| Brain tumor is divided into malignant tumor and benign tumor, each accounting for roughly 50 per cent of the total. Glioma, metastases, etc. are the most common malignant tumor, and meningiomas, pituitary adenoma, schwannoma, etc. are the most common benign tumor. Tumor can also be distinguished primary from secondary according to the cause of it. Neuroepithelial tumor, also known as glioma accounts for approximately 40% of brain tumors, which is the highest disease incidence. The symptoms of brain tumors are various, and include nausea, headache, vomiting, fatigue, probably mental disorders, and even decreased vision or even blindness. Although the treatment means in brain tumors has been improved rapidly in recent years, but the high death rate is frequently observed.Recently, MRI (Magnetic Resonance Imaging) and CT (Computed tomography) have been widely applied in tumor imaging diagnosis. MRI has practical value in the medical image analysis, such as the merits of high resolution as well as high quality, the super-resolution ability to human soft tissue, the accurate description of the brain anatomy. Hence, it becomes one of hotspots for study in the field of medical image analysis. Besides, it is also used as an important auxiliary means in the diagnosis of brain tumor. In addition, it’s of great significance to brain tumor treatment and the surgical guide. In order to make full use of anatomical information in the brain image, provide quantitative and intuitive reference for clinical diagnosis, the key point is to assure the validity and accuracy of tumor image segmentation. Manual segmentation is the most common approach mainly depends on clinical experts, with professional constraints, for instance that various experts can give different segmentation schemes to the same tumor image, and that although the same expert can’t obtains the same segmentation effect for the same patient at different time. Complicated procedure and poor repeatability are both disadvantages in manual segmentation. How to segment automatically rely on computers has been long considered by people. However, due to the changing shape of brain tumor, the complicated structure and uneven gray, the existence of edema in the tumor boundary, each mode emphasizes different information, for example:FLAIR modal emphasizes the grayscale difference between normal tissue and the tumor, while TIC modal emphasizes the different features in boundary-texture area. There also exists great difference of image information in the same modal to the various patients. How to excellently segment brain MRI image by the way of computer is still a hotspot worthy of further research.In recent years, the segmentation of brain MR images has been extensively researched at home and abroad. The proposed representative methods can be classified into the following groups:region segmentation; threshold segmentation; pixels-classifier segmentation; model-driven segmentation...etc. This paper analyzes in detail these segmentation methods:The threshold segmentation is relatively simple, which is usually used in the early work of brain tumor segmentation; the region segmentation, mainly including region growing technique, watershed algorithm and so on, is simple, easy to implement, and excellent to connectively the segment the image with clear texture as well as non-complicated gray levels, however, it’s easily trapped into over-segmentation problem with the influence of noise or uneven local gray; model-driven segmentation, such as parametric deformable models and level set and so on, has been widely drawn close attention by its adaptation to the variability of the anatomical structures, however, it’s not suitable for complex boundary as to it usually demands a high computational complexity, pixel-classifier segmentation, such as fuzzy C means, artificial neural network, Markov random fields, support vector machine and so on, the popular method in image segmentation especially for brain tumor MR images, can make full use of the gray information of neighborhood pixels and local texture features, and realize the nonlinear discrimination and classification, however, it also exists some drawbacks like the change of the parameters significantly influences the segmentation effects, the high time complexity and space complexity of the algorithm. The SVM method based on statistical learning theory, proposed By Vapnik, has made comprehensive consideration for expected risk and empirical risk. Besides, it’s the most widely used pixel classifier for brain tumor MRI image segmentation with snail number of sample, and non-linearization, and the capacity to overcome dimensional disaster, and high generalization performance. In this paper, emphasis is put on the research of the classification model based on SVM for brain tumor image segmentation.It has long been a hot research task for segmentation of MRI brain tumor image based on pixels, especially the aspects of feature extracting, feature selecting, classifier designing, and it proposes that better feature extraction can make the classifier work easier. Feature extraction can be mainly classified into three groups: the statistical methods, the model methods, the signal processing methods. Each of these methods has its advantages or disadvantages. The statistical method is simple, easy to implement, and some superiority to small images. However, it makes inadequate use of global information, and it’s out of touch with human’s vision model. The model method is very flexible with consideration to both the randomness of local texture and overall regularity. However, the model coefficient is difficult to solve, and the procession of parameter adjustment is complex. The signal processing method is good at capturing the texture details and expressing it in space and frequency domain. While the wavelet tend to lose too many high frequency information, and it’s unsuitable to extract the irregular texture. MRI images with various modal can provide different textures-boundary information. Due to individual differences, the information from the same modal for various patients is also quite different. Therefore, no one feature extracting method is suitable for the segmentation of all MRI brain tumor images.Convolution neural network, put forward by Yam L. C., is a kind of depth supervised learning methods, that has achieved great success in many fields, such as image recognition, speech recognition, natural language processing and so on. CNNs can directly extract some features helpful to classification from the raw input data by sub-sampling under cyclic convolution and obtaining the convolution weights in the way of supervised-training. The useful features mainly include texture, shape, structure...etc. Considering that CNNs requires multiple convolution as well as sub-sampling, and that the input object is usually an image with high neighborhood size value, it’s not suitable for feature extracting of various MRI brain tumor image with rich texture.This paper will extend CNNs to the 3D model, combined MRI multi-modal information which is also the raw data to the 3D model, and the 3D model shall be used to extract features benefiting to segmentation, and classify these features with the help of SVM classifier. Firstly, this paper introduces CNNs as the feature extracting method for MRI brain tumor image, and the supervised method automatically classify features according to the differences among various patients and overcome the shortcoming of non-supervised method only confine to some special features; Secondly, this paper improves the 2D-CNNs to multi-modal 3D-CNNs, which can obtain texture from the three directions at the same time, thus to solve problem that the raw input of 2D-CNNs requires large neighborhood data, and better to extract the difference information at the same time, and to enable a wider range of image segmentation for MRI brain tumor.The segmentation method based on SVM pixels-classifier, mainly for extracting and selecting image feature, directly applies the based-RBF SVM in brain tumor image segmenting. Although the based-RBF SVM has a good effect on tumor segmentation, it remains the need for further deep study on tumor MR images in which there exists the brain edema surround the tumor and the boundary of tumor is blur, in order to achieve the ultimate goal for the clinical application.A novel SVM method based on hybrid kernel function used for tumor MR image segmentation is presented in this paper, for making up the shortage of mononuclear. Hybrid kernel function, put forward by Smits et al. for the first time in 2002, shows a better performance in the learning ability and generalization ability of SVM than the performance of the single kernel function by combining local kernel function and the global kernel function, and is widely used in face recognition and palm recognition. AS the shape, size and location of brain tumor is so changeable and the boundary of malignant tumor is always blurred, the classical mixed kernel function can’t greatly improve segmentation effects relative to the single kernel function, but it increases the parameter number, leading to the poor use of the classical mixed kernel function in the brain tumor segmentation.This paper presents an improved mixed kernel function, which combines the RBF local kernel function and Sigmoid global kernel function both evolved from the neural network and expand the weight coefficient of the mixed kernel function. Firstly, the expansion of weight coefficient increases the distance of sample points in the new mapping space, weakens the influence of penalty factor C, fixing it in the process of parameter optimization and reducing the number of parameters to be optimized; Secondly, the expansion of weight coefficient changes the correction factors in SMO algorithm, affecting the selection of support vector, obtaining better classification interval, and eventually improving the segmentation accuracy of brain tumor.The key techniques presented in this paper mainly include the following parts:(1) the multimodal 3D-CNN feature extraction method; (2) the adaptive weighted hybrid kernel function method.(1) The multimodal 3D-CNN feature extraction method. Considering that the unsupervised feature extraction method can’t adapt to the differences of brain image in MRI brain tumor segmentation, this paper proposes a multimodal 3D-CNN feature extraction method for tumor segmentation in MRI brain tumor image. The 3D-CNN method, by grouping the 2D multimodal MRI images into the 3D original feature, is more conducive to extract the difference information between the modals, remove the redundancy interference information between modals, narrow the size of original feature neighborhood at the same time, adapt to the different changing size of tumor in various image layers of the same patient, and further improve the segmentation accuracy of MRI brain tumor. Experimental results show that the proposed method can not only adapt to the diversity and dynamics among various modal of different patients, but also improve the accuracy of brain tumors segmentation.(2) The adaptive weighted hybrid kernel function method. Firstly, this paper constructs the adaptive weighted hybrid kernel function, adjusting the distance between the space sample points adaptively, changing the modifying factor in the process of sequential minimal optimization, weakening the influence of the punishment factor, changing the value of the Lagrange multiplier, optimizing the selection of support vector, getting a better interface of classification, and ultimately improving the classification ability of SVM. Secondly, this paper first applies a hybrid kernel function SVM brain tumor image segmentation as well as guarantees the learning ability and generalization ability of it, which is an adaptive weighting method composed by the optimal local kernel function (radial basis kernel function) and a global kernel function (Sigmoid kernel function) adaptive weighted combination. Experimental results show that the proposed method obtains more efficient and accurate segmentation effects of brain tumors.In this paper, the established model is tested with image data of 30 patients with glioma from the online image library MICCAI2012.Besides, this paper evaluates the effects of the proposed model by some quantitative analysis compared with the segmentation result carried by the clinical experts. The assessment results show that the proposed algorithm can arrive at accuracy rate of 91.29% which is very close to the ideal value in the test, and demonstrates the feasibility and effectiveness of the proposed method. |
PDF Full Text Request