3, pp. 343-368, 1995. [2] R. P. Woods, S. T. Grafton, J. D. G. Watson, N. L. ... vol. 11, pp: 1-16, 2001. [9] R. A. Bermel, J. Sharma, C. W. Tjoa, S. R. Puli, and R.
Proceedings of the WSEAS International Conference on ENVIRONMENT, MEDICINE and HEALTH SCIENCES
A New Approach for Segmentation of Brain MR Image Alaa M. Riad, Ahmed Atwan, Hazem M. El-Bakry, Reham R. Mostafa Faculty of Computer Science & Information Systems, Mansoura University, EGYPT
Hamdy K. Elminir National Research Institute of Astronomy and Geophysics, Helwan, Cairo, Egypt
Nikos Mastorakis Technical University of Sofia, BULGARIA
Abstract: In the frame of medical imaging, accurate segmentation of brain MR images is of interest for many brain disorders. However, due to several factors such noise, imaging artifacts, intrinsic tissue variation and partial volume effects, tissue segmentation remains a challenging task. So, in this paper, a full automatic framework for segmentation of brain MR images is presented. The framework consists of three-step segmentation procedure. First, segmentation of brain/non-brain tissue is performed by using Hybrid watershed algorithm (HWA). Then the intensity inhomogeneity correction method is applied to MR image. Finally, Fuzzy Kohonen's Competitive Learning (F-KCL) Algorithms are used for MRI tissue segmentation. The efficiency of the proposed framework is demonstrated by extensive segmentation experiments using simulated MR images. Keywords: Magnetic resonance imaging, MRI segmentation, Skull stripping, Intensity in homogeneity correction.
1. Introduction Magnetic resonance (MR) imaging has been widely applied in biological research and diagnostics, primarily because of its excellent soft tissue contrast, non-invasive character, high spatial resolution and easy slice selection at any orientation. In many applications, its segmentation plays an important role on the following sides: (a) identifying anatomical areas of interest for diagnosis, treatment, or surgery planning paradigms; (b) preprocessing for multimodality image registration; and (c) improved correlation of anatomical areas of interest with localized functional metrics [1]. Intracranial segmentation commonly referred to as skullstripping, aims to segment the brain tissue (cortex and cerebellum) from the skull and non-brain intracranial tissues in magnetic resonance (MR) images of the brain. Skullstripping is an important preprocessing step in neuroimaging analyses because brain images must typically be skullstripped before other processing algorithms such as registration, tissue classification or bias field correction can be applied [2- 6]. In practice, skull-stripping is widely used in neuroimaging analyses such as multi-modality image fusion and inter-subject image comparisons [2], [3]; examination of the progression of brain disorders such as Alzheimer’s Disease [7, 8], multiple sclerosis [9- 12] and schizophrenia [13], [14]; monitoring the development or aging of the brain [15], [16]; and creating probabilistic atlases from large groups of subjects [2]. Numerous automated skull-stripping methods have been proposed [17], [18], [19], [20], [5], [21] are widely used. However, the performance of these methods, which rely on signal intensity and signal contrast, may be influenced by numerous factors including MR signal inhomogeneities, type of MR image set, gradient performance, stability of system electronics, and extent of neurodegeneration in the subjects studied [21].
ISSN: 1790-5125
74
Fully automatic brain tissue segmentation of magnetic resonance images (MRI) is of great importance for research and clinical study of much neurological pathology. The accurate segmentation of MR images into different tissue classes, especially gray matter (GM), white matter (WM) and cerebrospinal fluid (CSF), is an important task. Moreover, regional volume calculations of these tissues may bring even more useful diagnostic information. Among them, the quantization of gray and white matter volumes may be of major interest in neurodegenerative disorders such as Alzheimer disease, in movements disorders such as Parkinson or Parkinson related syndrome, in white matter metabolic or inflammatory disease, in congenital brain malformations or perinatal brain damage, or in post traumatic syndrome. The automatic segmentation of brain MR images, however, remains a persistent problem. Automated and reliable tissue classification is further complicated by the overlap of MR intensities of different tissue classes and by the presence of a spatially smoothly varying intensity inhomogeneity. Here a full automatic framework for brain MRI segmentation is presented. The system combines the HWA method [20] for non brain tissue removal, the bias field correction schema by Wells et al. [22], and Fuzzy Kohonen's Competitive Learning Algorithms for tissue segmentation. Extensive experiments using simulated MR image data show that the proposed method can produce good segmentation results. The reminder of this paper is organized as follows. Section 2 presents the proposed segmentation framework. Simulation results are introduced in section 3. Finally conclusion is given in Section 4.
2. THE PROPOSED FRAMEWORK The proposed full automatic framework for brain MR image segmentation consists of a sequence of processing
ISBN: 978-960-474-170-0
Proceedings of the WSEAS International Conference on ENVIRONMENT, MEDICINE and HEALTH SCIENCES
steps are shown in flow diagram form in Figure 1 and includes: (1) Input the MRI image data; (2) Removal of nonbrain tissue (3) intensity inhomogeneity (IIH) correction; (4) Fuzzy Kohonen's Competitive Learning Algorithms for brain tissue segmentation.
automated processing of both legacy and contemporary image sets: anisotropic filter = 5 iterations with 5.0 diffusion constant; edge detector kernel = 0.8 sigma. These parameters were utilized in this study. (2) Brain Extraction Tool (BET): BET [21] employs a deformable model to fit the brain’s surface using a set of “locally adaptive model forces”. This method estimates the minimum and maximum intensity values for the brain image, a “centre of gravity” of the head image, and head size based on a spherical equivalent, and subsequently initializes the triangular tessellation of the sphere’s (head’s) surface. BET v. 1.2 is freely available in the FMRIB FSL Software Library (http://www.fmrib.ox.ac.uk/fsl/). The developer recommended the default parameters for automated processing of both the legacy and contemporary images. The parameters utilized in the application herein are the default parameters, described as follows: fractional intensity threshold = 0.5; vertical gradient in fractional intensity threshold = 0.
MRI data
Skull Stripping
Intensity inhomogeneity correction
Tissue segmentation
(3) Minneapolis Consensus Strip (McStrip):
Figure 1: The flow chart of the proposed framework of automatic MR image segmentation A. Skull stripping process Segmentation of brain/non-brain tissue is one of the most time-consuming preprocessing steps performed in neuroimaging laboratories. As a result, numerous brain extraction algorithms (BEAs) have been developed to perform this step automatically. While BEAs speeds up overall image processing, their output varies greatly and can affect the results of subsequent image analysis. Here we provide a quantitative comparison of the performance of four BEAs – Brain surface Extractor(BSE), Brain Extraction Tool(BET), hybrid Watershed Algorithm (HWA) and McStrip - against the ‘‘gold standard’’ of expert manual brain extraction using high-resolution T1-weighted MRI brain volumes. (1) Brain Surface Extractor (BSE): BSE is an edge-based method, uses a sequence of anisotropic diffusion filtering, Marr-Hildreth edge detection, and morphological processing to segment the brain within whole head MRI. Briefly, BSE works in three steps: first, the image is smoothed to reduce noise using anisotropic diffusion filtering. Second, edge detection (Marr-Hildreth edge detector) is applied to the smoothed image. Finally, the edge image is further processed to identify the largest connected region and to smooth the surface of this region. The largest remaining connected region is assumed to represent the brain. An additional dilation and erosion is performed to fill in surface pits and small holes [19], [5]. BSE v. 3.3 is freely available for download from http://neuroimage.usc.edu/brainsuite/, the BrainSuite website, the developers recommended the following parameters for ISSN: 1790-5125
75
McStrip[23], [24] is an automatic hybrid approach to brain extraction from T1-weighted MR volumes that uses a hierarchy of masks created by different models to form a consensus mask. The algorithm (McStrip) incorporates atlasbased extraction via nonlinear warping, intensity-threshold masking with connectivity constraints, and edge-based masking with morphological operations. McStrip is an automatic hybrid algorithm implemented in IDL that incorporates BSE and requires no user intervention; it relies on warping to a template, intensity thresholding, and edge detection. (4) Hybrid Watershed Algorithm - Version 1.21 (HWA): HWA [20] method is a hybrid of a watershed algorithm [18] and a deformable surface model [17] that was designed to be conservatively sensitive to the inclusion of brain tissue. In general, watershed algorithms segment images into connected components, using local optima of image intensity gradients. Deformable surface-model is then applied to locate the boundary of the brain in the image. HWA v. 1.21 is freely available as a component of the FreeSurfer software package at http://surfer.nmr.mgh.harvard.edu/. HWA developers recommended the default parameters for automated processing of both legacy and contemporary images. The parameters utilized in this study are the hard-coded default parameters of HWA. Performance Measure: To compare the performance of various segmentation techniques, we compute different coefficients reflecting how well two segmented volumes match. The manually segmented brains are used as a gold standard, and the automatically extracted brains are compared to them. To provide fair comparison between methods, we use a different performance measure: ISBN: 978-960-474-170-0
Proceedings of the WSEAS International Conference on ENVIRONMENT, MEDICINE and HEALTH SCIENCES
datasets#1, and the “gold standard” brain mask. The brief definition of the evaluation methods and the average evaluation result are summarized in Table 1, Table 2.
1. Jaccard similarity coefficient [35]: The Jaccard similarity coefficient JSC is formulated as JSC = vol ( S ∩ S 1
2
)
vol ( S ∪ S 1
2
)
(1)
where S 1 is the automatically skull stripped region, S is the 2
brain
region of the manually stripped image, and vol ( X ) denotes the volume of the region X . A Jaccard similarity coefficient of 1.0 represents perfect overlap, whereas an index of 0.0 represents no overlap. JSC values of 1.0 are desired. 2. Dice Similarity Coefficient: Dice Similarity Coefficient [36] is used to show the similarity level of automatically stripped region to manual stripped image. The Dice coefficient is defined as D = 2vol ( S ∩ S 1
2
)
vol ( S + S 1
2
) = 2 JSC
(1 + JSC )
(2)
where S1 is the automatically skull stripped region, S 2 is the brain region of the manually stripped image, and vol ( X ) denotes the volume of the region X . A Dice similarity coefficient of 1.0 represents perfect overlap, whereas an index of 0.0 represents no overlap. D values of 1.0 are desired. 3. Sensitivity and Specificity [37]: We also compute the sensitivity and specificity coefficient of the automated segmentation result using the manually segmented brain mask. The Sensitivity is the percentage of brain voxels recognized by the algorithm (Equation 3). The Specificity is the percentage of non-brain voxels recognized by the algorithm (Equation 4).
Sensitivity =
TP TP + FN
Specificity =
TN TN + FP
Table 1: Dataset#2, BEA performance on repeat scans of a single object vs. manual mask Performance measure
JSC
Dice
Sensitivity
Specificity
time
BEA BET
0.79
0.76
0.90
0.902
BSE
0.82
0.88
0.907
0.97
McStrip
0.93
0.95
0.94
0.98
HWA
0.96
0.98
0.97
0.99
10 sec 2 min 6 min 4 min
(3)
(4)
Where TP and FP stand for true positive and false positive, which were defined as the number of voxels correctly and incorrectly classified as brain tissue by the automated algorithm. TN and FN stand for true negative and false negative, which were defined as the number of voxels correctly and incorrectly classified as non-brain tissue by the automated algorithm. 4. Processing Time: The time necessary for generating the final brain mask for each Brain Extraction Algorithm was recorded. The four Brain Extraction Algorithms (BEAs) are tested using two datasets of brain T1-wieghted MR images and evaluated the result against manual skull- stripped image that used as a “gold standard”. Moreover, Figure 2 shows the brain mask that obtained from the four BEAs on the slice in ISSN: 1790-5125
Figure 2: Original MRI, manual mask, BSE mask, BET mask, McStrip mask, and HWA mask.
76
Table 2: Dataset#2, BEA performance on repeat scans of a single object vs. manual mask Performance measure
JSC
Dice
Sensitivity
Specificity
time
BEA BET
0.80
0.75
0.903
0.90
BSE
0.82
0.87
0.90
0.963
McStrip
0.93
0.94
0.94
0.97
HWA
0.95
0.98
0.973
0.987
15 sec 2.5 min 6 min 3.67 min
HWA algorithm performed best when compared to manual mask with regard to JSC, Dice, sensitivity and Specificity. But BET performed fastest; creating mask in less than 1 min.
ISBN: 978-960-474-170-0
Proceedings of the WSEAS International Conference on ENVIRONMENT, MEDICINE and HEALTH SCIENCES
The comparison of four BEAs techniques against expertly hand stripped T1-wieghted MRI volumes revealed that HWA algorithm, a hybrid approach incorporating watershed algorithm and deformable model, outperformed BSE, BET, and McStrip, all of which rely on a single algorithmic strategy.
Figure 3 presents the results of the estimated bias field. The method consists two steps: 1: Estimation of the tissue type for all the pixel y i of the image. 2: Estimation of the inhomogeneity β i = mean ( R j ) at each location i in the image, and correction according to Eq.6. The method is iterative, and is initialized with inhomogeneities equal to zero.
B. Intensity Inhomogeneity (IIH) correction: Inhomogeneity in magnetic fields during image acquisition and magnetic susceptibility variations in scanned subjects cause intensity non-uniformities, also described as bias fields. These artifacts prevent characterization of voxel tissue content based solely on image intensity. As a result, segmentation as well as quantitative studies of MR images requires compensation for these non-uniformities. The method we implement is due to Wells [22]. Let Y be the observed image, X the ideal image, B the inhomogeneity, and N the noise present in the image. The interactions between those fields can be described as:
Y = X ×B+N
The noise is considered negligible. Taking the logarithm of Eq.5 gives: (6)
Assuming that the pixel intensities of a tissue type are normally distributed, the probability for a pixel to belong to a class, in absence of inhomogeneities, can be expressed as:
1 yi − µ k 2 p ( yi / x k ) = exp − 2 σ 2πσ k k = Gσ ( yi − µ k ) 1
k
(7)
where y i represents the intensity of image pixel, k number of a single class, µ k mean of class k, σ k represents the standard deviation of class k. In the presence of inhomogeneities, and according to Eq.6 and Eq.7, the same probability can be expressed as:
p ( yi / xk , β i ) = Gσ
βi
k
( xi − µ k − β i )
(8)
i th pixel location. To estimate the inhomogeneity, the residual term R i is
where
is the inhomogeneity at this
computed for each pixel:
Ri = ∑ p ( xk / yi ) k
yi − µ k
σk
(9)
i th pixel location, βi ,is the average of the R i in a 3 × 3 neighborhood of pixel i . The inhomogeneity at the
ISSN: 1790-5125
(b)
(c)
Figure 3: Illustration of estimation bias field. (a) Input image (b) estimated bias field (c) corrected image.
(5)
log (Y ) = log( X ) × log( B )
(a)
77
C. Segmentation: The goal of brain magnetic resonance image segmentation is to accurately identify the principal tissue structures in these image volumes. There are many methods that exist to segment the brain. One of these, conventional methods that use pure image processing techniques are not preferred because they need human interaction for accurate and reliable segmentation. Unsupervised methods, on the other hand, do not require any human interference and can segment the brain with high precision. In the light to this fact, we compare here the performance of four image segmentation techniques in the subject of brain MR image. Results show that Fuzzy Kohonen's Competitive Learning (F-KCL) Algorithms performs better in terms of segmentation accuracy, while Fuzzy C-Means ( FCM) performs better in terms of speed of computation. Methods (1) The Self-organizing Feature Map algorithm (SOFM) With the aim of obtaining adaptive image processing, researchers have tried to employ neural network (NN) approaches. Here, the basic objective is to emulate the human vision processing system which is highly robust and noise insensitive and hence can be applied even when information is ill defined and/or defective/partial. A self-organizing map (SOM) is a type of artificial neural network that is trained using unsupervised learning to produce a low-dimensional (typically two dimensional), discretized representation of the input space of the training samples, called a map. The map seeks to preserve the topological properties of the input space. Every input is connected extensively to every output node via adjustable weights. Let X =[x0 , x1, x2 ,....., xn−1] be a set of T
ISBN: 978-960-474-170-0
Proceedings of the WSEAS International Conference on ENVIRONMENT, MEDICINE and HEALTH SCIENCES
N inputs in R features). Let P be
N
such that each
x i has N dimensions (or
the
of
number
output
node
and
W j = [ w0 j , w1 j ,......, w( N −1) j ] denote the weights or reference T
vectors.
x i denotes the input to output node j and w ij is the
weight from input node i to the output node j. W j is the vector containing all of the weights from N input nodes to output node j. Updating weights for any given inputs in SOFM form is done only for output units in a localized neighborhood. The neighborhood is centered on the output node whose distance d ij is minimum. The measurement of
d ij is an Euclidean distance, defined as: d ij = min j || x i − w ij ||
In fuzzy segmentation, each pixel is assigned a membership value in each of the c regions. If the memberships are taken into account while computing properties of regions, we obtain more accurate estimates of region properties. One of the known techniques to obtain such a classification is the FCM algorithm [29, 30]. The FCM algorithm is an unsupervised technique that clusters data by iteratively computing a fuzzy membership function and mean value estimates for each class. The fuzzy membership function, constrained to be between 0 and 1, reflects the degree of similarity between the data value at that location and the prototypical data value, or centroid, of its class. Thus, a high membership value near unity signifies that the data value at that location is close to the centroid of that particular class. Fuzzy c-means algorithm
2
(10)
The neighborhood decreases in size with time until only a single node is inside its bounds. A learning rate, σ ij (t ) , is also required which decreases monotonically in time. The weight updating rule is as follows:
w ij (t + 1) = w ij (t ) + σ ij (t )( x i − w ij (t ))
(11)
The FCM algorithm, also known as Fuzzy ISODATA, is one of the most frequently used methods in pattern recognition. The FCM algorithm assigns pixels to each category by using fuzzy memberships. Let X = ( x 1 , x 2 ,......., x N ) denotes an image with N pixels to be partitioned into c clusters, where x i represents multispectral (features) data. The algorithm is an iterative optimization that minimizes the cost function defined as follows: N
The algorithm works as shown in [26], [27] and [28]. However, SOFM algorithms are, firstly, highly dependent on the training data representatives and the initialization of the connection weights. Secondly, they are very computationally expensive since as the dimensions of the data increases, dimension reduction visualization techniques become more important, but unfortunately the time to compute them also increases. For calculating that black and white similarity map, the more neighbors we use to calculate the distance the better similarity map we will get, but the number of distances the algorithm needs to compute increases exponentially. (2) Fuzzy c-means for image segmentation The objective of image segmentation is to divide an image into meaningful regions. Errors made at this stage would affect all higher level activities. Therefore, methods that incorporate the uncertainty of object and region definitions and the faithfulness of the features to represent various objects are desirable. In an ideally segmented image, each region should be homogeneous with respect to some predicate such as gray level or texture, and adjacent regions should have significantly different characteristics or features. More formally, segmentation is the process of partitioning the entire image into c crisp maximally connected regions { Ri } such that each Ri is homogeneous with respect to some criteria. In many situations, it is not easy to determine if a pixel should belong to a region or not. This is because the features used to determine homogeneity may not have sharp transitions at region boundaries. To alleviate this situation, we can inset fuzzy set concepts into the segmentation process. ISSN: 1790-5125
78
c
J = ∑∑ u ijm || x j −v i ||2
(12)
j =1 i =1
Where
u ij represents the membership of pixel x j in the ith
cluster, v i is the ith cluster center, ||.|| is a norm metric, and m is a constant. The parameter m controls the fuzziness of the resulting partition, and m =2 is used in this study. The cost function is minimized when pixels close to the centroid of their clusters are assigned high membership values, and low membership values are assigned to pixels with data far from the centroid. The membership function represents the probability that a pixel belongs to a specific cluster. In the FCM algorithm, the probability is dependent solely on the distance between the pixel and each individual cluster center in the feature domain. The membership functions and cluster centers are updated by the following:
uij =
1
|| x − v || ∑ || x j − v i || k =1 j k
2 / ( m −1)
(13)
c
and N
∑u vi =
m ij
xj
j =1
(14)
N
∑u
m ij
j =1
Starting with an initial guess for each cluster center, the FCM converges to a solution for v i representing the local minimum or a saddle point of the cost function. Convergence can be detected by comparing the changes in the membership ISBN: 978-960-474-170-0
Proceedings of the WSEAS International Conference on ENVIRONMENT, MEDICINE and HEALTH SCIENCES
function or the cluster center at two successive iteration steps. (3) An Adaptive Neuro-Fuzzy System for Automatic Image Segmentation. Auto adaptive neuro-fuzzy segmentation architecture is presented [31]. The system consists of a multilayer perceptron (MLP)-like network that performs image segmentation by adaptive thresholding of the input image using labels automatically pre-selected by a fuzzy clustering technique. The system's architecture is feedforward, but unlike the conventional MLP the learning is unsupervised. The output status of the network is described as a fuzzy set. Fuzzy entropy is used as a measure of the error of the segmentation system. Given an input image, the system is forced to evolve toward a minimum fuzzy entropy state in order to obtain image segmentation. The system is capable to perform automatic multilevel segmentation of images, based solely on information contained by the image itself. No a priori assumptions whatsoever are made about the image (type, features, contents, stochastic model, etc.). Such a “universal” algorithm is most useful for applications that are supposed to work with different (and possibly initially unknown) types of images.
Figure 4: The block diagram of our comparative study process
Figure 5: The planar simulated T1-weghted brain images
(4) A Fuzzy Kohonen's Competitive Learning Algorithms for MR Image Segmentation Kohonen’s self-organizing feature map (SOFM) is a twolayer feedforward competitive learning network, and has been used as a competitive learning clustering algorithm in brain MRI image segmentation. However, most brain MRI images always present overlapping gray-scale intensities for different tissues. In a Fuzzy Kohonen's Competitive Learning Algorithms [32], fuzzy methods are integrated with Kohonen’s competitive algorithm to overcome this problem (the name of algorithm F_KCL). The F_KCL algorithm fuses the competitive learning with fuzzy c-means (FCM) cluster characteristic and can improve the segment result effectively. The comparison between different techniques mentioned above in this paper will doing as shown in figure 4. The MRI image is first segmented using one of the segmentation techniques listed above in this paper. Then the segmented image is separated into three images corresponding to WM, GM, and CSF. Then these images will be compared to the references image using mean squared error to measure the segmentation accuracy. The MR images used in this paper are obtained from the http://www.bic.mni.mcgill.ca/brainweb web site in Montreal Neurological Institute, McGill University, McConnell Brain Imaging Centre (McBIC) [25]. The database is the result of a research work developed at McBIC and contains quantitative 3D investigation of brain structure and function. The brain phantom and simulated MR images have been made publicly available and can be used to test algorithms such as classification procedures which seek to identify the tissue ‘type” of each image pixel. The modality, T1weighted, are downloaded from the website as our experimental data shown in Figure 5. And Figure 6 shows the brain MR image Phantoms. They are considered as the true segmented tissues used in this paper. ISSN: 1790-5125
79
Grey Matter
CSF
White matter
Figure 6: The brain MR image Phantoms The segmented Grey matter, CSF, and White Matter using SOFM neural network are shown in Figure 7, the segmented images using the Fuzzy c-means are shown in Figure 8, the segmented images using neuro-fuzzy system are shown in Figure 9, the segmented images using a Fuzzy Kohonen's Competitive Learning algorithm are shown in Figure 10.
Grey Matter
CSF
White matter
Figure 7: The segmented tissues using SOFM neural network
Grey Matter
CSF
White matter
Figure 8: The segmented tissues using Fuzzy c-means
ISBN: 978-960-474-170-0
Proceedings of the WSEAS International Conference on ENVIRONMENT, MEDICINE and HEALTH SCIENCES
Grey Matter
CSF
and CSF segmentation owes to the Fuzzy Kohonen's Competitive Learning (F-KCL) Algorithms. The segmentation accuracy was measured by using an average overlap metric (AOM) [33], which is quantitative evaluation of performance. Overlap metric is defined for a given voxel class assignment as the sum of the number of voxels that both have the class assignment in each segmentation divided by the sum of voxels where either segmentation has the class assignment. This metric approaches a value of 1.0 for results that are very similar and is near 0.0 when they share no similarly classified voxels. The AOM measurements for results obtained in different situations are given in Table 4.
White matter
Figure 9: The segmented tissues using neuro-fuzzy system
Grey Matter
CSF
White matter Image 1 3% Noise 20% bias
Figure 10: The segmented tissues using Fuzzy Kohonen's Competitive Learning Algorithm Table 3: shows the errors between the segmented images using the techniques mentioned above in Figure 7, 8, 9, 10 and the brain MR image Phantoms (standard image) in Figure 6. Because the number of pixels in different segments (such as Gray matter, CSF, and White Matter.) is different, the segmentation error is compared individually for each segmented image. The performance measurement is the sum of mean squared errors pixel by pixel. n
∑ (o − s ) i
2
i
(15)
i =1
Where o i is the intensity value in a Phantom image at pixel
i . s i is the intensity value in the corresponding segmented
Image 3 3% Noise 40% bias
Image 2 5% Noise 20% bias
Image 4 5% Noise 40% bias
Figure 11: The simulated MR images under four conditions So, the four simulated T1-wieghted brain MR images with different noise and intensity in-homogeneity level that are used as our experimental data are shown in Figure 12. In addition, Figure 12 shows the brain MR image phantom (binary segmentation) that are considered as the true segmented tissues.
image at pixel i. n is the total pixel number for the concerned tissue type in an image. Table 3. The mean squared error errors between the segmented images and Phantoms Tissue Gray CSF White Techniques type matter matter MSE 1.9 1.1 2.1 Self-organizing map MSE 1.5 0.7 1.7 Fussy c-means MSE 0.7 0.4 0.9 Neuro-fuzzy system MSE 0.6 0.2 0.7 F_KCL Result show that F-KCL performs better than other techniques in term of segmentation accuracy.
3. EXPERIMENTAL RESULT To demonstrate and evaluate the performance of the proposed framework, we mainly applied to simulated MR images. These data were downloaded from brainweb [25]. Experiments have been done under four different conditions that are shown in Figure 11. We first removed the non brain tissue using HWA algorithm, and the IIH correction was used to bring out corrected result. The final result on WM, GM, ISSN: 1790-5125
80
White Matter
Gray Matter
CSF
Figure 12: The brain MR image Phantoms The segmented Grey matter, CSF, and White Matter using the proposed framework for the first image (Image1: 3% noise and 20% bias) are shown in Figure 13. The segmented Grey matter, CSF, and White Matter using the proposed framework for the second image (Image2: 5% noise and 20% bias) are shown in Figure 14. The segmented Grey matter, CSF, and White Matter using the proposed framework for the third image (Image3: 3% noise and 40% bias) are shown in Figure 15. The segmented Grey matter, CSF, and White Matter using the proposed framework for the fourth image (Image4: 5% noise and 40% bias) are shown in Figure 16.
ISBN: 978-960-474-170-0
Proceedings of the WSEAS International Conference on ENVIRONMENT, MEDICINE and HEALTH SCIENCES
Table 4: Segmentation results on T1-weighted data under different conditions of noise and IIH
Image 1 Image 2 Image 3 Image 4
White Matter
Gray Matter
Noise 3% 5% 3% 5%
Bias 20 % 20 % 40 % 40 %
WM 0.95 0.91 0.93 0.89
GM 0.88 0.85 0.83 0.82
CSF 0.96 0.94 0.92 0.90
According to Zijdenbos’ statement [34] that AOM indicates excellent agreement when it is above 0.7, it can be seen that our system framework is very robust to IIH and noise. Though the accuracy decreased when noise and bias increased, the variance was small and acceptable. Its superiority mainly depended on preprocessing (IIH correction and skull stripping) before segmentation.
CSF
Figure 13: The segmented tissues of Image1 using proposed framework.
4. CONCLUSION
White Matter
Gray Matter
In this paper, a full automatic framework for brain MR image segmentation has been presented. Furthermore, a validation process for MR image segmentation has been demonstrated. It aims to obtain more accurate different tissues with the presence of intensity inhomogeneity (IIH) and noise. The proposed system which firstly receives MRI data, removes non brain tissues using Hybrid Watershed Algorithm (HWA) algorithm, and then correct the intensity inhomogeneity. Then the corrected stripped image is segmented into GM, WM, and CSF using Fuzzy Kohonen's Competitive Learning Algorithms. Under the quantitative comparison of the performance of four Brain Extraction Algorithm (BEAs) – Brain surface Extractor (BSE), Brain Extraction Tool (BET), Hybrid Watershed Algorithm (HWA) and McStrip - against the ‘‘gold standard’’ of expert manual brain extraction using high-resolution T1-weighted MRI brain volumes, simulation results have shown that HWA algorithm, a hybrid approach incorporating watershed algorithm and deformable model, outperforms BSE, BET, and McStrip, all of which rely on a single algorithmic strategy. Also under quantitative comparison of the performance of four brain tissue segmentation algorithms – Self organizing Map (SOM), Fuzzy c-means (FCM), Adaptive Neuro-Fuzzy System, and Fuzzy Kohonen's Competitive Learning Algorithms (FKCL)- against the “gold standard”. Simulation results have shown that FKCL outperforms other algorithms. Finally, the proposed system, that firstly receives MRI data, removes non brain tissues using HWA algorithm, then corrects the intensity inhomogeneity. After that, the corrected stripped image is segmented into GM, WM, and CSF using Fuzzy Kohonen's Competitive Learning Algorithms. Tests have been performed and the results have shown that our system framework is very robust to intensity inhomogeneity (IIH) and noise.
CSF
Figure 14: The segmented tissues of Image2 using proposed framework.
White Matter
Gray Matter
CSF
Figure 15: The segmented tissues of Image3 using proposed framework.
References White Matter
Gray Matter
CSF
[1] Clarke L.P, Velthijizen R.P, Camacho M.A ,Heine J. J, " MRI segmentation methods and applications", Magnetic Resonance Imaging, Vol. 13, No. 3, pp. 343-368, 1995. [2] R. P. Woods, S. T. Grafton, J. D. G. Watson, N. L. Sicotte, and J. C. Mazziotta, “Automated image
Figure 16: The segmented tissues of Image4 using proposed framework.
ISSN: 1790-5125
81
ISBN: 978-960-474-170-0
Proceedings of the WSEAS International Conference on ENVIRONMENT, MEDICINE and HEALTH SCIENCES
registration: II. Intersubject validation of linear and nonlinear models”, J. of Computer Assisted Tomography, vol. 22, pp: 139-152, 1998. [3] R. P. Woods, M. Dapretto, N. L. Sicotte, A. W. Toga, and J. C. Mazziotta, “Creation and use of a TalairachCompatible atlas for accurate, automated, nonlinear intersubject registration, and analysis of functional imaging data”, Human Brain Mapping, vol. 8, pp: 73-79, 1999. [4] J. Van Horn, T. M. Ellmore, G. Esposito, K. F. and Berman, “Mapping Voxel-based Statistical Power on Parametric Imaging”, NeuroImage, vol. 7, pp: 97-107, 1998. [5] D. W. Shattuck, S.R. Sandor-Leahy, K. A. Schaper, D. A. Rottenberg, and R. M. Leahy, " Magnetic Resonance Image Tissue Classification Using a Partial Volume Model", NeuroImage, vol. 13, pp: 856-876, 2001. [6] S. Strother, S. La Conte, L. Kai Hansen, J. Anderson, J. Zhang, S. Pulapura, and D. Rottenberg, “Optimizing the fMRI data-processing pipeline using prediction and reproducibility performance metrics: I. A preliminary group analysis”, NeuroImage, vol. 23, pp:196-207, 2004. [7] H. Rusinek, M. J. de Leon, A. E. George, L. A. Stylopoulos, R. Chandra, G. Smith, T. Rand, M. Mourino, and H. Kowalski, “Alzheimer disease: measuring loss of cerebral gray matter with MR imaging”, Radiology, vol. 178, pp: 109-114, 1991. [8] P. M Thompson, M. S. Mega, R. P. Woods, C. I. Zoumalan, C. J. Lindshield, R. E. Blanton, J. Moussai, C. J. Holmes, J. L. Cummings, and A. W. Toga, “Cortical change in Alzheimer’s disease in detected with a diseasespecific population-based brain atlas”, Cerebral Cortex vol. 11, pp: 1-16, 2001. [9] R. A. Bermel, J. Sharma, C. W. Tjoa, S. R. Puli, and R. Bakshi, “A semiautomated measure of whole- brain atrophy in multiple sclerosis”, Neurological Sciences, vol. 208, pp: 57-65, 2003. [10] M. A. Horsfield, M. Rovaris, M. A. Rocca, P. Rossi, R. H. B. Benedict, M. Filippi, and R. Bakshi, “Whole-brain atrophy in multiple sclerosis measured by two segmentation processes from various MRI sequences”, Neurological Sciences, vol. 216, pp: 169-177, 2003. [11] R. Zivadinov, F. Bagnato, D. Nasuelli, S. Bastianello, A. Bratina, L. Locatelli, K. Watts, L. Finamore, A. Grop, M. Dwyer, M. Catalan, A. Clemenzi, E. Millefiorini, R. Bakshi, and M. Zorzon, M., “ Short-term brain atrophy changes in relapsing-remitting multiple sclerosis”, Neurological Sciences, vol. 223, pp: 185-193, 2004. [12] J. Sharma, M. P. Sanfilipo, R. H. B. Benedict, B. Weinstock-Guttman, F. E. Munschauer, and R. Bakshi, “Whole-brain atrophy in mul tiple sclerosis measured by automated versus semi-automated MR imaging segmentation”, Neuroradiology, vol. 25, pp: 985-996, 2004. [13] K. L. Narr, P. M. Thompson, P. Szeszko, D. Robinson, S. Jang, R. P. Woods, S. Kim, K. M. Hayashi, D. Asunction, A. W. Toga,and R. M. Bilder, “Regional specificity of hippocampal volume reductions in firstepisode schizophrenia”, NeuroImage, vol. 21, pp: 15631575, 2004. [14] P. Tanskanen, J. M. Veijola, U. K. Piippo, M. Haapea, J. A. Miettunen, J. Pyhtinen, E. T. Bullmore, P. B. Jones, and M. K. Isohanni, “Hippocampus and amygdala ISSN: 1790-5125
82
volumes in schizophrenia and other psychoses in the Northern Finland 1966 birth cohort”. Schizophrenia Research, vol. 75, pp: 283-294, 2005. [15] T. L. Jernigan, S. L. Archibald, C. Fennema-Notestine, A. C. Gamst, J. C. Stout, J. Bonner, and J. R. Hesselink, “Effects of age on tissues and regions of the cerebrum and cerebellum”, Neurobiology of Aging, vol. 22, pp: 581594, 2001. [16] R. E. Blanton, J. G. Levitt, J. R. Peterson, D. Fadale, M. L. Sporty, M. Lee, D. To, E. C. Mormino, P. M. Thompson, J. T. McCracken, and A. W. Toga, “Gender differences in the left inferior frontal gyrus in normal children”, NeuroImage, vol. 22, pp: 626-636, 2004. [17] A.M. Dale AM, B. Fischl , and M. Sereno, “Cortical surface-based analysis. I. Segmentation and surface reconstruction”, Neuroimage, vol. 9, pp: 179-94, 1999. [18] H. Hahn, and H-O. Peitgen, “The skull stripping problem in MRI solved by a single 3D watershed transform”, MICCAI, vol. 1935, pp: 134-143, 2000. [19] S. Sandor, and R. Leahy, “Surface-based labeling of cortical anatomy using a deformable database”, IEEE Trans. Med. Imag., vol. 16, pp :41-54, 1997. [20] F. Segonne, A. M. Dale, E. Busa, M. Glessner, D. Salat, H. K. Hahn, and B. Fischl, “A hybrid approach to the skull stripping problem in MRI”, Neuroimage, vol. 22, pp :1060-75, 2004. [21] S. M. Smith, “Fast robust automated brain extraction”, Human Brain Mapping, vol. 17, pp: 143-55, 2002. [22] W.M.Wells, R.Kikinis, W.E.L.Grimson, and F.Jolesz, “Adaptive segmentation of MRI data”, IEEE Transactions on Medical Imaging, vol. 15, pp. 429–442, 1996. [23] K. Rehm, D. Shattuck, R. Leahy, K. Schaper, and D.A. Rottenberg, "Semi-automated stripping of T1 MRI volumes: I. Consensus of intensity- and edge-based methods", NeuroImage, vol. 9, 1999. [24] K. Rehm, K. Schaper, J. Anderson, R. Woods, S. Stoltzner, and D. Rottenberg, "Putting our heads together: a consensus approach to brain/ non-brain segmentation in T1-weighted MR volumes". NeuroImage, vol.22, pp: 1262– 1270, 2004. [25] http://www.bic.mni.mcgill.ca/brainweb: Simulated Brain Database, McConnell Brain Imaging Centre, Montreal Neurological Institute, McGill University. [26] C. A. Parra, K.Iftekharuddin and R. Kozma, “Automated brain data segmentation and pattern recognition using ANN,” in the Proceedings of the Computational Intelligence, Robotics and Autonomous Systems (CIRAS 03), December, 2003. [27] J. Alirezaie, M. E. Jernigan and C. Nahmias, “Automatic Segmentation of Cerebral MR Images using Artificial Neural Networks,” in IEEE Transactions on Nuclear Science, Vol. 45, No. 4, August, 1998. pp. 2174-2182. [28] J. Vesanto and E. Alhoniemi, “Clustering of the SelfOrganizing Map”, IEEE Transactions on Neural Networks, Vol. 11, No. 3, pp. 586-600, May 2000. [29] J. C. Dunn, "A fuzzy relative of the ISODATA process and its use in detecting compact well-separated clusters", Vol. 3, pp. 32– 57, 1973. [30] J. Bezdek, "A convergence theorem for the fuzzy ISODATA clustering algorithms", IEEE Trans. Pattern Anal. Mach, 1980. ISBN: 978-960-474-170-0
Proceedings of the WSEAS International Conference on ENVIRONMENT, MEDICINE and HEALTH SCIENCES
[31] G. R. Reddy, E. Suresh, S. Uma Maheshwar, and M. S. Reddy, "A Neuro-Fuzzy System for Automatic MultiLevel Image Segmentation using KFCM and Exponential Entropy ", in IFIP, Vol. 228, pp: 367-372, 2006. [32] J. Kong, W. Lu, J. Wang, N. Che, and Y. Lu. "A Modified Fuzzy Kohonen's Competitive Learning Algorithms Incorporating Local Information for MR Image Segmentation", IEEE trams. Pattern Anal., 2007. [33] http://www.cma.mgh.harvard.edu/ibsr/ [34] P.Z. Alex, "MRI segmentation and the quantification of white matter lesions", PhD thesis, Vanderbilt University, Electrical Engineering Department, Nashville, Tennessee; December 1994.
ISSN: 1790-5125
83
[35] K. H. Zou, S. K. Warfield, A. Bharatha, C. M. Tempany, M. R. Kaus, S. J. Haker, W. M. Wells, and R. Kikinis, “Statistical validation of image segmentation quality based on a spatial overlap index”, Acad Radiol, vol. 11, pp: 178-89, 2004. [36] L. Dice, “Measures of the amount of ecologic association between species”, Ecology, vol. 26, pp: 297302, 1945. [38] P. Perona, and J. Malik, " Scale-space and edge detection using anisotropic diffusion", IEEE Trans. Pattern Anal., vol. 12, pp: 629–639, 1990.
ISBN: 978-960-474-170-0
