06-11-2012, 05:05 PM
CLASSIFICATION OF BRAIN TUMORS IN MR IMAGES
1CLASSIFICATION OF BRAIN.pdf (Size: 166.58 KB / Downloads: 78)
ABSTRACT
We study the problem of classifying brain tumors as benign or malignant using information from magnetic resonance (MR) imaging and magnetic resonance spectroscopy (MRS) to assist in clinical diagnosis. The proposed approach consists of several steps including segmentation, feature extraction, feature selection, and classification model construction. Using an automated segmentation technique based on fuzzy connectedness we accurately outline the tumor mass boundaries in the MR images so that further analysis concentrates on these regions of interest (ROIs). We then apply a concentric circle technique on the ROIs to extract features that are utilized by the classification algorithms. To remove redundant features, we perform feature selection where only those features with discriminatory information (among classes) are used in the model building process. The involvement of MRS features further improves the classification accuracy of the model. Experimental results demonstrate the effectiveness of the proposed approach in classifying brain tumors in MR images.
INTRODUCTION
Early detection and classification of brain tumors is very important in clinical practice. Many researchers have proposed different techniques for the classification of brain tumors based on different sources of information [1, 2, 3]. In this paper we propose a process for brain tumor classification, focusing on the analysis of Magnetic Resonance (MR) images and Magnetic Resonance Spectroscopy (MRS) data collected for patients with benign and malignant tumors. Our aim is to achieve a high accuracy in discriminating the two types of tumors through a combination of several techniques for image segmentation, feature extraction and classification. The proposed technique has the potential of assisting clinical diagnosis.
Necessary preprocessing steps prior to characterization and analysis of regions of interest (ROIs) are segmentation and registration. Image registration is used to determine whether two subjects have ROIs in the same location. However, in this work we do not take into account the location of the tumor in the classification model so we do not employ registration. Image segmentation is required to delineate the boundaries of the ROIs ensuring, in our case, that tumors are outlined and labeled consistently across subjects. Segmentation can be performed manually, automatically, or semi-automatically. The manual method is time consuming and its accuracy highly depends on the domain knowledge of the operator. Extensive work on medical image segmentation (either automatic or semiautomatic) has been done; methods can be divided into two broad groups: those that incorporate prior spatial information and those that are solely signal-intensity based; reviews can be found in [4, 5, 6, 7]. Specifically, various approaches have been proposed to deal with the task of segmenting brain tumors in MR images [8, 9]. The performance of these approaches usually depends on the accuracy of the spatial probabilistic information collected by domain experts. In previous work [10], we proposed an automatic segmentation algorithm that is based on the fuzzy connectedness concept [11]. The main idea is to assign to every pair of voxels, x, y, in the image, a real number between 0 and 1 indicating their connectedness. Starting with several seed points, all the voxels are automatically assigned to the structure to which they have the highest connectedness value. Utilizing the statistical information cumulated during the segmentation process, this method can provide satisfying results even in cases where the boundaries of the ROIs cannot be easily identified.
METHODOLOGY
There are four major steps in the proposed approach for brain tumor classification: (a) ROI segmentation: delineating the boundary of the tumor (ROI) in an MR image; (b) feature extraction: getting meaningful features of the ROI identified in the previous step; © feature selection: removing the redundant features; (d) classification: learning a classification model using the features.
Segmentation
We apply a segmentation algorithm (Fuzzy-Connectedness segmentation (FCS)) based on the concept of fuzzy connectedness to detect all the pixels on the MR images belonging to the tumor and discriminate from normal tissue. The main idea of fuzzy-connectedness is to assign to every pair of voxels, x, y, in the image, a real number between 0 and 1. This number is the fuzzy-connectedness of x to y and is utilized to denote the strength of the link between x and y. The segmentation algorithm begins with a given set of seed points, having at least one seed point for each type of tissue. The selection of seed-points is more effective using some prior knowledge about the domain; they can be either manually chosen by an expert or automatically selected by a computer algorithm. The automatic segmentation process utilizes the strength of fuzzy-connections between points to construct a structure. Each point is assigned to the structure having a neighboring point with the highest fuzzy-connectedness value. The strongest connection is detected first, and the process repeats until the weakest connection is calculated. In that sense, the grade of membership in an object of an arbitrary point is its fuzzy-connectedness to a pre-defined seed point. A sequence of points is called a chain, where its links are the ordered pairs of consecutive points in the sequence. The strength of a chain is the length of its weakest link. The segmented object is defined by the number of points that are connected through a chain to the selected seed point of the object. Here, we give several definitions related to Fuzzy-Connectedness segmentation (FCS). More details can be found in [10].
Feature Extraction
Before the classification model can be built, meaningful features of the ROIs delineated during the process of segmentation, need to be extracted and used as input in the model learning process. The feature extraction we used in our framework was introduced in [12] and has been successfully applied to medical images [14]. The basic idea of this technique is to construct a series of 1,…,k concentric circles with regular increments of their radius radiating from the ROI’s center of mass. For each radius increment, the fraction of the region occupied by the circle or the fraction of the circle occupied by the region is measured. In both cases, a feature vector consisting of k attributes is formed to represent the ROI. Based on the assumption that different classes of brain tumors tend to have different internal density characteristics and growth models, this intuitive method provides discriminative features. Figure 3 shows an example of the extraction of features using the concentric circle approach from a tumor.
Model Learning
There are many techniques available for constructing a classification model. Neural networks, decision trees and Bayesian networsk are among the most popular choices. While each technique has its strengths and constraints, in this work, we choose to utilize the decision tree model because of its nice interpretability.
A decision tree is a tree structure where leaf nodes represent classifications while branches show the related features leading to the classifications. Besides the ease of understanding, the decision tree model provides several other advantages over the other machine learning methods: it can deal with both nominal and categorical data; it is possible to
be validated with statistical tests; its training time is short and it is scalable to large datasets. All these characteristics are ideal for the analysis of medical data. When building a decision tree for a given dataset, all the features involved in the learning process are selected in the order of their discriminative power to add new branches to the tree until all the data samples under a node have the same class label. While a complex tree may fit perfectly the data samples used for training, it tends to provide a poor performance on new instances. This effect is called overfitting. In order to avoid overfitting, after a tree is built, usually a post-processing step of pruning is necessary to make the tree simpler and make the trained model more reliable and robust.
EXPERIMENTAL RESULTS
In our experiments, we used a dataset consisting of 50 MR brain images: 25 with benign tumors and 25 with malignant tumors. MRI was performed at 1.5 tesla and images consisted of unenhanced T1-weighted spin-echo [repition time (TR)/time to echo (TE)/number of excitations (NEX): 400/16/1], followed by fast spin-echo (FSE) T2-weighted (2800/90/1) sequences. The matrix was 256x128 for T1-weighted images and 256x160 for T2-weighted images. A 22cm field of view and a slice thickness of 5 mm with a 1.5 mm interslice gap were used with all imaging sequences. The axial T1-weighted sequence was then repeated following intravenous administration of 0.1 mmol/kilogram of intravenous gadolinium. The gadolinium-enhanced T1-weighted sequence was used for localization of MR spectroscopy voxel. The MRS was performed utilizing the multi-TE technique, consisting of short (TR/TE/excitations: 1150/35/128), intermediate (1250/144/128), and long (1350/288/128) TE. The size of the MRS localizer voxel was selected at our discretion, depending on the size of the tumor; its volume ranged from 4.0 – 8.0 cm3. The voxels were typically placed over the most homogeneous solid enhancing portion of the tumor. Automated spectral processing was performed using the commercially available GE SAGE software and consisted of zero filling and Fourier transformation of the free induction decay signal followed by zero and first order phasing and baseline correction of the frequency spectra.
DISCUSSION AND CONCLUSIONS
In this paper, we propose an approach for classifying brain tumors in MR images. The purpose is to develop tools for discriminating malignant tumors from benign ones assisting decision making in clinical diagnosis. The proposed approach utilizes a combination of different techniques and is composed of several steps including segmentation, feature extraction and model learning. We also demonstrate that the fusion of data from other sources (MRS data in our case) can further improve the system’s performance and provide encouraging results. Even though in this paper, we discuss mainly the application of the proposed approach to 2D MR images, it can be easily extended to 3D MR volumes, since both the segmentation and feature extraction techniques have been successfully applied to 3D volumes.