Combining stationary wavelet transform and self-organizing maps for brain MR image segmentation

https://doi.org/10.1016/j.engappai.2010.09.008Get rights and content

Abstract

This study presents an image segmentation system that automatically segments and labels T1-weighted brain magnetic resonance (MR) images. The method is based on a combination of unsupervised learning algorithm of the self-organizing maps (SOM) and supervised learning vector quantization (LVQ) methods. Stationary wavelet transform (SWT) is applied to the images to obtain multiresolution information for distinguishing different tissues. Statistical information of the different tissues is extracted by applying spatial filtering to the coefficients of SWT. A multidimensional feature vector is formed by combining SWT coefficients and their statistical features. This feature vector is used as input to the SOM. SOM is used to segment images in a competitive unsupervised approach and an LVQ system is used for fine-tuning. Results are evaluated using Tanimoto similarity index and are compared with manually segmented images. Quantitative comparisons of our system with the other methods on real brain MR images using Tanimoto similarity index demonstrate that our system shows better segmentation performance for the gray matter while it gives average results for white matter.

Introduction

MR imaging is a widespread method that is used to obtain high quality medical images. It is one of the popular, painless and noninvasive brain imaging techniques. Especially for brain imaging MR imaging reveals a unique view by providing high level spatial and contrast resolution. It is used for diagnosis of many diseases and gives high quality informative images of the inside structure of the brain. Multispectral images of tissues that have different contrast values such as T1-weighted, T2-weigted, PD and FLAIR are provided by diverse magnetic resonance parameters (Robb, 2000).

Image segmentation refers to the process of partitioning a digital image into multiple regions. The goal of segmentation is to simplify and change the representation of an image into something that is more meaningful and easier to analyze (Shapiro and Stockman, 2001). In medical image segmentation, different image components are used for analysis of different structures and tissues, spatial distribution of functional activities and pathologic regions. Segmentation of brain MR images consist of peeling the brain from skull and classifying as brain/not-brain or segmenting tissue parts as white matter, gray matter, cerebrospinal fluid or suspicious region (Aubert-Broche et al., 2006b, Acharyya et al., 2003).

For a great many clinical applications, it is a classical problem to segment brain images into the tissue types such as white matter, gray matter and cerebrospinal fluid. In addition to this, regions of the brain such as tumor, edema and lesion for diseased or injured people should be determined for the purpose of diagnosis and planning surgical operations (Alhoniemi et al.,, Boudraa and Zaidi, 2005, Boer et al., 2010, Balafar et al., 2010).

Medical image segmentation is usually performed manually. Areas of interest are drawn by expert radiologists and doctors. It is a time consuming and tiring process. Manual segmentation is not objective. Different segmentations done by different experts can be very different. The segmentation that is done by the same expert can be different under different circumstances. The brightness and contrast of the screen can affect the segmentation accuracy and the following analyses. Usage of computers for segmentation can overcome these problems (Güler et al., 2009).

In the medical image segmentation; thresholding, region-based segmentation, edge-based segmentation and classification based segmentation are the most frequently used techniques. Two main limitations are incorporated with thresholding: firstly, only two classes are generated; secondly it cannot be applied to multichannel images. In addition to these, thresholding typically does not take into account the spatial characteristics of an image. This causes it to be sensitive to noise and intensity inhomogeneity, which can frequently occur in MR images. Therefore, variations on classical thresholding have been proposed for medical image segmentation that combines information based on local intensities and connectivity (Pham et al., 2000). Region based segmentation algorithms are region growing algorithm (Chen and Pavlidis, 1980, Chaplot et al., 2006), divide-and-combine (Chen and Pavlidis, 1980) and watershed algorithm (Grau et al., 2004). The main drawback of histogram-based region segmentation is that the histogram provides no spatial information. Region growing algorithm approaches exploit the important fact that pixels which are close to each other have similar gray values (Boudraa and Zaidi, 2005). The requirement of manual interaction to obtain a seed point is the main disadvantage of region growing algorithm. A seed point must be planted for each region to be extracted. Region growing can be sensitive to noise which causes extracted regions to have holes or become disconnected. Divide-and-combine is an algorithm related to region growing, but it does not require a seed point, and has the same drawbacks (Pham et al., 2000). The watershed transformation constitutes one of the most powerful segmentation tools provided by mathematical morphology. But there are two disadvantages in the watershed algorithm. Firstly, classical watershed algorithm on gray images such as tissue images causes oversegmentation. Secondly, there are some regions which are not divided completely particularly in the transition regions of gray matter and white matter, or cerebrospinal fluid and gray matter (Kong et al., 2006). Edge based algorithms work with edge finders. Traditional Sobel and Laplace detectors can be used for this purpose (Gonzalez and Woods, 2002). Edge detection works well on images with good contrast between regions. However, the detection is limited in regions with low contrast. Furthermore, it is difficult to find correlation between the detected edges of the regions of interest (Boudraa and Zaidi, 2005). Classification based algorithms can be constructed according to brightness similarity, contour energy and curvilinear continuousness. Examples of classifiers are the k-nearest neighbor (Boer et al., 2010) and support vector machine (Chaplot et al., 2006). Classification based segmentation needs training. Performance of the classification based segmentation depends on inputs of the classifier and training parameters (Toulson and Boyce, 1992). Knowledge-based expert systems are another methodology that is used for segmentation of images (Gerig et al., 1992, Gonzales et al., 2004). These systems require knowledge engineering and domain experts for building and maintaining the rules and the system. Brain MR images are very complex to be efficiently described and handled using rules. Another widely used method for brain image segmentation is the fuzzy c-means algorithm. This algorithm is not efficient by itself, because it fails to deal with the significant property of images that neighbor pixels are strongly correlated. Ignoring the correlation of neighboring pixels leads to strong noise sensitivity and several other imaging artifacts (Szilágyi et al., 2007).

Neural networks are one of the classification based segmentation methods. They perform classification by a method that learns from data, instead of using a given rule set. They organize themselves in a data driven manner. Neural networks draw attention consistently due to their ability of self-learning, fault tolerance and capability of searching for the optimum. Neural networks are constituted from lots of nonlinear computation elements that work in parallel and they are organized in a design that is similar to biological neural networks. Neural networks change their response according to the environmental conditions, learn from their experience and perform generalizations from old samples. SOM is an unsupervised neural network that use competitive learning algorithm. It is one of the most popular networks in the neural network field. SOM networks have advantages like they can automatically form similarity diagrams according to the input data (Chaplot et al., 2006). SOM maps high dimensional inputs to one or two dimensional discrete lattice of neurons. It organizes input data into several patterns according to a similarity factor like Euclidean distance. SOM learns both the distribution and the topology of input data. In other words, the network preserves topological relationships in its input and maps neighbor inputs to neighbor neurons (The Internet Brain Segmentation Repository (IBSR), 2010, Jiang and Zhao, 2003). Several studies are performed that use SOM for the segmentation of brain MR images (Acharyya et al., 2003, Bayram et al., 2001, The Internet Brain Segmentation Repository (IBSR), 2010, Kaus et al., 1999, Kong et al., 2006). Following an unsupervised learning process accomplished by a SOM network, usually a vector quantization method, LVQ, is used to calibrate the neurons of the network to find their best location (Kohonen, 2002).

The nonlinear character of MR images makes classical statistical methods particularly inappropriate for use for the segmentation process. The major advantage of neural networks over classical statistical pattern recognition techniques is their relative insensitivity to the selection of the training sets. This property has been shown to be of importance both for single slice and multiple slice classification. Neural networks do not rely on any assumption about the underlying probability density functions, thus possibly improving the results when the data significantly depart from normality (Ozkan et al., 1993). Based on the success of the neural network algorithm in pattern recognition and classification of different tissues in terms of texture, intensity or contrast, SOM algorithm is used in this study for the segmentation task.

Variations in the magnetic fields of the scanner cause image intensity inhomogeneities called bias field or non-uniformity. The result is multiplication of image intensities with a bias that slowly varies spatially. This causes problems for the tasks such as registration and segmentation. To overcome this problem space-invariant filtering techniques like low-pass filtering or neighborhood averaging are applied to the image. However, these filtering methods blur the important features in the image while suppressing the noise. Space-variant filtering techniques aim to address this limitation by using local, feature-dependent strategies. The approaches include recursive low-pass filtering with adaptive coefficients, linear least-squares error filtering, local shape-adaptive template filtering and anisotropic diffusion filtering. Among these techniques anisotropic diffusion filtering technique improves the signal-to-noise ratio, blurs homogeneous regions and preserves the boundaries and interesting structures (Kohonen et al., 2000, Li et al., 1993, Morra et al., 2003). Therefore to eliminate bias field and random noise, anisotropic diffusion filtering is used as a preprocessing step in this study.

Accomplishing a true and reliable segmentation of the brain MR images depends on the selection of the best features that represent different tissue types and that will be used as input to the system. Multiresolutional wavelet analysis provides the subimages of an image localized in different spatial frequencies (Gonzalez and Woods, 2002). It divides the 2D frequency spectrum of an image into one lowpass (approximation) and three highpass (horizontal, vertical and diagonal) subimages. Due to multiresolutional modeling capacity of the wavelets, obtained subimages are in different scales and frequencies. Because these images have different characteristics, it is quite convenient to use them for distinguishing tissue types from each other and tissue analysis (Tsai and Hsiao, 2001).

Wavelet coefficients alone are insufficient for describing tissue properties exactly. Even though they are useful for splitting the textured information into different frequency channels, they lack the local statistical information around a given pixel. Therefore, usually a spatial filtering operation is applied to the wavelet coefficients by sliding a window through the coefficients. Thus local statistics, which are mean absolute deviation (MAD), entropy, energy and standard deviation that describe texture properties are calculated. MAD is the mean of absolute deviations of the data. It represents the uniformity of the textures. Entropy measures the randomness of the texture. Energy indicates if the texture is broader, coarser or finer. Energy and entropy can differentiate homogeneous and non-homogeneous regions. Standard deviation is a measure of average contrast.

In this study, an anisotropic filtering preprocess is applied to brain MR images for the purpose of segmentation. By this way, the qualities of the images are improved. Then SWT is applied to the images. Diverse features are extracted from the subimages obtained from the transform, by applying the spatial filtering process. These features are combined together with the raw wavelet transform coefficients to obtain a feature vector. This feature vector is applied to the SOM network. The network is trained using unsupervised training methodology. LVQ procedure is utilized for calibrating the map. Results obtained from the system are evaluated using similarity indexes and compared with the manually segmented images.

The rest of this paper is organized as follows. In Section 2 materials and methods utilized in this study are described. Brain MR images used in the study are introduced. Employed methods, which are anisotropic diffusion filtering, stationary wavelet transform, self-organizing maps and learning vector quantization are presented. Feature extraction methods and evaluation metrics are also explained. Section 3 is dedicated to the computational experiments. Details and parameters of the system are given in this section with the results and discussions. Section 4 gives concluding remarks and addresses the future work.

Section snippets

Brain MR images

Simulated and manually segmented MR images play an important role for development of medical image analysis algorithms, especially for segmentation algorithms. Image analysis methods must be tested and evaluated in a controlled environment. Simulated and manually segmented images are very useful tools for validation because the ground truth is known. Segmentation accuracy can be evaluated by comparing the results with these images (Ozkan et al., 1993, Pham et al., 2000, Perona and Malik, 1990,

Results and discussion

Images that are used for the training and testing processes are obtained from the IBSR database. This database contains T1-weighted MR images of 20 normal people and the segmentations of these images as gray matter, white matter and other, which is performed by experts. Each MR image shows a brain slice of 3.1 mm thickness and the size of the voxels are 1.17×1.17×3.1 mm3. The size of the images is 256×256 pixels. IBSR data sets have various difficulty levels and they include low contrast scans.

Conclusion

This paper proposes an automatic brain MR image segmentation method. The proposed method contains preprocessing, feature extraction, segmentation and evaluation stages. Images obtained from the IBSR database are used for the training and testing processes. An anisotropic filtering preprocess is performed before segmentation to improve the quality of the brain MR images. Then SWT is applied to the images to obtain subimages that contain multiresolution information for distinguishing different

Acknowledgements

This study has been supported by Gazi University Scientific and Research Project Fund (Project no.: 07/2009-04).

References (48)

  • C. Vijayakumar et al.

    Segmentation and grading of brain tumors on apparent diffusion coefficient images using self-organizing maps

    Computerized Medical Imaging and Graphics

    (2007)
  • M. Acharyya et al.

    Extraction of features using m-band wavelet packet frame and their neuro-Fuzzy evaluation for multitexture segmentation

    IEEE Transactions on Pattern Analysis and Machine Intelligence

    (2003)
  • Alhoniemi, E., Himberg, J., Parhankangas,J., Vesanta, J., 2010. SOM Toolbox for MATLAB,...
  • J. Alirezaie et al.

    Automatic segmentation of cerebral MR images using artificial neural networks

    IEEE Transactions on Nuclear Science

    (1998)
  • B. Aubert-Broche et al.

    Twenty new digital brain phantoms for creation of validation image data bases

    IEEE Transactions on Medical Imaging

    (2006)
  • M.A. Balafar et al.

    Review of brain MRI image segmentation methods

    Artificial Intelligence Review

    (2010)
  • Bayram, E., Ge, Y., Wyatt, C.L., 2001. Confidence based anisotropic filtering of magnetic resonance images. In:...
  • P. Berkhin

    A survey of clustering data mining techniques

  • R. Boer et al.

    Accuracy and reproducibility study of automatic MRI brain tissue segmentation methods

    NeuroImage

    (2010)
  • A. Boudraa et al.

    Image segmentation techniques in nuclear medicine imaging

  • C.A. Cocosco et al.

    Brainweb: online interface to a 3D MRI simulated brain database

    NeuroImage

    (1997)
  • R.O. Duda et al.

    Pattern Classification

    (2000)
  • G. Gerig et al.

    Nonlinear anisotropic filtering of MRI data

    IEEE Transactions on Medical Imaging

    (1992)
  • J.O. Glass et al.

    Improving the segmentation of therapy-induced leukoencephalopathy in children with acute lymphoblastic leukemia using a priori information and a gradient magnitude threshold

    Magnetic Resonance in Medicine

    (2004)
  • Cited by (80)

    • Clustering ensemble-based novelty score for outlier detection

      2023, Engineering Applications of Artificial Intelligence
    • Unsupervised brain tumor segmentation using a symmetric-driven adversarial network

      2021, Neurocomputing
      Citation Excerpt :

      An unsupervised approach offers the advantage that it can exploit the abundant amount of unlabeled data generated during routine clinical imaging. Our work is related to unsupervised brain tumor segmentation methods that can be separated into 2 main groups [10]: a) Local based methods [11–22], where local features e.g., intensity values, are first calculated and then used for classification by an unsupervised classifier. Commonly used features and unsupervised classifiers include: thresholding [11], region growing [12,13], K-means [14], Fuzzy C-means [15–18], Markov Random Field (MRF) methods [19,20] and methods where multiple methods are combined [21,22].

    • A new active contour model driven by pre-fitting bias field estimation and clustering technique for image segmentation

      2021, Engineering Applications of Artificial Intelligence
      Citation Excerpt :

      It is widely used in many fields, including industrial automation production, medical image analysis, intelligent monitoring and traffic management. In the past decades, various image segmentation methods were proposed (Wu and Kang, 2021; Mousavirad et al., 2020; Sarkar et al., 2016; Çataloluk and Çelebi, 2018; Li et al., 2006; Gamino-Sánchez et al., 2018; Demirhan and Güler, 2011; Jin and Weng, 2019; Cai et al., 2018; Chan et al., 2018; Banerjee and Maji, 2019). Active contour model (ACM) is one of the most representative image segmentation methods since Kass et al. (2011) proposed.

    View all citing articles on Scopus
    View full text