non-rigid registration approach to segment the prostate in 3D MR images. The design of atlas data has attracted considerable attention. The simplest approach is to use a single atlas image. However, because individual patients’anatomy varies, and image registration algorithms are not perfect, the likelihood of atlases based on a single-subject is not constructed to represent the diversity of human anatomy. To better characterize the variability of anatomical structures, multiple atlases better account for anatomical variability. Several studies have shown that multi-atlas segmentation methods can give results significantly better than using a single registration. In multi-atlas-based segmentation, multiple labeled images are registered to the target image independently, and then combine their segmentation, a strategy known as label fusion. Examples of label fusion schemes are label voting, STAPLE using expectation maximization with binary volumes, voting with globally or locally weighted votes, Spatial STAPLE represents an extension to the traditional STAPLE framework that allows for the estimation of a smooth spatially-varying performance level field instead of global performance level parameters and has been shown to provide robust and accurate multi-atlas segmentations. SIMPLE using a selective and iterative method for performance level estimation in atlas-based segmentation.Traditional multiple atlas-based segmentation methods incorporate prior knowledge about atlas’ anatomy and shape information. However, in the fusion process, most current label fusion methods complete the label fusion only use the label information. They again neglected the structures information of target image, thus lacking a mechanism to help correct possible labeling errors. In order to resolve the limitations in the conventional label fusion framework described above, we present a novel brain image segmentation based on multi-weighted probabilistic atlas in this paper. Firstly, the local similarity measure between the registered atlas image and the target image is used as a weight to compute the weighted average probabilistic atlas. Secondly, the distance field of the atlas labels is used as weight to incorporate the locality information of the atlas. Thirdly, the self-similarity of the target image is used as the weight to make small corrections and approving the segmentation. In the experiment part, many brain MR images were used to segment the hippocampus region. Experimental results indicate that the proposed algorithm outperforms the common brain image segmentation methods.