| Brain tumor refers to the cancerous substance grown in the cranial cavity, including primary tumor caused by the lesions of brain parenchyma such as brain, cranial meanings, blood vessels, nerves or brain accessories, and secondary tumor shifting from other parts of the body. The incidence of brain tumor has a high rate in patient population. It can happen at any ages, most commonly seen from20to50years old. And once the tumor occupying a certain space in the cranial cavity, regardless of its nature is benign or malignant, that will be bound to constrict the brain tissue, which may cause high intracranial pressure, central nerve damage and threaten patients’ life.In recent years, result from the increase of environmental pollution, too much stress in people’s life, the incidence of brain tumors is on the rise. It shows that the incidence of brain tumors account for1.4%of systemic cancers, and death rate is more than2.4%according to the latest cancer epidemiology investigation and research results. Brain tumor which is still a kind of disease with a higher morbidity and mortality has become one of the most life-threatening diseases. Glioma is the most common primary malignant brain tumor, especially in the central nervous system tumors. Glioma, which is infiltrating and diffuse, has variable shapes and its boundary is difficult to apart from the surrounding tissue. On the basis of overall neurologic examination, as well as the use of appropriate auxiliary imaging examination method, can doctors diagnose and classify tumor accurately and effectively.Brain tumor develops slowly, gradually, and the course of different period. Its clinical symptoms include such as headache, vomiting and papilledema caused by high intracranial pressure. Tumors at different parts of the body can also be characterized by focal symptoms, signs such as hemiplegia, aphasia, paralytic symptom caused by disorder of spirit and consciousness, and irritation symptoms such as muscle twitching and epilepsy. Clinical doctors rely on medical history and reliable data, based on the neural anatomy, physiology and the law of development of various diseases to diagnose comprehensively, analyze objectively, and then select further auxiliary examination tools, to research the location, size, nature, blood supply of the tumor and the surrounding tissue may be involved comprehensively. Finally, clinical doctors detemine more accurate position and qualitative differential diagnosis to the tumor.Surgery combined with chemotherapy, radiotherapy or biological treatment is the best therapy method for brain tumor. The precondition of surgical treatment is a reasonable and effective means of detecting and monitoring of brain tumors. The surgery is just for the hope of removing the tumor as much as possible. Because malignant tumor is infiltrating to diffuse growth and its boundary is difficult to apart from the surrounding tissue, the segmentation results between individuals performing the same brain tumor or within individuals repeating at different times are different. So using computer image processing technology to identify tumor effectively is the inevitable developing trend of clinical application.Brain Imaging techniques commonly used in clinical include Computed tomography (CT), Magnetic Resonance Imaging (Magnetic Resonance Imaging, MRI), etc. The development and wide application of these imaging techniques have greatly increased the detection rate of brain tumor, and brought a lot of help to doctors and patients.Medical imaging equipments provide variety images for computer processing, such as MRI (magnetic resonance imaging, MRI). MRI is a kind of important anatomical imaging diagnostic techniques, with high resolution of soft tissue. As an imaging method which is noninvasive, nonradiative, multi-parameters, sensitive to the organization form and pathological changes, MRI has now become an important tool for diagnosing brain tumors. Different modalities focuses on the performance of different information, for example, FLAIR focuses on the obvious grayscale difference between tumor and normal tissue; TIC focuses on texture feature difference. Single modality MRI image is difficult to fully provide identify information. At the same time, the variable shapes and fuzzy boundaries with surrounding edema area make it difficult to accurately segment the tumor. Experienced doctors commonly combined with multimodal MRI images, using CAD software, manually delineate the tumor area layer by layer. That results in strong subjectivity and poor repeatability. Thus using the machine to recognize tumor is an inevitable trend in the development of clinical application.Common tumor segmentation methods mainly include fuzzy clustering (FCM) method based on image gray level information, the level set method, the neural network method, AdaBoost iteration method and Support Vector Machine (SVM) method. FCM method is easy and fast, but due to the complexity of medical image information and indistinct edge, the selection of the seed points had a great influence on the clustering results, and it is difficult for FCM method to use the image’s spatial information. For level set method, the biggest advantage is that the method processes the change of the curve’s topology naturally and stably. It is sensitive to initialization, so the segmentation result is easy to fall into local extremum. Neural network learning ability is powerful, but it requires proper training samples. The training process is prone to local optimality, which makes its generalization performance poor, especially in the case of small sample. AdaBoost algorithm’s segmentation accuracy is high, no over fitting phenomenon, but when use this method for multimodal MRI segmentation image, it requires more training samples, longer training time. The above methods have their own advantages, but there is a certain gap for the clinical application.The SVM method based on statistical learning theory shows many advantages. SVM can still obtain good generalization ability in condition of the relatively small samples, and the higher dimension of feature. At the same time, SVM can deal with non-linear data effectively by means of introducing kernel function. Single RBF kernel function may achieve good application in some documents for multi-modality MRI images. But RBF kernel function often use the local information of the sample, it can obtain good segmentation results only for tissues in the input images are distinct. But for gliomas, with variable shapes and blur boundaries, a single RBF kernel function’s performance has certain limitations. The new kernel function combines the nature of local and global kernel function can overcome these problems. It has been widely confirmed in Face recognition and palmprint recognition that the optimal combination parameters of mixed kernel function is better than a single kernel function. The structure of face and palmprint is simple, relatively fixed, but the images of the brain tumor tissues, especially the low grade gliomas, with diffuse infiltrative growth, signal strength between normal tissue, variable tumor shape, location, size, blur boundaries, complex texture structure and so on. How to make full use of image’s multimodal information, and to find the optimal combination parameters of SVM model are the difficulties of the SVM method applied to tumor segmentation.This paper proposes a multimodal MRI image segmentation method based on parameters optimization of SVM model to improve the existing multimodal MRI brain tumor segmentation method, making full use of the MRI image of multimodal information, at the same time combining with SVM method. This method firstly analyzes different multimodality images, and finds the prominent support vectors’ differences in tumor tissue and normal tissue information. Secondly optimizes kernel function of support vector machine classifier. SVM classifier skillfully solve the problem of nonlinear data by introducing kernel function.Kernel function includes local and global kernel function. Different types of kernel function focus on different information, so there are differences in performance. Combine the advantage of local RBF kernel function and global Sigmoid kernel function. Then, train SVM classifier based on the optimal mixed kernel function for a single modality image, which only needs a small training set, and the performance is better than a single RBF kernel function. The classification results became different because there was some different information in the selected support vectors of mono-modality image. By means of modifying weight values of the error data points, chose the best weight values of the sub-classifier. The method would get a weighed combination SVM classifier of multi-modalities, which would enhance the segmentation performance, and used it to MRI image segmentation. Experiments show good generalization performance, strong feasibility and practicabilit. The method can realize high accurate of brain tumors segmentation.In this approach, there introduce and put forward the key techniques include:(1) image denoising algorithm;(2) the mixed method of kernel functions;(3) combination of sub-classifiers.(1) Noise in MRI images can reduce image quality, affect the image visual observation, and some details are often submerged in noise. So the machine usually gets less useful information, even gets error messages, which affects the accuracy of image processing algorithm. Therefore we present a novel pre-processing method to solve such problem. For additive noise in MRI images, the difference between homogenous areas is only related to the noise. We introduce the improved increment-dimensional bilateral filtering algorithm, on the premise of guaranteeing the filter performance to make bilateral filtering realize quickly. That can effectively prevent the important information during the denoising process, and speed up the implementation of the algorithm.(2) Optimize combination coefficients of mixed kernel function. By adaptively adjusting each sample points’distance in new mapping space, weakens the classifier punishment factor’s influence on the classification results, thus makes the punishment factor can be fixed during optimization process, without affecting the segmentation precision. At the same time, the Optimization of weight coefficients can change SMO(Sequential Minimal Optimization) algorithm’s correction factors, which affect the selection of support vectors, in order to get better margin, then greatly improve the segmentation accuracy of brain tumors in the end.(3) With the aid of weighted combination of multi-modality sub-classifiers, the method makes full use of various information of different modality. SVM is a learning method with its advantage, but medical images have complex information, so the optimal classifier trained by finite samples cannot satisfy the requirement of high precision. Therefore using the theory of integration, construct new classifier by combining more diversity, better performance and independent sub-classifiers, to improve the generalization performance of the final classifier. Every modality has a corresponding set of new samples with multimodal information. Using these samples train sub-classifier respectively. Then combine sub-classifiers and optimize the segmentation results.The dataset for testing and validation consists of34cases of glioma patient images which were provided by the online image library MICCAI2012. The results were reviewed by clinicians and quantitative analyze. The average classification accuracy is available for92.50%, and the segmentation results are closed to the ground truth. Thus verify the feasibility and practicability of this method. |
PDF Full Text Request