| Brain Tumors are the lesions of the local tissue dysplasia of the brain caused by tumor-causing factors. Brain tumors can oppress normal brain tissue, cause ischemic degeneration, result in dysfunctional in normal tissue. With the increase in tumor volume, hydrocephalus and cerebral edema can be caused, lead to increased intracranial pressure, affect the patient breathing and heartbeat central, result in deaths of the patients. In the mortality statistics of the top ten in tumor diseases, brain tumors came in ninth place in four deaths per100,000people, which is a really high mortality rate, so brain tumors can do great harm to the social and the human race.The imaging data of brain tumors has great significance as the basis for a series of operations throughout the clinical course of diagnosis and treatment of brain tumors. In today’s several major medical imaging equipments such as the X-Ray, CT, MRI and et al, MR is especially suitable for the diagnosis and treatment of brain tumors due to its no radiation injury and high soft tissue resolution. After the Nuclear Magnetic Resonance phenomenon was first found in1938, the nuclear magnetic resonance technology has gone through a rapid development and until1977the technology was first used for human body scan. After development for decades, the nuclear magnetic resonance devices are installed in medical institutions at all levels all over the world, MRI has become one of the most popular methods in medical imaging. With the MRI examination, the doctors can discover the lesion in time at the early stage of brain tumors, the MRI image data can provide the doctors with full and detailed image information for the diagnosis and is convenient for doctors to design the effective treatment options for patients. In the stage of the cancer treatment and the prognosis, MRI can-also guide the operation process, measure the size and the position of tumor, bring the doctors a lot of convenience, make the diagnosis and the treatment of brain tumor more convenient and more effective. But at nowadays, brain tumor segmentation is mostly carried out by experienced doctors manually, manual segmentation has a very high demand of knowledge and clinical experience of medical professionals, the segmentation process takes a long time and is easy to introduce human error, the segmentation results is different from person to person. Therefore, to meet the huge demand for clinical and research, MR image automatic segmentation of brain tumors has been the research focus in recent years.But algorithm research of the MR images automatic segmentation for brain tumors has been facing some difficulties, which include:limiting principle and technology of MR imaging itself; factors which affect the image acquisition process, the external environment, the device itself and the operating personnel and other factors; the complex structure of the brain tissue itself; huge differences between individuals, etc. In order to solve these problems, researchers from different aspects, proposed a number of related algorithms including:the segmentation method based on edge such as the Differential operator and the Deformation model; the segmentation method based on regional such as the Threshold method and the Fuzzy clustering method; other typical methods such as Atlas method and the method based on classification and so on.The. principle of the segmentation method based on the edge is that because of changes in the gray value of the pixel on the edge area is more severe, so the problem can be solved by detecting the edge of different local regions as the image segmentation. The Differential operator method is using a first-order or second-order countdown at step edge and roof edge of a image value different characteristics to detect the edge of the image, cause the method is simple and the edge detection result is good, this method is one of the essential methods to learn through image processing. However, due to edge detection operators are sensitive to noise, in terms of medical image, image segmentation simply rely on differential operator are often not satisfied with the results. The Deformation model method obtains the image segmentation result by combining internal forces and external forces to gradually approach the contours of the target, but the deformation modeling approach relies heavily on the initial contour and is easy to fall into local minima, and thus limited its applications.The principle of the segmentation method based on the region is that classify the pixel which having the same attribute to the image classification, apply image segmentation based on similarity of the set of pixels. The Threshold method is the easiest method of region-based segmentation, by designing different threshold selection criteria to meet the different needs of the division, but the medical image, especially the MR brain images cause it contains a lot of different organizations and each organization has its own gray value range, so the ranges are easy to overlap between these organizations, and thus the effect of threshold method is not satisfactory. The Fuzzy clustering method introduced fuzzy clustering algorithm theory, by assigning different degree of membership in the class of similarity while minimizing the similarity between classes and maximizing the similarity within the class, in order to determine the ownership of the object class to finalize segmentation. The disadvantage is that the fuzzy clustering algorithm performance depends on the initial cluster centers, the algorithm in terms of medical images easily into local optima, not global optimal solution.The Atlas method completes the segmentation by pre-prepared maps, then the target image is registrated on the pre-prepared maps to complete the split. As for medical image, because the anatomical structure knowledge of the image is determined, thus the atlas method can effectively use a priori knowledge to complete the split. But the result of the atlas method greatly affected by registration accuracy, and because of the complexity of the patients, how to build a adequately atlas which is big enough to meet the need is also a challenge facing researchers.The segmentation method based on classifier treats the segmentation problem as pixel classification problem by extracting the relevant features of pixels, the pixels will be classified into different categories, due to the classification algorithm can effectively use high-dimensional features of images, by improving the discernment of the pixel the image segmentation accuracy can be improved, so it is very popular in recent years as a segmentation method. In the segmentation method based on classifier, the ability to identify the strength of image features have a direct impact on the quality of the segmentation results, how to extract effective feature is the key issue the classification algorithm facing. The gray features is the most commonly used in medical image segmentation, because of the high degree of standardization of the medical image acquisition, allowing a direct comparison between the gray scale image. However, brain tumors in MR images, due to the presence of gray distribution of the tumor area and gray distribution of the background area overlap, and therefore rely on a simple gray feature is difficult to effectively dividing up the tumor area. Meanwhile, unable to provide spatial information is also inadequate for the gray feature. The texture features describes the image smooth, sparse and other features, such as GLCM feature can be used as an effective feature of pixels, but this feature is not ideal for a large local area. Also, because brain tumor has many types, and the texture information of different types tumors are quite different thus limited the use of the traditional texture feature on the segmentation of the brain tumor.For the traditional gray features has a low distinguish ability, lack of spatial information and the traditional texture features lack of generalization, the Bag of Words(BoW) model is introduced as the feature of the pixel. The BoW model has a wide range of applications in text processing and computer vision fields. The BoW model is be seen as a collection of visual words of images, all the visual words together constitute the visual word dictionary. Features of each pixel on the image are extracted and features can be quantizing, so that a pixel can be visually represented on the dictionary, and thus classifying the pixels to complete the image segmentation. Since the visual words are representative of similar characteristics and distinguish between different characteristics, tihus expressing pixels out of a dictionary as the feature can enhance the ability to distinguish features.The pretreatment for MR images is the first step of the image segmentation of brain tumor, due to inherent limitations of MR imaging principle and various factors affecting image acquisition process like the environment, the equipment, doctors and patients, etc., resulting in poor image quality generated MR images, therefore the MR images need fo go through pre-processing after which can be used for image segmentation. The pretreatment for MR images includes de-noising, non-uniformity correction, normalization and brain tissue extraction. In the pretreatment procedure, the brain tissue extracts is the most important part. Commonly used algorithms include BET brain extraction algorithm, BSE algorithms, graph cut method and the extraction method based on the classifier.The BET algorithm is a classic representative of the variable surface model, which combines the forces based on morphological operations and will expand the image edge contour points to brain tissue, resulting in better segmentation of brain tissue. The BSE method of the Boundary detection techniques is also a classic algorithm that combines anisotropic diffusion filter, Marr Hildreth edge detection algorithm and morphological operations together, in order to get a brain contour mask. The Graph cut method obtains brain tissue by removing the weak link between non-brain tissue with brain tissue. However, these methods are more or less facing the existence of over-segmentation or-under-segmentation problems, this paper presents a brain tissue extraction method based on the AdaBoost classifier. The AdaBoost classifier apply classify through continuous training, more attention is payed to the samples which are difficult to classify and enhance the ability of weak classifiers, and then combine a strong classifier through a cascade of weak classifiers to build the final classifier. By selecting effective gray features, texture features and context features in the data set consists of20TI-weighted MR image composition experiment was taken with the BET, BSE and GCUT three methods to compare, the brain tissue extraction method based on the AdaBoost classifier has improved to some extent on the accuracy of segmentation.After the pretreatment of the MR images, it is necessary to work on the image segmentation problem. The use of the BoW model for image segmentation is generally divided into three steps;1.feature extraction and description;2.build visual dictionary;3.classifiers.In the first phase, we select the image patch-based feature as the describer of the pixel, the patch-based feature can combine both pixel grayscale characteristics and texture features, but also can provide sufficient contextual information for the pixel, thus helping to improve the ability to distinguish features. In the process of building a visual dictionary, this paper uses K-means clustering method to the feature set of normal tissue and diseased tissue separately clustering, two separate dictionaries are generated from the lesion area and the background area Dictionaries are then combined to get the final joint visual dictionary, the joint dictionary can provide the relative spatial position information as a feature to improve the ability to express pixel from the dictionary. While classifying, we use a sliding window technique, the characteristics is represented by pixels of the neighborhood in a sliding window to further strengthen co-expression features of the target pixel, in order to increase the ability to distinguish. Finally, to test the MR brain tumor image segmentation method based on bag of words model through experiments, the logistic regression classifier was trained and tested on a data set consisting of160MR images and got a segmentation accuracy rate of90.42%.Finally, a summary was done on the end of this paper while doing some future work prospects. |
