14
Methods of MRI Brain Tumor Segmentation
Amit Verma
School of Computer Science, UPES, Dehradun, Uttarakhand, India
Abstract
Till today, doctors manually see the MR images of the tumor to predict the presence of a tumor or not based on their experience. Automatic detection of the presence of a tumor and identifying various other properties of brain tumors is still a challenge. Identifying a tumor automatically requires the segmentation of the tumor in MRIs. Researchers have developed various state-of-the-art methods to carry out the process of segmenting the brain tumor in MRI, and almost every method of segmentation can be broadly classified in two ways that are generative and descriptive models. In this chapter, the requirement and the importance of brain tumor segmentation in MRIs and the basic methods of doing brain tumor segmentation are discussed. Further, region-based and generative models with weighted aggregation methods for performing brain tumor segmentation using MR images are also discussed.
Keywords: MRI, brain, tumor, segmentation, generative, automatic
14.1 Introduction
Magnetic Resonance Image (MRI) is an imaging technique used for medical purposes for generating pictures of various organs and tissues of the body based on strong magnetic fields and radio waves. MRI has proven as a great advantage for the doctors to diagnose the patient and to separate the healthy and the infected tissues. MRI is the most useful way to diagnose a brain tumor; for diagnosis, the MRI of the patient is manually analyzed by the doctor to detect the presence of the infected area in the brain. Since the development of the MRI technique, much research has been done to date to automate or semi-automate the process of detecting the tumor in the brain with much higher accuracy. For automating the process of detecting the tumor in the brain, segmentation of MRI images plays a vital role.
Basically, image segmentation is the process of partitioning or segmenting the image into various segments so the analysis of the image becomes simpler. Further, the segmentation makes it simpler to analyze the area of interest in the image and leave the rest. The process of image segmentation is used in the detection of brain tumor by analyzing the MRI images of the brain, and this process is named as Brain Tumor Segmentation. It is the process of segmenting the MRI image based on region, pixel, etc., to separate the tumor from the part of healthy tissues in the brain and to automate the process of tumor detection with higher accuracy. Based on segmentation the affected region, shape, location, texture, size, and many other features of the tumor can be detected with less or no intervention of the human being. Till now, many researchers are working on the segmentation process of brain tumors to automate the way of tumor detection with much high accuracy. BTS can be done on various features of the tumor-like the temperature of the infected tissues remains higher as compared to the healthy tissues. Still, it is a much challenging area of research to automate the process of analyzing MRI for tumor detection. Therefore, various states of arts have been done for proving the method of segmenting the MRI image for brain tumor detection. Various popularly known methods are discussed further in this chapter.
14.2 Generative and Descriptive Models
Broadly, the methods of brain tumor segmentation can be divided into two categories specifically generative and discriminative [1]. Majorly, both the techniques are based on conditional probability. Generative methods use probabilistic models for segmenting the tumor based on size and shape. Generative models do not require any manually labeled training data for detecting the brain tumor; basically, it is used for unsupervised training data. Let us consider the various features of brain tissues and compare them with the prior knowledge, its group, or label various brain tissues according to these features. More specifically, each pixel of the MRI is compared with some features of the healthy or infected tissue to categories the pixel accordingly. Progressively the pixels of the image are grouped accordingly in some shaped boundaries (regions). If the discriminative model is used, then the pre-labeled data set is required for making a hypothesis model. Considering the MRI image of a brain tumor, each pixel of the image would be already labeled. On the basis of the pre-labeled training data set, it is broadly classified into some regions such that mean squared error with respect to the training data set would remain minimum [2–5]. Now, we will discuss the generative and descriptive models in a more general way.
Basically, the generative model is used for unsupervised or non-labeled training data sets for making a model or to train a machine. Based on some feature(s), data is divided into the region(s). In this way, the training data set is used to train the machine for making a hypothesis model predict the new data according to the region the new data would belong to. Taking an example for mathematical understanding about the generative model, let us consider a training data set in which each data (tuple) could be either in category A or B. That means it is still not known which data belong to which category more specifically data is unsupervised data (non-labeled). In this case, each data is checked on the basis of some feature X whether the data belong to category A or B. According to Equation (14.1) based on Bayes Theorem, P(X|Y = A) will be calculated for any data, say, D1. Here, data D1 is checked for feature X that how strong is the probability for data D1 to be in category A.
Similarly, P(X|Y = B) will also be calculated for data D1, and if P(X|Y = A) > P(X|Y = B), then the data D1 will be labeled as it belongs to category A. Progressively, all the data in the training data set are categorized as A or B and boundary accordingly as shown in Figure 14.1 specifying two regions named A and B.
Figure 14.1 Training data set is categorized in A or B based on some feature X.
Now, these specified regions are used for making a hypothesis of any new data say D2 as shown in Figure 14.2 with a green colored dot. Please note that green color is not specified as a property or feature of data, it is just for making difference with other given data. The new data D2 is unknown and the hypothesis is to be made on the basis already learned machine. As the data, D2 is falling in the red colored boundary region so the data D2 is more likely to be in category B.
The discriminative model basically requires pre-labeled data set to train the machine according to the training data set for making a hypothesis model [6–9]. For example, we are having a training data set in which each data is already labeled with A or B as shown in Figure 14.3. For ease, red colored dots are representing data with label A and blue colored dots are representing data with label B.
Now, a decision boundary is drawn to broadly separate the data in such a way that the mean squared error would be minimum as shown in Figure 14.4. By the term error, we mean the decision boundary should divide the training data in such a way that the probability of data labeled A to be in a region belonging to data labeled B should be minimum. According to the data, linear or logistic regression with gradient descent can be used for making a decision boundary with a minimized error.
Now, as the hypothesis model is built according to the training data set using the discriminative model, it can be used for making a hypothesis on new data arrival. Let us say new data D2 (which is not pre-labeled) is shown with the green colored dot in Figure 14.5.
According to the hypothesis model prepared according to training the data set, data D2 is lying in a region belonging to category B. So, it would be labeled as B.
Figure 14.2 Showing arrival of new data (green colored dot) for making hypothesis.
Figure 14.3 Pre-labeled data of training data set.
Figure 14.4 Decision boundary separating data.
Figure 14.5 New data D2 is shown with green colored dot.
Based on the generative and discriminative approach, multiple states of arts were proposed for segmenting the brain tumor (abnormality) in MR images. Many researchers contributed to this field for automating the process of brain tumor detection in the MRIs. The most popular ways for doing segmentation of the MRI are as follows.
14.2.1 Region-Based Segmentation
In this approach, pixels of MRIs are group together based on some characteristics. These groups of similar pixels are considered as regions [10–14]. This approach can be used for partitioning the brain lesion structure into separate regions according to the characteristics of the pixels in MRIs. Kim et al. [15] used seeded region growing-based approach for segmentation in which the seed pixel(s) are chosen randomly. The surrounding pixels are compared with the seed pixels and categorize on the basis of homogeneity with the seed pixels. In the proposed work, region-based segmentation has been done, seeds are automatically selected, and the spatial domain is divided into small clusters on the basis of thresholding. The clusters are recursively divided into smaller regions to minimize the error and increase a better hypothesis for predicting the brain tumor in the MRIs.
14.2.2 Generative Model With Weighted Aggregation
Corso et al. [16] used multilevel segmentation for segmenting edema (swelling due to fluid leakage) and the tumor. The work has been done specifically on glioblastoma tumor of the brain. Glioblastoma initiates from glial cells, which are responsible for maintaining the good health of the brain. It is a fast-growing tumor. The tumor is majorly comprised of a dead part which is known as the necrotic and active part. Adema is also the part of the brain which is basically the swelling caused by the leakage of the plasma. Figure 14.6 shows the region of various parts of growing glioblastoma tumor on the basis of MRI (T1 and T2 weighted).
Figure 14.6 The different parts of brain tumor on the basis of T1 and T2 MRIs [16].
Figure 14.7 Flowchart representing steps in SWA [19].
Figure 14.8 Coarsening of the pixel representation of the portion of MR image.
T1 and T2 are the types of MR images which differ on the basis of contrast and brightness to ease the manual prediction of brain tumor by the doctor. Glioblastoma is one of the common brain tumors which is nearly 40% [17] of all brain tumors patients of almost all ages and having a very short postoperative time of survival nearly 8 months [18]. Basically, in this method, the MR image is considered as a graph of pixels (vertex).
Each voxel in the MR image is considered as a node in a graph G connected with six neighbors. Now, the segmentation according to the weighted aggregation algorithm is used to coarsening the graph recursively until some portion of interest is encountered as shown in Figure 14.7. The flowchart of the process of segmentation on the basis of weight aggregation is shown in Figure 14.8.
The process of SWA (Segmentation by Weight Aggregation) is divided into two broad processes that are bottom-up in which the construction of graph pyramids is done. Followed by top-down, in which with the help of pyramids region of salient (interested) nodes are identified and defined in the boundary.
As we can see, image is just a grayscale representation of the portion of the MR image, and then in image 2, gray pixels are represented with colors. Then, iterative coarsening is done for getting the interesting region.
14.3 Conclusion
In this chapter, we discuss the importance of automatic analysis of MRIs for the detection of a brain tumor which could be achieved with the process of brain tumor segmentation. Many researchers have proposed various methods of segmentation, and almost every method of brain tumor segmentation can be categorized into two broad categories that are generative and discriminative models. These are the two basic approaches which are used in various state-of-the art methods for segmenting the brain tumor on MR images to automate the process of detecting tumor for the doctors. The generative model basically works on the data which are not pre-labeled, whereas discriminative model requires the pre-labeled data for the hypothesis.
References
1. Menze, B.H., Jakab, A., Bauer, S., Kalpathy-Cramer, J., Farahani, K., Kirby, J., Burren, Y., Porz, N., Slotboom, J., Wiest, R. et al., The multimodal brain tumor image segmentation benchmark (brats). IEEE Trans. Med. Imaging, 34, 10, 1993–2024, 2015.
2. Agn, M., Puonti, O., Law, I., af Rosenschöld, P., van Leemput, K., Brain tumor segmentation by a generative model with a prior on tumor shape, in: Proceeding of the multimodal brain tumor image segmentation challenge, pp. 1–4, 2015.
3. Corso, J.J., Sharon, E., Dube, S., El-Saden, S., Sinha, U., Yuille, A., Efficient multilevel brain tumor segmentation with integrated bayesian model classification. IEEE Trans. Med. Imaging, 27, 5, 629–640, 2008.
4. Menze, B.H., Van Leemput, K., Lashkari, D., Weber, M.A., Ayache, N., Golland, P., A generative model for brain tumor segmentation in multimodal images, in: International conference on medical image computing and computer-assisted intervention, Springer, pp. 151–159, 2010.
5. Prastawa, M., Bullitt, E., Ho, S., Gerig, G., A brain tumor segmentation framework based on outlier detection. Med. Image Anal., 8, 3, 275–283, 2004.
6. Bauer, S., Nolte, L.P., Reyes, M., Fully automatic segmentation of brain tumor images using support vector machine classification in combination with hierarchical conditional random field regularization, in: International conference on medical image computing and computer-assisted intervention, Springer, pp. 354–361, 2011.
7. Hamamci, A., Kucuk, N., Karaman, K., Engin, K., Unal, G., Tumor-cut: segmentation of brain tumors on contrast enhanced mr images for radiosurgery applications. IEEE Trans. Med. Imaging, 31, 3, 790–804, 2012.
8. Lun, T. and Hsu, W, Brain tumor segmentation using deep convolutional neural network, in: Proceedings of BRATS-MICCAI, 2016.
9. Pratondo, A., Chui, C.K., Ong, S.H., Integrating machine learning with region-based active contourmodels in medical image segmentation. J. Vis. Commun. Image Represent., 43, 1–9, 2017.
10. Chang, Y.-L. and Li, X., Adaptive Image Region-Growing. IEEE Trans. Image Process., 3, 6, 302, 1994.
11. Rays, S.P. and Udupa, J.K., Shape-based interpolation of multidimensional objects. IEEE Trans. Med. Imaging, 9, 1, 32–42, 1990.
12. Gong, I., Kulikowski, C., Mezrich, R., Valley-enhanced histogram computation for MR image segmentation. Presented at the Ann. Meeting of the American Society of Neuroradiology, Mar. 1991.
13. Adams, R. and Bischof, L., Seeded region growing. IEEE Trans. PAMI, 16, 6, 641–647, June 1994.
14. Reutter, B.W., Klein, G.J., Huesman, R.H., Automated 3-D Segmentation of Respiratory-Gated PET Transmission Images. IEEE Trans. Nucl. Sci., 44, 302, 6, 1997.
15. Kim, J., Feng, D.D., Cai, T.W., Eberl, S., Automatic 3d temporal kinetics segmentation of dynamic emission tomography image using adaptive region growing cluster analysis, in: Nuclear science symposium conference record, vol. 3, IEEE, pp. 1580–1583, 2002.
16. Corso, J.J., Sharon, E., Dube, S., El-Saden, S., Sinha, U., Yuille, A., “Efficient multilevel brain tumor segmentation with integrated bayesian model classification”. IEEE Trans. Med. Imaging, 27, 5, 629–640, 2008.
17. Smirniotopoulos, J.G., The newWHO classification of brain tumors. Neuroimaging Clin. N. Am., 9, 4, 595–613, Nov. 1999.
18. Patel, M.R. and Tse, V., Diagnosis and staging of brain tumors. Semin. Roentgenol., 39, 3, 347–360, 2004.
19. Neilson, R. and B.N.M., Image segmentation by weighted aggregation with gradient orientation histograms, in: Southern African Telecommunication Networks and Applications Conference, SATNAC, 2007.
- Email: [email protected]