Two fully-unsupervised methods for MR brain image segmentation using SOM-based strategies

Abstract

Image segmentation consists in partitioning an image into different regions. MRI image segmentation is especially interesting, since an accurate segmentation of the different brain tissues provides a way to identify many brain disorders such as dementia, schizophrenia or even the Alzheimer's disease. A large variety of image segmentation approaches have been implemented before. Nevertheless, most of them use a priori knowledge about the voxel classification, which prevents figuring out other tissue classes different from the classes the system was trained for. This paper presents two unsupervised approaches for brain image segmentation. The first one is based on the use of relevant information extracted from the whole volume histogram which is processed by using self-organizing maps (SOM). This approach is faster and computationally more efficient than previously reported methods. The second method proposed consists of four stages including MRI brain image acquisition, first and second order feature extraction using overlapping windows, evolutionary computing-based feature selection and finally, map units are grouped by means of a novel SOM clustering algorithm. While the first method is a fast procedure for the segmentation of the whole volume and provides a way to model tissue classes, the second approach is a more robust scheme under noisy or bad intensity normalization conditions that provides better results using high resolution images, outperforming the results provided by other algorithms in the state-of-the-art, in terms of the average overlap metric. The proposed algorithms have been successfully evaluated using the IBSR and IBSR 2.0 databases, as well as high-resolution MR images from the Nuclear Medicine Department of the “Virgen de las Nieves” Hospital, Granada, Spain (VNH), providing in any case good segmentation results.

Introduction

Many current problems in image-guided surgery, therapy evaluation and diagnostic tools strongly benefit from the improvement on the medical imaging systems at reduced cost [1]. In this way, magnetic resonance imaging (MRI) has been widely used due to its excellent spatial resolution, tissue contrast and non-invasive character. Moreover, modern medical imaging systems [2] usually provide a vast amount of images to be analyzed. The study and evaluation of these images are usually developed through visual ratings performed by experts and other subjective procedures which are time-consuming and prone to error.

Generally, MRI images are qualitatively analyzed by experts based on their own experience and skills, but it is always limited by the human vision system which it is not able to distinguish among more than several tens of gray levels. However, as current MRI systems can provide images up to 65,535 gray levels, there is much more information contained in a MRI than the human vision is able to extract. This way, computer aided tools (CAD) play an important role for analyzing high resolution and high bit-depth MRI images, as they provide an important source of information for radiologists when diagnosing a disease or looking for a specific anomaly.

Segmentation of MR images consists in identifying the neuro-anatomical structures within medical images or “splitting an image into its constituent parts” [31]. Brain segmentation techniques, as a part of CAD systems, can be used to characterize neurological diseases, such as dementia, multiple sclerosis, schizophrenia and even the Alzheimer's disease (AD) [3]. In the case of AD, there is no a well-known cause and it is very difficult to diagnose. With the improvements of MR imaging systems, the image processing techniques as well as the discovery of new biomarkers, neurological disorders such as AD are expected to be diagnosed even before the manifestation of any cognitive symptoms [57], [58]. Thus, segmentation of brain MRI enables finding common patterns in AD patients such as hippocampal volume or cortical gray matter density reduction. Furthermore, these techniques could help to find other causes of brain disorders or anomalies. In fact, the segmentation algorithms presented in this paper are part of a larger study performed by the authors on the use of tissue distribution for the early diagnosis of AD [57], [58].

The development of effective tools for grouping and recognizing different anatomical tissues structures and fluids, is a field of growing interest with the improvement of the medical imaging systems. These tools are usually trained to recognize the three basic tissue classes found on a brain MR image: white matter (WM), gray matter (GM) and cerebrospinal fluid (CSF). All of the non-recognized tissues or fluids are classified as suspect to be pathological. In the same way the human expert has to learn to recognize different regions on the MR image, the segmentation algorithms have to be trained.

There are a wide range of brain segmentation techniques. They can be classified into manual, semiautomatic and automatic techniques. Manual techniques are the most common and have been used for years. They require a human expert to select the voxels belonging to a specific object individually. In semiautomatic segmentation, the human expert is usually aided by image processing techniques.

The most common image processing techniques used for semiautomatic segmentation are histogram-based techniques [8], [9], [16], [22], statistical classifiers, fuzzy classifiers, support vector machine (SVM) classifiers, and neural network-based classifiers. Histogram-based techniques are based on thresholding, which consists in determining the thresholds of the voxel intensity values in order to separate the voxels belonging to each class. This requires a previous training process of the system using expert segmentation images. Segmentation techniques based on histogram thresholding use the fact that the peaks on histogram can belong to a specific tissue [35]. Thus, the problem is reduced to classifying and modeling the peaks and valleys on the histogram. There are other histogram-based segmentation techniques which also take into account the relative position of the peaks or other statistics calculated from the histogram [36], [37]. Nevertheless, the histogram thresholding segmentation approaches usually do not take into account the spatial information contained on a MR image. On the other hand, spatial information is essential since anatomical regions of the brain [18] are more likely to accommodate a given tissue. As a result, different MR images could have similar histograms and then, similar thresholds.

Other segmentation approaches are based on contour detection techniques [10], [11], using the boundaries among different tissues for segmentation. Edge detection algorithms such as Sobel, Prewitt, Laplacian or Canny filters [31] select the border voxels among different objects. These filters generally perform the preprocessing for active contour algorithms [12].

In region-based techniques [13] once a voxel is marked, the algorithm starts to add more voxels surrounding it, preserving some properties such as homogeneity or intensity level. Other algorithms search for voxels belonging to the initial class following a specific geometrical model [38].

Statistical classifiers use some previous learned rules to perform the grouping. These are called clustering techniques which classify voxels in an unsupervised manner, since they group similar voxels into the same class. Thus, a similarity criterion has to be established or learned in order to determine whether a voxel belongs to a given class. Then the classifier [4] will generate different classes which contains group of voxels with the same properties. Some of the statistical classifiers are based on the expectation-maximization algorithms (EM) [1], [14], [15], maximum likelihood (ML) estimation [16] or Markov random fields [13], [17]. K-means and its variants such as Fuzzy k-means are widely used as they avoid abrupt transitions in the classification process [19].

Support vector classifiers [40] are a new type of classifiers based on statistical learning theory which have been successfully applied to image segmentation [20] due to its generalization ability. Other segmentation techniques are based on artificial neural network classifiers [21], [22], [23], [24], [25], [26], [27], such as self-organizing maps (SOM) [23], [24], [25], [26], [28].

As mentioned before, segmentation of MR images can be seen as a pattern classification and recognition problem. Thus, a pre-processing stage is necessary in order to make the segmentation more effective as well as a post-processing stage for ensuring the voxel grouping algorithm is performed correctly. On the other hand, all the above segmentation methods use some a priori knowledge from reference images [5].

Several fully-automated segmentation methods have been proposed, but most of them also use reference images for training [1], [5], [6], [7]. In [5], an automatic segmentation framework which works in three steps is presented. It uses a combination of different techniques: (i) skull-stripping, (ii) intensity inhomogeneity correction and, finally, (iii) classification of the brain tissue by means of a Fuzzy Kohonen's Competitive Learning algorithm (F-KCL). In [6] a two-step algorithm is presented. In the first step, the noise is removed and in the second step, an unsupervised image segmentation method based on fuzzy C-means clustering algorithm is applied to the MR image in order to partition it into distinct regions.

In this work, we present two different MR image segmentation methods. The first one uses information from the volume image histogram to compose feature vectors to be classified by a SOM, referred as HFS-SOM method in the following. Then, SOM prototypes are clustered by a k-means algorithm. In order to compute the optimum k value for the best clustering performance, the Davies–Boulding index (DBI) [47] is computed for several trials. Thus, the SOM prototypes are grouped into k clusters providing the lower DBI, and each of these clusters corresponds to a different tissue in the image.

The second method splits the acquired images into overlapping windows, and computes first and second order statistical features from each window. A feature selection process is performed by means of multi-objective optimization using a genetic algorithm [32], [55] in order to select the most discriminative set of features. The selected features compose the feature vectors used as inputs to train the SOM. During the training stage, the SOM projects the input vectors into a two dimensional space and computes a number of prototypes. These prototypes are a generalization of the input space in a lower number of vectors, meaning the quantization the input space. SOM is a clustering algorithm itself, and it considers each SOM unit as a cluster [33]. Nevertheless, similar units have to be grouped as voxels belonging to the same tissue can be represented by similar prototypes [33]. This way, the SOM prototypes have to be clustered and the borders between clusters have to be defined. The computation of these borders can be addressed by a generic clustering algorithm such as k-means, or by a specific SOM clustering algorithm [46]. In this work, a specific algorithm (i.e. EGS-SOM) is devised to improve the segmentation performance which uses the accumulated entropy to cluster the SOM units.

The methods described in this paper do not use any a priori knowledge about the voxel classification, and result in fully-unsupervised methods for MRI image segmentation. In addition, it is not necessary to indicate the number of tissue classes to be found, as the algorithms compute this number maximizing the goodness of the overall clustering process.

The paper is organized as follows: Section 2 presents the materials and methods used in this work. It is divided into four subsections; Section 2.1 describes the image databases used in this work; Section 2.2 shows the pre-processing stage which is common to the two segmentation approaches; Section 2.3 presents a faster implementation of the method which uses information extracted from the image histogram for segmentation of the whole volume and; Section 2.4 shows a high resolution approach for image slices. Section 3 depicts the experimental results obtained from the evaluation of the proposed methods using the IBSR database and discusses the main questions derived from them and. Finally, conclusions are drawn in Section 4.