Research Of Liver Image Segmentation Based On Statistical Shape Models

Medical image segmentation is the technique that divides the image into different regions which the inter-regions have different characteristics and the intra-region is similar, or extracts the interested objects from the image. The segmentation takes an important role in the field of medical image processing and analyzing. However, medical image segmentation is put more emphasis on the extracting special regions, such as organs or cancers. As the same time, the different regions or objects should be identity with the anatomical structure. Based on medical image segmentation, the image contents can be better analyzed, which will assist doctors to diagnose and make the treatment plan correctly. Up to now, there are lots of segmentation methods for all kinds of purposes, but no one can generally apply to different modes of images and different organs. Moreover, There are a number of factors that cause current segmentation algorithms fail to satisfy the need of clinical practice, including the individual differences in the tissue anatomy; slow calculating speed and inaccuracy; and poor image quality affected by noise, intensive in homogeneity and partial volume effect(PVE) etc. In addition, it is difficult to describe the practical problem completely using mathematical models; the dissimilarity of the segmentation objectives; image degradation and so on. There is no such an algorithm of segmentation which is competent for all kinds of the problems. So we only can take the given problems and the special requires into account to develop different algorithms, and balance between the precision, the speed and the robustness. For these reasons, the medical image segmentation has been an important field in medical image processing and analysis.The liver is the largest gland in the body (approximately1500grams) which is vital organ. It has a wide range of functions, including detoxification, excretion, blood storage, protein synthesis, and production of biochemicals necessary for digestion. And the liver is necessary for survival; there is currently no way to compensate for the absence of liver function long term. So it is the important in metabolism to sustain life. Because the liver diseases are common which is dangerous, it is particularly important for liver diseases treatment.The liver segmentation is the first step for the diagnosis, research and surgical planning of the liver diseases. Therefore, the accuracy and robustness of the liver segmentation is particularly significant and has important sense. In clinic, liver manual segmentation is difficult and time-consuming for expert. So the research of automatic and precise liver segmentation is valuable. As the same time it is a difficult and hot point.At present, a number of automatic segmentation techniques for liver have been proposed, such as Statistical Shape Models, Region Growing, Atlas Registration, Level Sets and Graph cuts etc. There is a testing for all the algorithms of liver in the MICCAI2007Grand Challenge, where16teams evaluated their algorithms on a common database. A collection of20clinical images with reference segmentations was provided to train and tune algorithms in advance. Participants were also allowed to use additional proprietary training data for that purpose. All teams then had to apply their methods to10test datasets and submit the obtained results. However, the best automatic methods (mainly based on statistical shape models with some additional free deformation) could compete well on the majority of test images. Mostly, the shape models have better result need to combine the appearance models. While Cootes described the Active shape models and the Active Appearance models in1990s, wide-spread utilization of Statistical Shape Models appeared. These models integrate the shape and the appearance information. The features in those models varied and straightforward divide into region feature and edge feature. The edge features are profile, Gabor Wavelets, Bayesian and so on. Region appearance modes and in-out region histograms are features based on regions. However, the methods have their limitations and drawbacks.The challenge in liver segmentation is lying in the large variations in liver shape and in the intensity pattern inside and along liver boundaries, and the complicated texture patterns:1) the volumes have diseases as to the change of the shape and the inconsistency of the intensity;2) Some volumes are enhanced by contrast agent while some not;3) The volume dimension and liver position in the volume substantially vary;4) The inter-slice resolution changes from0.33mm to8.0mm. The staircase effects will be in the low resolution of the volumes.To solve the problem of the liver segmentation, a novelty method is proposed, which is based Statistical Shape Models integrated a new intensity model. In this way, they incorporate a priori information extracted from the training set and can flexibly represent object shapes. Two central problems need to be resolved in designing these methods:(1) How to build a compact Statistical Shape Model to capture the shape variation more accurately.(2) How to select the image features and design the cost function to accurately and robustly match the deformable surface onto the object in an image.For the second point, active shape models (ASM) has supplied compact forms to capture shape variability, including both inter-and intra-patient variation. In this work, we use a point distribution model (PDM) similar to ASMs.The second point aims to characterize image features richly and to design the cost function to guide the model deformation. Numerous features have been proposed to build the appearance model. They can be divided into four categories:(a) Profile-based features. The features are vectors of photometric variables, such as intensity or derivatives, acquired along profiles normal to the object boundary. These schemes are effective in situations where objects have distinctive patterns along the normal of the object boundary, such as a high-contrast edge, and these patterns are consistent across the corresponding landmarks. However, for CT liver images, there is not always a consistent tissue mixture at corresponding locations, because of organ motion. Further, because this kind of feature only depends on a few samples along the boundary normal, it is sensitive to noise and prone to local minima.(b) Region-based features. Probability distribution functions for some photometric variables taken over the entire object or in local regions near boundary landmarks are used as features. This kind of approach could detect objects whose boundaries are difficult to be defined by gradients, perhaps lacking sharp edges. At same time, correspondence between the landmarks need not be precise (for the local region case) or even present (for the global region case). Also, it uses statistical knowledge about the distributions of some photometric variables of the object and background without imposing a specific parameterization, e.g. a Gaussian distribution usually used in the profile-based approaches.(c) Combination of several candidate features. For this kind of approach, feature selection strategies, such as support vector machines (SVMs) and K-nearest neighbor classifiers (kNNs), usually are adopted to get optimal combination features for specific segmentation task.(d) Other local image descriptors. For example, the scale invariant feature transform (SIFT) has been employed for chest radiography segmentation. Steerable features have been used to segment the liver and heart chambers, respectively.In this work, we want to develop a kind of features which can produce robust and accurate segmentation. As mentioned above, region-based features and profile-based features often suffer from a variety of limitations. Here we propose a method which endeavors to integrate these two kinds of features in an effort to form combination features, called region-edge combination features, which are robust to noise and poor initialization and produce accurate segmentation.In this paper, traditional region and edge features are first learned. Then according to the specific characteristic of liver, we propose the novelty region and edge features to integrate into the appearance models to segment the liver images based on the Statistical Shape Models. Afterward, the shortcomings using Statistical Shape Models and appearance models are intensive studied. Liver segmentation using a Statistical Shape Model based on learning local objective functions are proposed. In this algorithm, we firstly extract the appropriate features which are suitable for liver. Secondly in order to adjust these features automatically and intelligently, we build an ideal objective function through learning. We endeavor to segment the liver robustly and precisely:(1) Novelty region and edge featureEdge is essential in the medical image, which occupies the most image information. Image edge directly contact with the image content. For the CT of liver in abdomen, it exists widely in organs, and can effectively apply to image segmentation. Canny edge detector uses a multi-stage algorithm to detect a wide range of edges in images. Canny detector is the optimal edge detection algorithm, which have good detection, good localization and minimal response to noise.Canny detector only can detect the clear edge which is continuous. So we need some feature to distinguish the other kind edges. Gaussian Mixture Models is a useful technique in Mathematical Modeling or imaging processing Modeling. We use the Gaussian Mixture Models to fit the intensity distribution of the liver region to make a distinction between the liver and surrounding organs. According to the main gray information in liver area, we can easily calculate the probability of the estimate. In this paper, the estimate can be probability difference between points belong the profile (called edge feature), also can be the region around landmaks (called region feature)(2) The learned cost functionAfter extracting appropriate features, how to efficiently combine the region and edge feature is the key of this paper. There are two factors to determine the Shape models to fit the image data, termed objective function and search algorithm, respectively. Previous objective function is built on domain-dependent knowledge, with further local minima which fitting algorithms might incorrectly classify as the global minimum. And its iterative nature also makes it a time-consuming process of unpredictable duration. This last consequence is an especially direct cause of the complexity and sophistication of fitting algorithms:Determining the optimum of complex search spaces requires complex search algorithms. To overcome the weakness in the traditional algorithm, we learn the weight of the different features in different landmasks. In this paper, the appearance models in the segmentation locate quickly, and detail in the edge. The experiment results demonstrate the approach not only produces excellent segmentation accuracy, but also increases the robustness.